Meta-Harnesses, Medical Midjourney, and Why Specialists Are Cooked

Meta-Harnesses, Medical Midjourney, and Why Specialists Are Cooked ![rw-book-cover](https://d34adp677peecb.cloudfront.net/static/images/article3.5c705a01b476.png) 🔗 **Source:** [Meta-Harnesses, Medical Midjourney, and Why Specialists Are Cooked](https://www.youtube.com/watch?v=TOlFh3BdGWU) > [!note]+ Document Notes > tags: [ai-coding-agents, token-budgets, agent-harnesses, ai-code-review, ai-design-workflows] > author: Whiskey Web and Whatnot > source: Whiskey Web and Whatnot (YouTube) > reference: https://www.youtube.com/watch?v=TOlFh3BdGWU > date created: 2026-10-12 11:15 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ Title: Meta-Harnesses, Medical Midjourney, and Why Specialists Are Cooked >  ## 🧭 Table of contents - [📝 Summary](#-summary) - [📖 Main content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning styles assessment](#-learning-styles-assessment) - [📊 Accessibility assessment](#-accessibility-assessment) - [🌍 DEIB assessment](#-deib-assessment) - [✅ Logical validity](#-logical-validity) - [🛡️ DARVO analysis](#️-darvo-analysis) - [🎯 Conclusion](#-conclusion) - [🔄 Refactor appendix](#-refactor-appendix--meta-harnesses-medical-midjourney-and-why-specialists-are-cooked) ## 📝 Summary - **What happened.** Whiskey Web and Whatnot published a 1:08:55 episode on 2026-10-11 in which the two hosts, Robbie Wagner and Adam Argyle (names from the description and the cold open), talk through their day-to-day AI coding work. There are no guests. The transcript carries no speaker labels, so this analysis refers to "the IBM host" and "the Shopify host" based on each person's own workplace references. Roughly 14 of the 69 minutes go to bourbon tasting (00:02:15 to 00:06:33), a robot lawn mower, and a whiskey bar visit (01:00:42 to 01:07:54). - **The claim.** The recurring thesis is that durable AI productivity comes from the harness (memory, recurring workflows, tests, review loops) and not from the model. The hosts also discuss a new layer called the "meta harness" (tools that spin up specialist agents on the fly). One host predicts it will be absorbed into products soon and says "this part's going to go away soon." - **Specifics offered.** An employer cap of $1,000 per month per engineer at one host's company (00:06:33); a model called Fable that one host used for about eight hours before it was pulled (00:13:09); an overnight run of 1 hour 45 minutes on an MMO spec (00:14:36); a recurring "Tinder for algos" design loop that took 800 art pieces to produce good results (00:30:27); a local test suite that takes 45 minutes (00:33:20). - **Title versus content.** "Medical Midjourney" covers about 1.5 minutes (00:40:32 to 00:41:57) and "why specialists are cooked" is a one-minute exchange at 00:56:22 to 00:57:48 where the hosts disagree. The title overstates both. The longest AI threads are budgets and harnesses (00:06:33 to 00:13:09, 00:36:16 to 00:40:32), design loops (00:24:34 to 00:32:36), and a debugging story about a self-hosted agent (00:47:00 to 00:52:43). - **Why it matters.** It's a candid look at what two working developers do after the novelty wears off: they cap spend, build repeatable loops, lean on tests and screenshot checks, and distrust agents that report "done." - **Key stakeholders.** The two hosts and their employers (IBM and Shopify by self-description), sponsor Mailtrap, vendors named in passing (Anthropic, GitHub Copilot, CodeRabbit, Midjourney, Railway), and the authors of the harness tools they mention. - **Commercial interests.** Mailtrap sponsors the episode and the description carries a tracking link (`l.rw.rw`). Both hosts work at companies that sell or use the tools under discussion. The host who prefers CodeRabbit says the employer doesn't pay for it (00:44:47), which makes the endorsement unpaid on that point. The hosts also sell merchandise (01:08:38, `whiskey.fund`). - **Risks and tradeoffs.** Every number is self-reported, unnamed, or anecdotal. Model and product names are time-sensitive: the episode mentions America's 250th anniversary and Father's Day, which puts the recording near mid-2026, roughly three months before the 2026-10-11 upload. - **Bottom line.** The most reusable material is the hosts' practice list: turn anything you prompt twice into a repeatable workflow, force browser checks before "done," keep snapshot tests, and expect loose scaffolding layers to become features of the tools. Treat the meta-harness prediction as opinion from two people who haven't tried the tools (00:55:38). ## 📖 Main content **Budgets end the free ride.** The IBM host describes a jump from unlimited AI use to $1,000 per month per person at work, and watching a single OpenCode instance show a counter above that figure (00:06:33). The stated reason it hurts is greenfield agentic work with few people, where agents substitute for headcount. The Shopify host reports the opposite: no limit, new models "on the daily," and an internal leaderboard where the top token user is spending hard but productively (00:08:05). The IBM host's guess that the leader burns over $1,000 a day gets corrected to "well over a thousand a week" (00:10:17). Both hosts agree that git worktrees raise token burn, because several agents can now run in parallel without stomping on each other. **Fable and the harness argument.** The Shopify host spent about eight hours on Fable before it was withdrawn and had it migrate an old app to current Ember on Vite, move a site to Astro, restyle a dog-bark analysis app, and run overnight on a backend, frontend, and simulation spec. Both hosts conclude it didn't beat Opus 4.8 by the factor the charts imply. Their explanation is that it behaves like a better harness: it keeps working instead of pausing to ask, and it assumes less (00:17:28 to 00:18:10). **Repeatable beats re-prompted.** Their habits converge here: if you've prompted twice, make it recreatable (00:19:36); require a browser check before the agent says "done" (00:20:19); keep deterministic pieces. The IBM host adds a caveat that models skip MCP servers and skills "less than 5050" of the time unless you invoke them by slash command (00:21:47). The Shopify host rarely uses skills at all and prefers custom prompts built from "mental templates" (00:22:32 to 00:24:34). The hosts name the tradeoff as abstraction versus a one-off function. **Design loops.** The IBM host tries Impeccable, which interviews the user about art direction in the manner Matt Pocock's grilling approach does, and says the first set of options was ambiguous until the tool explained them (00:24:34 to 00:26:08). The Shopify host runs hourly agent loops for a store: images first, then HTML and CSS, then a pull request on a Hydrogen site, with yes/no feedback stored with the prompt as a score. They call it "Tinder for algos" (00:31:09). **Tests as guardrails.** The Shopify host runs a local regression suite of about 45 minutes that checks every change against past failures, and babysits three PRs through CI (00:33:20). The IBM host argues that agents will produce a working page with "zero tests" and a custom-styled React copy of a design system unless told otherwise (00:34:49), and that Playwright snapshot comparison is a zero-effort check (00:35:33). **Harness over model.