Don't Ignore Your SOUL.md With Erik Hanchett  🔗 **Source:** [Don't Ignore Your SOUL.md With Erik Hanchett](https://www.youtube.com/watch?v=NOXLNvfG2IU) > [!note]+ Document Notes > tags: [ai-coding-agents, agent-configuration, context-engineering, developer-tools, model-routing] > author: Erik Hanchett > source: O'Reilly > reference: https://www.youtube.com/watch?v=NOXLNvfG2IU > date created: 2026-07-16 08:50 > > # 🏷️ Title: Don't Ignore Your SOUL.md with Erik Hanchett > <!-- deep-read:sections start --> ## 🧭 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) ## 📝 Summary - Developer and educator Erik Hanchett argues that SOUL.md — a persona-definition file used in AI coding-agent workspaces (OpenClaude, per the video description; the auto-captions render it as "Open Plot"/"OpenClaw") — is the most-skipped and most-valuable config file in an agent setup. - SOUL.md acts as a "character sheet": personality, voice, temperament, and behavioral boundaries for the agent, including explicit model-routing rules (cheap model for routine work, frontier model only when asked for deep analysis). - Companion files matter too: MEMORY.md holds long-term preferences/standing rules injected into every session; AGENTS.md/USER.md round out the workspace. Overlap between files is normal and not worth agonizing over. - Session-hygiene commands (`/new`, `/compact`, `/btw`) manage context-window bloat and token cost as conversations grow. - Backing up the whole agent workspace to a private GitHub repo (with auto-commit after each session) is presented as a near-mandatory hedge against corruption, bad updates, or reinstalls. - Stakeholders: individual developers/power users configuring agentic coding tools (Claude Code / OpenClaude-style CLIs); secondarily, teams standardizing on shared agent config conventions. - Risk/tradeoff: duplicating instructions across SOUL.md/AGENTS.md/MEMORY.md can create drift or conflicting guidance over time; the talk waves this off as acceptable but doesn't offer a reconciliation strategy. - Bottom line: treat the agent's config files as living infrastructure — persona, memory, and backup — not one-time onboarding boilerplate. ## 📖 Main Content - The core argument is that most users write (or accept) a generic onboarding SOUL.md once and never touch it again, losing the chance to shape agent behavior deliberately. - Hanchett frames SOUL.md less as documentation and more as a persistent system prompt: identity ("collaborator, not assistant"), interaction style (push back, ask clarifying questions), and cost-control policy (model tier by task type) all live there. - MEMORY.md is distinguished from SOUL.md by function, not just name: SOUL.md is who the agent *is*; MEMORY.md is what it *remembers* (preferences, standing rules, key contacts) and both get folded into context on every turn. - Session management commands are presented as practical token-economy tools: `/new` resets context while preserving files on disk; `/compact` summarizes and prunes an over-long session automatically (or on demand) but is lossy ("it dies on the compaction"); `/btw` lets you ask an aside question that never enters the persisted context at all. - The GitHub-backup recommendation is framed as insurance against the inherent fragility of fast-moving CLI agent tooling — reinstalls, corruption, and "bad updates" are treated as expected, not exceptional, events. - Uncertain/unverified: the tool name itself. Auto-captions render it variously as "Open Plot" and "OpenClaw," but the YouTube description confirms the actual product is "OpenClaude." This is a known auto-caption artifact, not a claim about a different product — see Accessibility Assessment. - No named competing frameworks or counter-arguments are presented; the talk is instructional rather than comparative. ### 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=NOXLNvfG2IU&t=0) SOUL.md as the agent's character sheet - SOUL.md defines personality, voice, temperament, and permitted actions — Hanchett calls it "basically the character sheet for the agent." - Practical instruction: after installing the agent tool, go to Files → soul (GUI) or `.openplot/workspace` (CLI — likely a caption mis-transcription of the actual OpenClaude workspace path) to find it. - Recommends writing more persona detail into SOUL.md than what ships in the default onboarding file. #### ⏱️ [00:00:32](https://www.youtube.com/watch?v=NOXLNvfG2IU&t=32) Identity and collaborator framing - Hanchett shares his own SOUL.md language: the agent is told it's "your collaborator, not an assistant," instructed to push back when the user is wrong, challenge ideas during brainstorming, and ask clarifying questions. - Presented as a copy-paste-able default ("you can screenshot this if you like"). #### ⏱️ [00:00:44](https://www.youtube.com/watch?v=NOXLNvfG2IU&t=44) Model-routing defaults for cost control - Default policy: only escalate to a more powerful model when explicitly asked for deep analysis; use the cheapest capable model for routine tasks. - With multiple models configured as