# Joey Kudish — how I use AI these days (thread) > [!question] Lingering Questions > - Would the month-long-thread persistence claim survive an actual benchmark, or is it selection bias from threads that happened to go well? (author explicitly admits it's unbenchmarked) > - What exactly is the "self-learning" mechanism — durable memory, curated re-injection, or retrieval over the decision log? Undescribed. > - How much of the claimed quality is attributable to the workflow versus the author's prior expertise as an experienced consultant with custom tooling? > - Does "never sloppy anymore" survive contact with an adversarial reviewer who didn't write the plan? > [!tip] Article Ideas > - "Treat software as drafts" tension: rewriting is cheap for the author and potentially expensive for dependents relying on software they maintain — worth a piece on where the rewrite-as-default thesis breaks. > - "Raise your ambition" as an answer to layoff fear risks individualizing a structural labor-market problem — worth a careful, sympathetic-but-critical piece. > [!link] Cross Reference > - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — plan-stage adversarial review, multi-model workflows > - [[3. Resources/Permanent Notes/The Maker Should Not Grade Its Own Homework]] — already cites this source (added 2026-08-01) for the "review the model's plan, not just its code" upstream-review move > - [[3. Resources/Permanent Notes/Intent Written on the Outside Compounds While Re-Derived Intent Is Debt]] — closely related to this source's persistent-state claim (decision logs surviving compactions) > [!quote] Vocabulary > - **compaction** — context-window summarization event in a long-running agent thread; the source's central untested claim is that persistent external state survives hundreds of these without losing track > - **orchestrator agent** — an agent that spawns and coordinates implementation/review sub-agents once a plan has passed adversarial review > - **slop** — low-effort, unreviewed AI-generated output; the source concedes casual one-off requests can still produce it even under a disciplined workflow ## Metadata - Source: [Gonna share a few general thoughts and tips about how...](3.%20Resources/Readwise/Full%20Document%20Contents/Tweets/Gonna%20share%20a%20few%20general%20thoughts%20and%20tips%20about%20how....md) - Author: Joey Kudish (consultant, open-source tool author) - URL: https://twitter.com/jkudish/status/2083390828249002195/ ## Author's thesis Kudish describes his current agentic-development workflow as inverted from common practice: long planning sessions where the agent maintains planning docs and decision logs and interviews him, adversarial review agents applied to the *plan* rather than the code, persistent external state that (he reports but hasn't benchmarked) survives hundreds of context compactions across month-long threads, and cheap multi-model cross-review assigning different models to different roles. He's explicit about the workflow's limits — output can still be "sloppy" on casual one-off requests, and the retention claim is self-reported and unmeasured — and closes by urging developers anxious about AI to raise their ambition. ## Key arguments (≤3) 1. **Adversarial review belongs at the plan stage, before any code exists.** Multiple review agents inspect generated plans for defects and inconsistencies, looping until clear, before orchestrators spawn implementation — moving review to where defects are cheapest to fix, which inverts most published practice of reviewing generated code. 2. **Persistent external state is what makes long-horizon agent work survive** — several Codex threads with a month or more of near-daily interaction and hundreds of compactions reportedly lose no context, attributed to a save-as-you-go planning-doc/decision-log system. Immediately and honestly conceded as unbenchmarked. 3. **Both Codex and Claude default toward preserving backwards compatibility**, requiring explicit instruction to discard it — a specific, checkable behavioral claim distinct from the thread's broader generalities, offered in service of a "software as drafts" argument that rewriting has gotten cheap. ## Claim atoms ### Review agents audit plans before any code exists > [!quote] ID: rw_01kyymp1c20zf355agmz5rzfr5 > I have multiple review agents look over generated plans for defects, inconsistencies, discrepancies Where/why: mid-thread, immediately before the orchestrator handoff. Inverts common practice of reviewing generated code; defects are cheapest to fix pre-implementation. ### Persistent state survives hundreds of context compactions > [!quote] ID: rw_01kyymp1c20zf355agmz5rzfr5 > 100s of compactions and 5.6 Sol doesn't lose track of anything Where/why: describing month-long near-daily Codex threads. Attributed to save-as-you-go planning docs and decision logs — explicitly not benchmarked; the thread's central untested hypothesis. ### Models default to preserving backwards compatibility > [!quote] ID: rw_01kyymp1c20zf355agmz5rzfr5 > both Codex and Claude really like to be backwards compatible so you have to be quite explicit Where/why: following the software-as-drafts argument. The most falsifiable claim in the thread, testable in an afternoon. ### Quality gains land in surrounding systems, not code > [!quote] ID: rw_01kyymp1c20zf355agmz5rzfr5 > the systems around it (testing, security, validation, documentation, monitoring etc) are better than they would have ever been Where/why: the verification-loops passage, immediately after the slop concession. The more interesting and more measurable claim, distinct from a bare code-quality assertion. ### Ask the model to contradict you > [!quote] ID: rw_01kyymp1c20zf355agmz5rzfr5 > ask the model to contradict you, or to form its own informed opinion Where/why: offered tentatively. A cheap, single-session adversarial technique complementing the multi-model review pattern at lower cost. ## Permanent candidates - **Review the plan, not the output** — adversarial review belongs at the earliest inspectable artifact, since defect cost rises with how much is built on top of it. No existing note captures this specific plan-stage-review claim. Promoted below. - **Externalized state substitutes for model memory** — closely related to and largely covered by the existing [[3. Resources/Permanent Notes/Intent Written on the Outside Compounds While Re-Derived Intent Is Debt]]. Connect rather than duplicate; that note could add this source's specific "survives hundreds of compactions" evidence line if revisited. ## Processing checklist - [ ] Test the backwards-compatibility default claim across Codex and Claude with identical refactor prompts - [ ] Check whether the "librarium" skill already present in this environment is the same tool as the author's open-source Librarium - [x] Connect into the relevant MOC — [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] ## Links - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — plan-stage review, multi-model workflows, agentic development practice - [[3. Resources/Permanent Notes/The Maker Should Not Grade Its Own Homework]] — already cites this source