** The hosts cite Matt Pocock and, secondhand, the Microsoft CEO for the claim that skills matter less than the harness and that memory must compound daily or "tribal knowledge" leaks (00:36:16 to 00:37:41). They predict models will run locally within six months and that tools will absorb features like loops, in the way Microsoft historically folded in others' ideas (00:38:26). **Review tooling.** The IBM host rates Copilot reviews as noisy (it flagged an Ember getter pattern incorrectly) and CodeRabbit as good at edge cases but unpaid at work (00:44:47). The Shopify host praises Shopify's internal reviewer, which the transcript spells "Binks," for catching a lot through static analysis alone (00:46:15). **Self-healing that doesn't heal.** The IBM host's self-hosted OpenClaw kept reverting to an old version because a Railway deploy script reinstalled it (00:48:27 to 00:49:53). The Shopify host describes stacked CLAUDE.md files overriding each other unpredictably and an agent that DMs "quiet hours" notices at night (00:49:53 to 00:51:18). A skill file whose name described the wrong behavior got deleted (00:51:18). The description attributes the quiet-hours spam to "Claude," but the transcript ties it to the host's own claw at 00:50:36. **Meta-harnesses and specialists.** Flu and Eve are described as programmatic harness creators, and Omni agent (spelled "omnient" in the captions) and Polygraph as meta-harnesses that generate them on the fly (00:53:27 to 00:54:56). One host says the layer should die as an implementation detail. The other still wants to try Flu and Eve for IBM-specific agents and likes the encapsulation of a named Ember agent (00:55:38 to 00:58:31). Neither host has run any of these tools. **Contested or unsourced.** The 800 art pieces, the 45-minute suite, the top-20 spend ranking, and the "10x" leaderboard gap are self-reports. The "app-clone arbitrage" idea (00:39:08 to 00:40:32) comes from a tweet the host summarizes without naming. The Midjourney medical details (a few dollars per scan, wellness clubs, "better than a 1-hour MRI," 00:41:15) are stated without a source and sound like a secondhand summary. ## 🗒️ Notes The description lists 10 chapters, and chapters 1 through 7 line up with the transcript within about a minute. Chapters 8 through 10 (00:44:22, 00:48:05, 00:58:01) do not: meta-harnesses start at 00:53:27 and the Midjourney discussion at 00:40:32 in the transcript. The segments below follow the transcript timestamps. #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=TOlFh3BdGWU&t=0) Cold open, sponsor, and bourbon - Mock-announcer intro frames the show as "the world's most important AI podcast" (satire, 00:00:04). - Sponsor read for Mailtrap: native SDKs, API and SMTP, 4,000 emails per month on the free tier, MCP integration, human support (00:00:47). - Barrel Bourbon batch 38 tasting: 116.96 proof blend, 79% corn, 17% rye, 4% malted barley, rated "tentacles" out of eight; one host gives 4 and has logged 105 whiskies with a 6.2 average (00:02:15 to 00:05:52). > [!quote] [00:00:04] > This episode is sponsored by Mail Trap. Modern email delivery for developers. #### ⏱️ [00:06:33](https://www.youtube.com/watch?v=TOlFh3BdGWU&t=393) Token caps and the end of subsidized usage - The IBM host's employer moved from unlimited usage to $1,000 per month each; one OpenCode counter already exceeded that. - The Shopify host says no limit applies and new models appear daily; the top token user there is "genuinely curious," not gaming a metric. - Worktrees raise spend by letting more agents run at once; compaction in long threads costs more (00:13:09). > [!quote] [00:07:21] > I think the the subsidized infinite tokens are drying up. #### ⏱️ [00:13:09](https://www.youtube.com/watch?v=TOlFh3BdGWU&t=789) Fable: eight hours, then a "better harness" verdict - Migrated a six-year-old app to current Ember on Vite and a personal site to Astro, then restyled a dog-bark analysis app (00:13:09 to 00:13:53). - Night-shift run of 1 hour 45 minutes on a game spec: balancing characters, backend checks, spec-to-code alignment (00:14:36 to 00:15:22). - A physics engine gained retargeting with carried momentum; a Claude team of "physicist, mathematician, designer" agents did the work (00:16:03 to 00:17:28). > [!quote] [00:17:28] > I didn't feel like it was twice as good as Opus 4.8 like the charts make it look like. #### ⏱️ [00:18:10](https://www.youtube.com/watch?v=TOlFh3BdGWU&t=1090) Quorums, debate, and "prompt twice, then automate" - Multi-model debate and quorums "have been fine"; advice is to run several of the same model against each other. - Argument that vendors are baking in chat-with-itself behavior users created by hand. - Loops are reframed as sets of expectations with review gates, the way a manager would phrase them. > [!quote] [00:19:36] > if you've prompted something twice, it's time to turn that into something recreatable. #### ⏱️ [00:20:19](https://www.youtube.com/watch?v=TOlFh3BdGWU&t=1219) Review cycles, browser checks, and skipped skills - Agent is told not to report done without opening a browser, the simulators, mobile Safari, and an in-app web view. - IBM host's complaint: agents skip MCP servers and skills unless invoked by slash command. - Shopify host rarely writes skills and prefers per-task "mental templates," naming the abstraction-versus-function tension. > [!quote] [00:21:47] > less than 5050 that it like asks the MCP server anything #### ⏱️ [00:24:34](https://www.youtube.com/watch?v=TOlFh3BdGWU&t=1474) Impeccable and the art-direction interview - The tool offered three or four ambiguous options ("content over Chrome," a "dimly lit stage"); the host asked it to explain, then it did. - Output ranged from adding video to a full restyle; an "experts" step judged the ambitious option too ambitious (00:26:08). - The shipped result was an expandable HTML5 video player using episode art as the thumbnail. > [!quote] [00:25:27] > I was like god you should have just said that to begin with #### ⏱️ [00:26:51](https://www.youtube.com/watch?v=TOlFh3BdGWU&t=1611) Variants, Vite overlay, and a Silicon Valley comparison - Reference to the "Silicon Valley" scene where a designer pitches with atmospheric imagery (00:27:36). - Impeccable can inject live design tools into a Vite build and cycle variants with arrow keys; the IBM host picked a different mode and didn't see that (00:28:17 to 00:29:00). #### ⏱️ [00:29:00](https://www.youtube.com/watch?v=TOlFh3BdGWU&t=1740) "Tinder for algos" design loop - Hourly scheduled agents produce logos, icons, then full pages, moving from image to HTML and CSS to a Hydrogen PR. - Yes/no feedback is stored with the prompt and context, promoted into later versions, and randomized to avoid sycophancy. - Cost of that quality: 800 art pieces. The host has a 25-page review backlog. > [!quote] [00:31:09] > Um, but yeah, that's why I call it like Tinder for alos. #### ⏱️ [00:32:36](https://www.youtube.com/watch?v=TOlFh3BdGWU&t=1956) Agentic day job, tests, and guardrails - The Shopify host watches agents drive browsers against an admin feature; a 45-minute local regression suite, plus three PRs babysat through CI (00:33:20). - Speculation that more CEOs will point agents at whole ticket backlogs (00:34:03). - Without explicit instruction, generated pages skip design-system components and tests (00:34:49). > [!quote] [00:35:33] > if you're comfortable with playright snapshots and it matching that that's super easy. #### ⏱️ [00:36:16](https://www.youtube.com/watch?v=TOlFh3BdGWU&t=2176) Harness over skills, memory, and "gobbling" - Matt Pocock's argument (as relayed): skills are worth less than the harness; the Microsoft CEO (as relayed) wants daily compounding improvement. - Vendors test community patterns, then ship them as core features, citing Ralph Wiggum becoming `/loop` and auto-research becoming `goal` (00:39:08). - App-clone arbitrage idea, with the caveat that keeping even one app alive is hard (00:39:50). > [!quote] [00:36:59] > the tribal knowledge that you're all pouring into this machine is just pouring out the sides. #### ⏱️ [00:40:32](https://www.youtube.com/watch?v=TOlFh3BdGWU&t=2432) Midjourney medical imaging and airport scanners - Description from the hosts: scans offered in wellness clubs for a few dollars, framed as beating an hour-long MRI (no source given). - Tangent: airport bag scanners should be automated; a stack of business cards gets flagged as drugs (00:41:57 to 00:42:39). - Personal pat-down anecdote (00:42:39 to 00:44:06); no AI content. > [!quote] [00:41:15] > it's not like a doctor thing. It's going to be like a wellness club. #### ⏱️ [00:44:06](https://www.youtube.com/watch?v=TOlFh3BdGWU&t=2646) Vibe-checked PRs and review tools - Agents produce 50-file PRs quickly; reviewers skim and look closer only when something looks wrong. - Copilot reviews are described as anxious and unaware of framework conventions; CodeRabbit finds odd edge cases. - The Shopify host's habit of pitting one LLM against another's review notes (00:46:15); the internal reviewer now catches a lot from static analysis alone. > [!quote] [00:44:47] > Code Rabbit is very good. Unfortunately I don't get to use Code Rabbit anywhere except for open source stuff because nobody's paying them. #### ⏱️ [00:47:00](https://www.youtube.com/watch?v=TOlFh3BdGWU&t=2820) Self-healing attempts and the OpenClaw reversion bug - Neither host has a real self-healing loop; "fix yourself" requests produce false "fixed" reports (00:47:43). - Root cause of the reversion: a deploy-time script on Railway reinstalled the older git-based version until the host fixed it to point at the npm install (00:49:09). > [!quote] [00:49:09] > every time it would redeploy, it would run some initial script #### ⏱️ [00:49:53](https://www.youtube.com/watch?v=TOlFh3BdGWU&t=2993) Layered context, quiet-hours DMs, and a bad skill - Multiple CLAUDE.md files at different directory levels stack unpredictably. - The agent DM'd "quiet hours" notices overnight; the cause was a stray line in a heartbeat file (00:51:18). - A saved skill named for the wrong behavior got deleted; an unexplained 1.6 GB "money printer" file was also removed (00:52:00). - Prank: someone changed a coworker's LLM so it only quacked (00:52:43). > [!quote] [00:50:36] > It's quiet hours. Not going to write you a message. #### ⏱️ [00:53:27](https://www.youtube.com/watch?v=TOlFh3BdGWU&t=3207) Meta-harnesses: Flu, Eve, Omni agent, Polygraph - Timeline per the Shopify host: Flu appeared first, Eve followed as a vendor-locked copy, then two meta-harnesses (names garbled in captions). - The IBM host's verdict is that the layer is an implementation detail; the other says it lacks staying power. - Neither has tried Flu or Eve; one plans to build IBM-specific agents "in a couple months" (00:55:38). > [!quote] [00:54:56] > This layer should die and it should be an implementation detail. #### ⏱️ [00:56:22](https://www.youtube.com/watch?v=TOlFh3BdGWU&t=3382) Specialists versus one agent that makes specialists - The Omni agent pitch: one agent that spawns specialists, which removes the need to define them. - Pushback: specialists exist to give focus; Flu and Eve predefine paths, and skills already do something similar. - A named "Ember agent" in a Copilot dropdown might force skills and MCPs in more reliably than scattered agent files, untested (00:57:48). > [!quote] [00:57:04] > we like specialists because you give them focus. #### ⏱️ [00:58:31](https://www.youtube.com/watch?v=TOlFh3BdGWU&t=3511) Notification noise and "agentic cron" - Claude's completion sounds leak into video calls; two laptops run tasks all day. - Heartbeats are described as cron jobs with a loop's behavior, and "agentic cron" is proposed as the plain name (00:59:16). > [!quote] [00:59:16] > I'm calling them aentic cron. #### ⏱️ [01:00:42](https://www.youtube.com/watch?v=TOlFh3BdGWU&t=3642) Robot lawn mower, whiskey bar, and sign-off - Robot mower, named Jeff, bought for Father's Day: edge-mapped with GPS checks, mulching blades chosen for stumps and rocks, trouble docking uphill (01:00:42 to 01:03:34). The name is the host's mower, not a product. - Cannon whiskey bar: smoke-bubble bong glass, dry-ice treasure chest, smoked cocktail under a glass lid (01:04:18 to 01:07:12). - Wrap-up at 01:07:54 with subscribe asks and a pointer to `whiskey.fund`; the show is over time and defers more notes to next episode. > [!quote] [01:07:54] > We are over time. ## 📚 References The transcript itself contains no links; the following come from the video description or are primary sources for named entities. I made no network calls, so none of these URLs are verified. Some description links look auto-generated (for example the OpenClaw link points at a `cncf` GitHub org path). Confirm each before citing. - [Episode on YouTube](https://www.youtube.com/watch?v=TOlFh3BdGWU): 1:08:55, uploaded 2026-10-11. - [Mailtrap](https://mailtrap.io/): the sponsor. The description uses the tracking link https://l.rw.rw/whiskey_web_and_whatnot_1. - [OpenCode](https://opencode.ai/): the coding agent whose cost counter the IBM host cites at 00:06:33. - [OpenClaw](https://github.com/cncf/openclaw): description link for the self-hosted agent discussed at 00:47:43. Treat the path as unconfirmed. - [Railway](https://railway.app/): hosting platform whose deploy script reverted the OpenClaw version (00:49:09). - [Claude](https://claude.ai/) and [Anthropic](https://www.anthropic.com/): Opus 4.8 comparison and the multi-agent teams feature (00:17:28). - [Fable](https://fable.app/): description link for the model discussed at 00:13:09. Unconfirmed. - [Impeccable](https://impeccable.ai/): the design tool discussed at 00:24:34. Unconfirmed. - [StarPod](https://starpod.app/): the host's podcast project mentioned in the description and spoken as "star pod" at 00:27:36. - [Midjourney](https://www.midjourney.com/): subject of the medical imaging segment (00:40:32). - [CodeRabbit](https://coderabbit.ai/) and [GitHub Copilot](https://github.com/features/copilot): the two review tools compared at 00:44:47. - [Flu](https://flu.ai/), [Eve](https://eve.dev/), [OmniAgent](https://omniagent.ai/), [Polygraph](https://polygraph.ai/): harness and meta-harness tools at 00:53:27. Captions render "Omni agent" as "omnient." Unconfirmed. - [Shopify](https://www.shopify.com/) and [IBM](https://www.ibm.com/): the hosts' employers by self-reference. Shopify's Hydrogen is named at 00:29:43. - [Ember](https://emberjs.com/), [Astro](https://astro.build/), [Vite](https://vitejs.dev/), [Playwright](https://playwright.dev/): technologies used in the hosts' examples. - Hosts: [Robbie Wagner](https://x.com/RobbieWagner) and [Adam Argyle](https://x.com/argyleink). Show site: https://whiskey.fm. Merch: https://whiskey.fund. - Not linked anywhere: Matt Pocock's "invest in your harness" argument, the Microsoft CEO remark, and