fallbacks, route expensive work to a stronger model (Claude is named specifically) and everything else to a cheaper one. #### ⏱️ [00:01:01](https://www.youtube.com/watch?v=NOXLNvfG2IU&t=61) Boundaries and cross-file duplication - Individual behavioral boundaries can also live in AGENTS.md — Hanchett explicitly says not to worry about duplicating information across the core config files; "that happens." #### ⏱️ [00:01:19](https://www.youtube.com/watch?v=NOXLNvfG2IU&t=79) "Aggressive note taker" tip (credited to "Tina") - Suggested SOUL.md instruction: act as an aggressive note-taker — after every significant session, write key decisions, preferences, and facts into daily logs so they roll into memory. - Attributed as borrowed from someone named "Tina" (no further identification given in the transcript). #### ⏱️ [00:01:38](https://www.youtube.com/watch?v=NOXLNvfG2IU&t=98) Workspace file rundown - Acknowledges there are "so many different workspace files" and narrows attention to the ones that matter most: SOUL.md, MEMORY.md, AGENTS.md, and possibly USER.md. #### ⏱️ [00:01:58](https://www.youtube.com/watch?v=NOXLNvfG2IU&t=118) MEMORY.md mechanics - MEMORY.md holds communication preferences and standing rules; it's long-term memory injected into context on every interaction (preferences, response length, standing rules, key contacts). - Notes this file may already be partially populated depending on the tool's onboarding flow. #### ⏱️ [00:02:29](https://www.youtube.com/watch?v=NOXLNvfG2IU&t=149) `/new` for session hygiene - `/new` starts a fresh session while keeping all workspace files intact and accessible. - Warns against running one long, ever-growing chat session: it re-sends the full conversation history to the model every turn, burning tokens. #### ⏱️ [00:03:01](https://www.youtube.com/watch?v=NOXLNvfG2IU&t=181) `/compact` vs `/new` - `/compact` runs automatically when a session gets very large, or can be triggered manually; it summarizes session memory but is described as lossy ("it dies on the compaction") even though it also condenses the conversation for continued use. - `/new` is offered as the clean alternative — a genuinely fresh window. #### ⏱️ [00:03:46](https://www.youtube.com/watch?v=NOXLNvfG2IU&t=226) `/btw` for side questions - `/btw` ("by the way") lets the user ask an aside question that doesn't get added to persisted session context — no token/context-window cost, no risk of overloading the working conversation. #### ⏱️ [00:04:15](https://www.youtube.com/watch?v=NOXLNvfG2IU&t=255) Back up the workspace to a private GitHub repo - Frames reinstalls, corruption, and bad updates as inevitable given how fast the tooling moves ("We're all playing around with OpenClaw" — captioned name for OpenClaude). - Gives a literal copy-paste prompt: ask the agent to put its own workspace files into a private GitHub repo and auto-commit changes after each session, so configuration survives a reinstall. ## 📚 References - [Video: Don't Ignore Your SOUL.md with Erik Hanchett (O'Reilly, YouTube)](https://www.youtube.com/watch?v=NOXLNvfG2IU) - [O'Reilly on LinkedIn](https://www.linkedin.com/company/oreilly/) - Per the video description, this talk is excerpted from an "AI Superstream" session — no direct URL for that program was confirmed in the fetched content, so it is not linked here to avoid guessing. ## 🗣️ Paraphrase - Hanchett's pitch: don't let SOUL.md sit as unread onboarding boilerplate. Treat it as an editable persona contract — who the agent is, how bluntly it should challenge you, and which tasks are worth paying for a stronger model. - His second pitch: session hygiene and backups are not optional housekeeping. `/new`, `/compact`, and `/btw` keep token spend under control day to day, while a private-repo backup of the whole workspace protects the persona/memory investment from tool churn. ## 📰 Newsletter - Quick one for anyone running an agentic coding CLI (Claude Code, OpenClaude, or similar): Erik Hanchett's "Don't Ignore Your SOUL.md" walks through turning your agent's SOUL.md into an actual persona spec — collaborator-not-assistant framing, push-back behavior, and a cheap-model-by-default routing rule — instead of leaving the onboarding default in place. - Two habits worth stealing: log key decisions to a daily file so they roll into MEMORY.md, and auto-commit your whole agent workspace to a private GitHub repo so a bad update or reinstall doesn't wipe months of tuning. - Bonus token-economy tip: `/btw` lets you ask a side question without it ever touching your persisted context — cheaper than you'd think. - Watch: [O'Reilly, YouTube](https://www.youtube.com/watch?v=NOXLNvfG2IU) ## ✅ Action Items #### DevRel - Developer experience: publish a starter SOUL.md/MEMORY.md template pair (collaborator framing + model-routing defaults) as a quickstart artifact for agentic-CLI onboarding docs. - Education: a short workshop or office-hours segment on session-hygiene commands (`/new`/`/compact`/`/btw`) would directly reduce support questions about token burn. - Developer advocacy, developer