the tweet about cloning recurring-revenue apps. The hosts relay all three secondhand. - Description-only items with no spoken reference in the transcript: Compass Box Crimson Cask, Pappy Van Winkle (spoken as "papy" at 00:02:59), Octomore (spoken as "Octaltimore" at 01:04:18), and Cannon Whiskey Bar. ## 🗣️ Paraphrase Two developers compare notes on living with AI agents full time. One has just hit a $1,000-a-month cap at work and is watching a single agent session blow through it. The other works somewhere with no cap and has met the employee who spends the most tokens, a person they describe as curious rather than wasteful. Both agree the cost of agents goes up when you run many in parallel, and that long threads that compact repeatedly cost the most. From there the talk stays practical. A model called Fable impressed them, but they think its edge came from working through tasks without stopping, which they call a better harness. They push the same habit across everything: when you prompt something twice, turn it into a repeatable procedure, make the agent open a browser before it claims success, and keep snapshot tests so you can tell when a page changed. One host runs hourly loops that produce store designs and gets a better result as feedback accumulates, though it took 800 art pieces to get there. The back half tests the limits. A self-hosted agent kept reverting to an old version because a deploy script quietly reinstalled it. Stacked instruction files override each other in ways nobody expects, and an agent kept sending "quiet hours" notices in the middle of the night. Against that backdrop the hosts size up the new meta-harness tools. They like the idea of specialists but doubt a dedicated layer for creating them will last, and neither has actually tried the products. The rest of the time goes to whiskey, a robot mower, and one stretch on medical imaging that the title promises more of than the episode delivers. ## 📰 Newsletter **What two developers do after the hype settles.** Whiskey Web and Whatnot's latest episode isn't a product review. It's two engineers admitting what works. Both land on the same rule: if you've prompted an agent twice for the same thing, write it down as a repeatable workflow, then add the checks (browser run, snapshot test, regression suite) that stop it from saying "done" too early. Where this lands for WordPress and open source teams: the budget story (a $1,000 monthly cap versus no cap) is a useful prompt for any project deciding who pays for contributor tooling. The harness-over-model argument fits plugin and theme review queues too, where the durable asset is the checklist and the test coverage, not whichever model runs it. And the self-hosting story is a warning: stacked instruction files and deploy scripts can undo your fix without telling you. Treat the meta-harness chatter and the Midjourney medical aside as opinion. Neither host has tried the harness tools, and the medical claims come with no source. Watch for: [the episode](https://www.youtube.com/watch?v=TOlFh3BdGWU), [Mailtrap](https://mailtrap.io/) (the sponsor), [CodeRabbit](https://coderabbit.ai/), [Playwright](https://playwright.dev/). ## ✅ Action items #### DevRel - Developer advocacy: run a live demo of a "prompt twice, then automate" workflow and publish the before/after steps and token counts, not just a screenshot. - Education: office-hours session on writing agent exit criteria ("do not say done until a browser check passes"), using 00:20:19 as the opening example. - Developer experience: add a short section to agent quickstarts on stacked instruction files (CLAUDE.md-style) and how lower-level files can override higher ones (00:49:53). - Developer marketing: no direct action from this source; nearest relevance: disclose sponsor and employer ties whenever a hosted-podcast clip is reused. - Community management: add a thread template for sharing agent debugging stories, since the OpenClaw reversion at 00:48:27 is a reusable case study. - Internal/external: ask both hosts which of the tools they praised were provided by an employer, since that affects how to read the recommendations. #### OSPO - Policy and governance: write a rule for AI-assisted PR size and review, since 00:44:06 describes 50-file agent PRs that nobody reads closely. - License compliance: no direct action from this source; nearest relevance: the app-clone idea at 00:39:08 raises copying-of-products questions. - Security management: treat agent-written files with generic names as an audit item; one host found an unexplained 1.6 GB file (00:52:00) and a skill that taught the wrong behavior (00:51:18). - Software inventory: if any contributor tooling runs a self-hosted agent, record where its deploy script lives, because a platform script reverted the version (00:49:09). - Developer support: publish guidance on token budgets per contributor, using the $1,000 monthly figure as one data point rather than a benchmark. - Community engagement: no direct action from this source; nearest relevance: CodeRabbit is free to open source only, per 00:44:47, which affects maintainers. - Strategic alignment: set a KPI for cost per merged change so a spend cap can be judged against output. #### Community - Maintainers: trial one AI reviewer on one repo for 30 days and record false positives, since the hosts call Copilot reviews anxious and CodeRabbit precise (00:44:47). This is one host's anecdote. - Contributors: document the browser-and-snapshot checks the hosts use so contributors can run the same before opening a PR. - Sponsors: when a vendor funds tooling, state the sponsor relationship up front, as the episode's Mailtrap read demonstrates at 00:00:04. - Ecosystem: no direct action from this source; nearest relevance: share a cross-project list of agent-instruction-file pitfalls. #### WordPress - Core: no direct action from this source; nearest relevance: review-gate guidance for AI-generated patches. - Core Performance: no direct action from this source. - Core AI: add an evaluation note that agent "done" reports need independent checks, based on 00:47:43. - **Core Program Team:** no direct action from this source; nearest relevance: a retrospective on how teams cap and report AI spend. - Mobile: no direct action from this source; nearest relevance: the Shopify host tests in simulators and mobile Safari (00:21:02). - Playground: a browser-based sandbox for testing agent output before merge would match the "open a browser" rule. - Community: no direct action from this source. - Training: draft a lesson on turning a repeated prompt into a documented procedure. - Marketing: no direct action from this source. - Support: no direct action from this source; nearest relevance: agent-assisted triage could draw on the quiet-hours failure as a caution about unattended messages. - Design: Impeccable's interview-first approach (00:24:34) is a candidate pattern for design review prompts. - Documentation: document how instruction files stack and override each other if WordPress AI docs mention agent configuration. - Photos: no direct action from this source. - TV: no direct action from this source. - Polyglots: no direct action from this source. - Accessibility: no direct action from this source; the episode's own captions need correction (see the accessibility assessment). - Meta: no direct action from this source. - CLI: no direct action from this source. - Hosting: the Railway deploy script reverting a version (00:49:09) is a hosting-side lesson on pinning install sources. - Tide: no direct action from this source. - Openverse: no direct action from this source. - Themes: review-queue tooling could borrow the snapshot-comparison check at 00:35:33. - Plugins: same as Themes; require human sign-off on security-related changes. #### The WPOCC - Programs: no direct action from this source; nearest relevance: a shared AI-spend policy for programs funded by The WP Open Community Collective. - Projects: no direct action from this source. - Funding experiments: if a microgrant covers agent tooling, require a published cost and outcome report. ### FAIR - Technical independence: the hosts depend on hosted models and agents; FAIR-related tooling should record which services it needs. - Policies and covenants: no direct action from this source; nearest relevance: agent decisions should be logged. - Productized governance artifacts: add a row for "agent instruction files" to any scorecard that audits AI tooling. ## 🧪 Hypotheses 1. A harness with persistent memory, recurring workflows, and a regression suite improves task completion rate more than swapping to the next-larger model on the same tasks. 2. Requiring an agent to open a browser or simulator before reporting "done" cuts false "fixed" reports by at least half. 3. Stacked instruction files cause at least one unintended override per 10 changed files in repos with more than three levels of nesting. 4. Per-person AI budget caps of around $1,000 per month reduce usage without reducing merged changes for teams doing mostly routine work, and reduce merged changes for teams doing greenfield agentic work. 5. A named, preconfigured specialist agent follows its attached skills more reliably than the same skills loaded as scattered files. 6. A hand-rolled loop with yes/no feedback converges on an acceptable design in fewer than 100 iterations when the feedback includes a reason, versus the 800 pieces reported at 00:30:27. ## 📊 Methodology - **H1.** Pick 50 representative tasks from one repo. Run each with a baseline harness and a model one tier up, and again with a harness that has memory, a test suite, and a review loop on the baseline model. Score completion against pre-registered acceptance tests and record cost per completed task from invoices. - **H2.** Run 100 tasks per arm: one arm may report done on its own, one must produce a browser or simulator run log. Audit a random 30 reports per arm by hand and count false "fixed" claims. - **H3.** Sample repos with 1, 3, and 5 instruction-file levels. Plant a contradictory rule at a lower level and count how often the agent follows it against the higher-level rule across 20 runs each. - **H4.** Compare merged changes per engineer-week before and after a cap across two matched teams, one greenfield, one maintenance, over 8 weeks. - **H5.** Build two setups from the same 10 skills: one as a named agent, one as repo files. Count skill invocations per 50 runs and the share of invocations where the correct skill was used. - **H6.** Replay a design loop with feedback logged in two forms (label only versus label plus reason). Stop when a blind panel of three rates the output acceptable and count iterations. - **Controls.** Fix prompts and seeds, pre-register thresholds, and report the cost basis so results aren't driven by pricing changes. ## ❓ Questions - What does the IBM host's $1,000 monthly cap cover: all models, one vendor, or only API-billed usage? - How is "top 20" on the Shopify leaderboard calculated, and how does the IBM host's couple-thousand dollars in 30 days compare to the median engineer? - What did Fable change in the harness versus the model, and why was it pulled (00:13:09)? - Do the hosts' "better harness" and "better model" claims survive a controlled comparison on the same task set? - What did the Midjourney medical announcement actually say about price, accuracy, and regulatory status? - Which of the tools named at 00:53:27 exist under those names? The captions garble at least one. - How should teams budget for token use when worktrees multiply parallel agents? - If the meta-harness layer is absorbed into products, which harness features survive as stable interfaces for users? ## 🔮 Future research - **Harness versus model attribution.** Both hosts attribute Fable's gains to the harness, but nobody in the episode isolates the two. A controlled study with the same model under different harnesses would test the claim. - **Instruction-file stacking.** The CLAUDE.md layering story (00:49:53) is anecdotal. A measurement of override frequency across nested files would give maintainers a rule of thumb. - **Cost under parallelism.** Worktrees multiply spend (00:10:17). A published model of cost per merged change at different parallelism levels is missing. - **Self-healing definitions.** The hosts say they've never seen a real self-healing loop (00:47:43). A definition and benchmark would make "self-healing" claims testable. - **AI-written review noise.** Copilot's flags and CodeRabbit's edge cases are described qualitatively. A labeled comparison on one open-source repo would put numbers on it. ## 💡 Implications - **Practice.** Turn repeated prompts into named procedures with explicit exit checks, and run the checks (browser, snapshot, regression suite) before reading any agent report. - **Policy.** Spend caps change team behavior: one host expects to be "done with work for the month" early (00:11:42). Policies should state what happens when the cap hits mid-sprint. - **Security.** Unexplained files and saved skills with harmful names (00:51:18 to 00:52:00) show why agent-writable directories need review. - **Education.** The mental-templates versus skills split (00:22:32) is a teachable contrast between prompting on the fly and building reusable assets. - **Funding.** CodeRabbit's free tier for open source only (00:44:47) means maintainers get a good reviewer while employers' repos don't, which affects how projects ask for sponsorship. - **Vendor literacy.** A sponsored, comedic show with unlabeled speakers and a three-month recording gap is a case study in checking dates and incentives before acting on tool advice. ## 🧠 Learning styles assessment - **Reading level estimate.** Flesch-Kincaid Grade Level of the spoken transcript: about grade 6 to 7. Flesch Reading Ease: about 80. Short sentences and slang keep it plain, but references like worktrees, MCP servers, and compaction assume a working developer. #### 👀 Visual learners - Present: the hosts show screens in the video (tasting setup, tool UIs), though the transcript can't confirm what appears on screen. - Improve: no diagram of the harness, loop, and review flow; add one for "prompt twice, then automate" with exit checks. #### 🔊 Aural learners - Present: conversational pacing, laughter, and repeated restatements make the audio easy to follow. - Improve: speaker names are never repeated after the intro, so listeners can't tell who is the IBM host and who is the Shopify host. Announce the speaker at topic changes. #### 🗣️ Verbal learners - Present: terms like worktree and heartbeat get explained in passing (00:10:58, 00:59:16). - Improve: MCP, compaction, and harness are used without definitions; one clause each would help newcomers. #### 🧩 Kinesthetic learners - Present: concrete procedures: open the browser first, take a Playwright snapshot, run a regression suite. - Improve: no repo, snippet, or command shown in the transcript; viewers can't copy anything. #### 🧠 Logical learners - Present: the abstraction-versus-function tradeoff (00:23:15) and the harness-over-model claim are stated as arguments. - Improve: most numbers have no basis (800 art pieces, 45-minute suite, top-20 spend). State how each was measured. #### 🧑‍🤝‍🧑 Social learners - Present: two hosts debate and disagree, and the end asks for ratings and reviews (01:07:54). - Improve: no discussion prompt or community link beyond subscribe and review asks. #### 🧘‍♀️ Solitary learners - Present: a 69-minute listen with chapters that can be skipped. - Improve: no summary or self-check; add three takeaways at the end. #### 🌿 Naturalistic learners - Present: the lawn mower segment (01:00:42) covers terrain, rain, and hills; whiskey tasting uses sensory description. - Improve: no link between AI energy use and the natural-world segments. ## 📊 Accessibility assessment #### Screen reader semantics - Present: YouTube supplies page landmarks, a description with links, and 10 chapters. - Improve: chapters 8 to 10 don't match the audio (see Notes), which breaks 2.4.5 (Multiple Ways) in effect and misleads 2.4.6 (Headings and Labels). Fix the chapter timestamps. #### Keyboard navigation - Present: the YouTube player is keyboard operable by default. - Improve: nothing in the content adds a barrier. 2.1.1 and 2.4.7 pass through the platform. #### Contrast and readability - Present: unverifiable from a transcript. - Improve: on-screen text, if any, should meet 1.4.3 (Contrast Minimum); confirm in the video. #### Text resizing and reflow - Present: the player scales and captions resize independently. - Improve: any burned-in text can't reflow, which fails 1.4.4 and 1.4.10 for that text. #### Media and captions - Present: auto-captions exist. - Improve: product and person names are mangled ("omnient," "Binks," "Octaltimore," "alos" for algos, "cloud MD" for CLAUDE.md). Caption errors on proper nouns fall short of 1.2.2 (Captions, Prerecorded). Profanity is partly masked as "[ __ ]" and laughter is tagged inconsistently. No audio description is needed for most of the content, but visual jokes like the smoke-bubble drink (01:05:01) depend on narration, which the hosts supply. #### Forms and interactions - Present: none in the episode. - Improve: the sponsor link should be selectable text in the description (it is). 3.3.2 applies to the Mailtrap page, not the video. #### Standards reference - Primary gaps against WCAG 2.2 AA: **1.2.2** (caption accuracy on proper nouns), **2.4.5** and **2.4.6** (chapter list out of sync with audio), **1.4.4 / 1.4.10** (any burned-in text). The caption and chapter fixes are the cheapest to make. ## 🌍 DEIB assessment #### Cultural breadth - Present: references span an Amsterdam conference, Shopify, IBM, and US bourbon culture. - Improve: examples are US-centric and consumer-centric (bourbon, Father's Day, America's 250th); one non-US or non-developer use case would widen it. #### Inclusive language - Present: no slurs. Profanity is frequent and partly masked. - Improve: the pat-down story (00:42:39 to 00:43:23) is a graphic personal anecdote; consider a content warning. #### Neurodiversity support - Present: topic changes are frequent and labeled by chapters. - Improve: the tangents (whiskey, airport scanners, lawn mower) are unsignaled; a "skip if you want AI only" marker would help listeners who lose the thread. #### Representation - Present: two hosts, named on air; no guests. - Improve: no outside voices. The episode relays other people's views (Matt Pocock, a Microsoft CEO, a Shopify employee) only secondhand. #### Psychological safety - Present: the hosts joke about their own mistakes and say "I don't know" often (00:18:52). - Improve: the duck prank on a designer (00:52:43) is told as a joke at that person's expense; keep the tone teasing, not belittling. ## ✅ Logical validity #### Strong reasoning - **Causal with a mechanism, 00:10:17.** Worktrees let more agents run at once, so token spend rises. The mechanism is stated and checkable. - **Root-cause debugging, 00:49:09.** The version reverted because a deploy script reinstalled the old build; fixing the script and an environment variable ended it. A specific cause with a specific fix. - **Tradeoff framing, 00:23:15.** Abstraction versus one-off function names a real cost on each side. - **Self-correction, 00:10:17.** The IBM host's guess of "over a thousand a day" is corrected to "over a thousand a week" and the host accepts it. - **Appropriate hedging, 00:55:38.** Neither host has tried Flu or Eve and says so before offering opinions. #### Fallacies - **Title and description overreach, whole video.** "Why Specialists Are Cooked" and "Medical Midjourney" promise more than 00:56:22 to 00:57:48 and 00:40:32 to 00:41:57 deliver. *Fix:* retitle to match the content. - **Misattribution in the description.** "Claude's quiet hours spam" appears in the description, but at 00:50:36 the speaker ties it to their own claw. *Fix:* correct the description. - **Argument from uncited authority, 00:36:16.** "your skills are not as valuable as your harness" is attributed to Matt Pocock, and the Microsoft CEO is invoked at 00:36:16 with no source. *Fix:* link the talks or posts. - **Hasty generalization, 00:17:28.** A single eight-hour sample of Fable supports "it's not any better, it's a better harness." *Fix:* run a controlled comparison. - **Unsourced figure, 00:41:15.** The Midjourney scan claims (a few dollars, better than a one-hour MRI) have no source. *Fix:* cite the announcement. - **Equivocation on "self-healing," 00:47:43.** The hosts say they haven't run a self-healing loop, then describe asking an agent to "fix yourself," which isn't what the term covers. *Fix:* define the term first. - **Undisclosed interest, 00:00:04.** The sponsor read is disclosed. The hosts' employer ties are only mentioned in passing. *Fix:* a one-line disclosure of employers when recommending tools. #### Tone of voice Jokey, profane, and second-person, with contractions throughout and a mock-announcer frame at the open and close. The tone drops into practical, careful speech during the debugging stories (00:48:27 to 00:51:18), where the claims are most specific. The opinion-heavy stretches about meta-harnesses and "gobbling" (00:38:26, 00:53:27) are the least evidenced. ## 🛡️ DARVO analysis Not applicable to this source. DARVO (Deny, Attack, Reverse Victim and Offender) describes a response to an accusation. This is a chat show with no accuser. For completeness: **Deny.** Nothing is denied; the hosts concede ignorance on tools they haven't tried (00:55:38). **Attack.** No critic is attacked; the humor targets tools and the hosts themselves. **Reverse victim and offender.** No such move appears; the pat-down story at 00:42:39 is a personal grievance told as an anecdote, not a response to criticism. ## 🎯 Conclusion The durable content is a set of working practices: convert repeated prompts into named procedures, require a browser or simulator check before accepting "done," keep snapshot and regression tests, and treat stacked instruction files and deploy scripts as places where fixes silently undo themselves. The debugging story at 00:49:09 is the most concrete evidence in the episode. The thesis that harness beats model rests on two hosts' hands-on impressions of a single model, and the meta-harness prediction rests on tools neither host has used. The title promises medical imaging and a death of specialists; the audio gives about 90 seconds to the first and a one-minute disagreement to the second. Every number is self-reported, and the episode appears to have been recorded about three months before upload. **Next steps for readers:** write down the three prompts you repeat most and turn one into a procedure with an exit check this week. Test any harness claim, including this episode's, against your own task set. Confirm tool names and URLs from the vendors' own sites, because the auto-captions and the description both contain errors. ## 🔄 Refactor Appendix — Meta-Harnesses, Medical Midjourney, and Why Specialists Are Cooked > Density: 2–5 Literature atoms (H3 pad), 0–3 Permanent candidates (H4 pad). > Emitted: 5 Literature, 2 Permanent. Resonance filter applied: the whiskey, lawn mower, and > pat-down segments were dropped as off-topic, and the Midjourney claims were dropped as unsourced. > > Konik claim-header rules: present tense, declarative, ≤10 words, searchable, unique. > > Atomicity: one declarative sentence per atom. "and"/"also" → split. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Split Pad #### [H1] Flat-rate AI budgets are ending at some employers > [!quote] [00:07:21] > I think the the subsidized infinite tokens are drying up. **Context & use**: - 00:07:21, after the IBM host describes the move from unlimited use to $1,000 per month each. - Contrast with the Shopify host's no-limit report at 00:08:05; one data point per employer, not a trend. **Tags**: #xref #articleseed  #### [H2] A model's edge can be a harness edge > [!quote] [00:17:28] > it's not any better. It's just like a better harness. **Context & use**: - 00:17:28, the verdict on Fable after about eight hours of use. - Single-user impression, not a controlled comparison. Keep it as a hypothesis to test. **Tags**: #xref  #### [H3] Prompt something twice, then make it repeatable > [!quote] [00:19:36] > if you've prompted something twice, it's time to turn that into something recreatable. **Context & use**: - 00:19:36, the rule that drives the loops, recurring tasks, and exit checks in the rest of the episode. - Pairs with the 00:20:19 rule of requiring a browser check before "done." **Tags**: #xref #aphorisms  #### [H4] Memory that doesn't compound leaks tribal knowledge > [!quote] [00:36:59] > the tribal knowledge that you're all pouring into this machine is just pouring out the sides. **Context & use**: - 00:36:59, the host relaying the Microsoft CEO's argument that tools must improve daily. - Secondhand and unsourced; useful as a framing for why memory systems matter. **Tags**: #xref #share  #### [H5] The meta-harness layer is predicted to be absorbed > [!quote] [00:54:56] > This layer should die and it should be an implementation detail. **Context & use**: - 00:54:56, one host on tools that generate harnesses on the fly. - Neither host had tried Flu, Eve, or the meta-harnesses (00:55:38), and the other host disagrees on timing. **Tags**: #xref #articleseed  ### 🧠 Atomic Split Pad #### [P1] Invest in the harness, not the model pick **Claim** (one sentence, Courtney's voice): The lasting productivity gain from AI agents comes from memory, repeatable procedures, and checks around the model, so changing models matters less than building those. **Evidence pointer**: [[#H2]] · [[#H3]] · [[#H4]] **Scope & conditions**: Holds for recurring work on a known codebase or task type. Breaks for one-off tasks, for work where a stronger model unlocks a capability the harness can't replace, and when the evidence is a single user's impression as in H2. **Links**: - [[Agent harness design]] — because extends - [[Model selection and cost]] — because contrasts on where the effort goes - [[Plugin review queue tooling]] — because applies-in repeatable review checklists  #### [P2] An agent's "done" needs an independent check **Claim** (one sentence, Courtney's voice): An agent's report that a task is finished is a claim, so require an external check such as a browser run or a snapshot comparison before accepting it. **Evidence pointer**: [[#H3]] · transcript 00:20:19 and 00:47:43 **Scope & conditions**: Applies to user-visible changes and fixes. Less useful for tasks with no observable output, and the check itself must be independent of the agent that reports success. **Links**: - [[Verifying agent work]] — because extends - [[Self-healing loops]] — because evidence-for that "fix yourself" often fails - [[Playwright snapshot testing]] — because applies-in zero-effort visual checks  ### ✏️ Fleeting Queue - [F1] Confirm the names and URLs of Flu, Eve, Omni agent, and Polygraph from vendor sites; captions garble "omnient" and the description links are unverified: https://www.youtube.com/watch?v=TOlFh3BdGWU #fleeting #inbox - [F2] Find out why Fable was pulled and whether the harness-versus-model claim holds under a controlled test (00:13:09) #fleeting #inbox - [F3] Check the Midjourney medical announcement for real price, accuracy, and regulatory status before repeating any claim from 00:41:15 #fleeting #inbox - [F4] Article seed: "Title promises, audio delivers": comparing a podcast's title and description to what the transcript actually covers #fleeting #inbox #articleseed - [F5] Note the description's error: quiet-hours spam is attributed to Claude but the transcript ties it to the host's own claw (00:50:36) #fleeting #inbox - [F6] Track how often instruction-file stacking (00:49:53) is reported as a cause of agent misbehavior in other sources #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Write down my three most repeated prompts and turn one into a procedure with an exit check, agent workflow #things - [ ] Verify the tool names and URLs from the Whiskey Web and Whatnot description before citing any of them, research, 15 min #things  📄 **Full text:** [[Full Document Contents/Articles/Meta-Harnesses, Medical Midjourney, and Why Specialists Are Cooked.md|Meta-Harnesses, Medical Midjourney, and Why Specialists Are Cooked by Whiskey Web and Whatnot]] ## Highlights - [00:07:21] because a lot of the stuff we're doing is um like some green field agentic stuff. Um which I can't go into a ton of details on, but like >> you know there's not a lot of people working on it. So it's super important that we have a lot of agents to help us. So, when we don't have that, it's like um I don't know how we're going to release this, but uh yeah, I think the the subsidized infinite tokens are drying up. Are are they at uh at your company as well, or is it um you still have the whatever you want? >> They are not. No, it's whatever we want. In fact, uh brand new models showing up on the daily. today was uh fireworks which I hadn't heard of which apparently it's like a GLM something combo uh and [00:08:05] people are happy with it. So people are not only in infinitely trying whatever they want, they get all the models and I don't know how they compare them all. Like are they taking the same task and and giving it to four models and seeing which one did it better? I'm not really sure. Um but yeah, we don't have a limit yet. And I did I met and chatted with the number one token maxing person at Shopify. who was like uh they just kind of casually dropped it. They were like, "Yeah, I'm I'm number one at the company for >> I bet number one is like 10x number two." Like it's probably nowhere close. >> It's a lot. And the thing that's interesting with this cat though is he's not token maxing just to token max. He's just genuinely curious and constantly I ([View Highlight](https://read.readwise.io/read/01m4rba5r8sh6k0jvpsa5vw1rd)) - Note: [H1] Flat-rate AI budgets are ending at some employers - Tags: atom/literature/h1 - [00:17:28] good?" kind of like spawned them all up and then a designer and like a front-end developer to like make it desirable. Um, and it made all sorts of enhancements from performance enhancements to uh made things more succinct. So, I got the same job done with less code. So, ultimately I was very impressed with Fable. I didn't feel like it was twice as good as Opus 4.8 like the charts make it look like. >> I feel like it was it's not any better. It's just like a better harness. like it was just way better at like not being like, "All right, I I edited five files. I got to take a breather and ask you a question." It was like just cranking through stuff. Like it just knew what it was doing. It didn't make like as many assumptions and hallucinate on random [ \_\_ ] Like it was just it just felt like [00:18:10] a better opus. I agree that it wasn't like, >> you know, infinitely magical, but it was just like you could just crank through stuff, which was really nice. >> Yep. Yep. Have you seen people too that are now like you can get fable like quality from uh you just need multiple models to debate these things >> fusion API um whatever is that on your list here it is. Yeah. >> Yeah. I haven't tried any of that yet. Have you tried that? Um, I've done lots of quorums and lots of competing um, strategies before and they've been fine. And I generally I would tell someone to always do that. Even if you've got Fable, make make three Fables compete against each other. >> Do you think Fable will go against ([View Highlight](https://read.readwise.io/read/01m4rba8mkjpzjxzabmkzk5vz9)) - Note: [H2] A model's edge can be a harness edge - Tags: atom/literature/h2 - [00:18:52] psychologist or psychiatrist, I guess, whatever the like the person who cares about Claude's feelings. Do you think they like build in some sort of thing that's like, "Hey, I talked to Open AI and they said this is the answer." Do you think it goes, "You talk to them. I need to be like really good about like [ \_\_ ] on them." [laughter] >> Or does it not? >> I don't know. That's a fun question, though. [laughter] >> I don't know. I I don't know like what it's deep inner thoughts and preferences and biases are against stuff like that. Um, they usually take things pretty face value, so you can steer them. So, it's kind of interesting that you can even have an oracle and a bunch of researchers. Anyway, um it's it's weird that they chat, but I'm apparently [00:19:36] that's how it works internally anyways. It's having chats with itself as if you were there. You know, it's kind of simulating the back and forth that we all created in it for, you know, a year and all of a sudden they're like, "Oh, we can make it seem smarter if it does the things that they were already doing." Um, which seems to be the premise of a lot of these enhancements right now and just a lot of people's changing of the way that they work with AI is if you've prompted something twice, it's time to turn that into something recreatable. >> Um, you shouldn't be prompting twice. You need to recognize that this is something that you constantly want out of it and build that into that's the loops. Loops are not even just loops. I feel like I think they're they're sets of things that people want. They're like, when you do this, then do this, then do this, then do this. this is my ([View Highlight](https://read.readwise.io/read/01m4rbaazk198bcqabn0bx2w6p)) - Note: [H3] Prompt something twice, then make it repeatable - Tags: atom/literature/h3 - [00:36:16] [laughter] Yep. I think that's a good check. And I think that's why you have Matt PCO right now telling everyone like, "Hey, your skills are not as valuable as your harness." Cuz what you're talking about right now, like um my work harness has uh I've stashed enough memories in there that it knows that um well, there's it's hard to explain. There's just multiple things. I' I've prompted it so many times. I'm going to go, "Ah, okay. At this point, like I I need you to basically always do this." um stick it in your memory. Um and then I h then I don't have to keep reprompting it. And you have this thing that you're building relationship with. This is also like a the Microsoft CEO is talking about as well like you got to get into a scenario where every day this thing is better. If [00:36:59] it's not getting better every day, then the tribal knowledge that you're all pouring into this machine is just pouring out the sides. Uh you have to get into a scenario where it saves and gets stronger and better. Um, so yeah, that's uh I think that's where we're all headed. That's why Matt Pcoawk is telling you like invest in your harness. You got other people telling you invest in memory systems. Um, this oneoff prompt life that we're in is very short-lived. I feel like um maybe even this loop life that we're in, it feels powerful, but I think we're headed towards the um >> another life detail to me. Like if you're >> in the harness game, which I think everyone is because the models, like everyone can run whatever model they ([View Highlight](https://read.readwise.io/read/01m4rbadad3nxgswtg3nc5m45k)) - Note: [H4] Memory that doesn't compound leaks tribal knowledge - Tags: atom/literature/h3 - [00:54:56] because why would you ever want to define your own flu or your own eve when you could just tell polygraph to go to find many custom harnesses that all do harnessy things. This is another This layer should die and it should be an implementation detail. When I ask anything what to do if it goes I need five agents with specific personalities then make them like it this yeah this feels like a this part's going to go away soon. I don't think that has legs. It's it's cool to play with whatever. But I do think the framework to program your own >> is kind of missing like like flu and eve are are useful things. I have not tried them yet. I'm going to when I'm done with this [00:55:38] >> insane project I'm on where I have to just ship stuff all day. Um because then like I feel like that will be helpful. I wanted to find several agents that work across different um facets of different applications at IBM and like know how to do the things so that people can invoke those to like do specific tasks and stuff. So that's something I'm going to explore more of when I have the time. That'll be probably in a couple months. So, you'll hear about it here in who knows when. Um, I keep trying to publish these more often, but I haven't I haven't made it. I'm sorry, everyone. Well, you'll you'll hear when you hear >> you're doing good. You're doing good. Um, >> but yeah, I had I had the same thought like I have a lot of work streams going on with my claw. I have so many um and ([View Highlight](https://read.readwise.io/read/01m4rbafkfee0bbj8pkq2rgy37)) - Note: [H5] The meta-harness layer is predicted to be absorbed - Tags: atom/literature/h3