marketing, community management, internal/external alignment: Not applicable to this source — the talk doesn't touch program design, campaigns, or cross-org SLAs. #### OSPO - Not applicable to this source — no license, governance, security-disclosure, or SBOM content is present; this is a single-user tooling configuration talk. #### Community - Maintainers/contributors: if a team standardizes on shared agent config files, a lightweight "config file conventions" doc (who owns SOUL.md edits, how to avoid AGENTS.md/SOUL.md drift) would head off the duplication problem Hanchett shrugs off. - Sponsors, ecosystem: Not applicable to this source. #### WordPress - Not applicable to this source — content is about general-purpose AI coding-agent configuration, not WordPress core, plugins, or community programs. #### The WPCC - Not applicable to this source. ### FAIR - Not applicable to this source — no funding, governance-artifact, or provenance content present. ## 🧪 Hypotheses - H1: Developers who explicitly define a "collaborator, not assistant" persona in an agent's system-prompt-equivalent file receive more pushback/clarifying questions from the agent than those using a default assistant persona. - H2: Explicit model-routing rules (cheap-by-default, escalate on request) measurably reduce per-session token/dollar cost compared to using a single model tier for all tasks. - H3: Teams that back up agent workspace config to version control recover from tool reinstalls/corruption faster than teams that don't. - H4: `/btw`-style side-channel questions reduce main-session context growth compared to asking the same question inline. - H5: Logging session decisions into a daily-log-to-memory pipeline improves cross-session consistency of agent behavior over time. ### 📊 Methodology - H1: A/B compare two agent configs (default persona vs. explicit collaborator persona) across matched brainstorming tasks; code push-back/clarifying-question frequency via transcript annotation. - H2: Instrument token usage and cost per session across a cohort using routing rules vs. a control cohort using a single model tier for equivalent task mixes. - H3: Time-to-recovery study — simulate a reinstall/corruption event for workspaces with and without a version-controlled backup; measure elapsed time and config-fidelity loss. - H4: Compare context-window growth (tokens) for sessions that use `/btw` for asides vs. sessions where the same asides are asked inline. - H5: Longitudinal case study of one or more users' MEMORY.md/daily-log pipeline, coding behavioral consistency (e.g., rule adherence) before and after adoption. ### ❓ Questions - Is "OpenClaude" (the auto-caption-garbled "Open Plot"/"OpenClaw") a specific named product, or a colloquial/internal name for a Claude Code-style CLI workflow? The transcript alone doesn't resolve this — only the video description supplies the spelling. - How do teams reconcile conflicting instructions when the same boundary is duplicated (and edited independently) across SOUL.md, AGENTS.md, and MEMORY.md? - What's the actual token/cost delta from disciplined use of `/new` and `/btw` versus running one long session, in a realistic workload? - Who is "Tina," credited as the source of the aggressive-note-taker tip — a named practitioner in this space worth following for more agent-config patterns? ### 🔮 Future Research - A comparative survey of SOUL.md/AGENTS.md/MEMORY.md-style convention files across different agentic CLI tools (naming, precedence rules, conflict resolution) — this talk assumes but never explains how conflicts are resolved. - Empirical work on optimal "daily log → long-term memory" pipelines: what should get promoted to MEMORY.md versus stay ephemeral. - Failure-mode research on `/compact`'s lossiness — what specifically gets dropped, and whether that's ever consequential for task continuity. ### 💡 Implications - Practice: agent configuration is being treated as a first-class, versioned artifact (persona + memory + backup) rather than disposable onboarding text — this is a meaningful workflow shift for developers adopting agentic CLIs. - Cost/economics: explicit model-routing-by-task-type in a persona file is a lightweight, low-effort lever for controlling LLM spend without building separate routing infrastructure. - Security/operations: recommending users hand their agent write access to create and auto-commit to a private GitHub repo has real operational implications (credential scope, accidental secret commits) that the talk doesn't address. - Education: as more tools standardize on similarly-named convention files (SOUL.md, AGENTS.md, MEMORY.md), onboarding material that explains the *distinctions* between them (as this talk does) becomes increasingly valuable to reduce config sprawl. ### 🧠 Learning Styles Assessment - Reading level estimate: conversational spoken transcript; approximate Flesch-Kincaid grade ≈ 7-8 (short clauses, informal connectors like "cuz," minimal jargon beyond tool-specific file names). #### 👀 Visual learners - Present: none — this is a transcript-only capture; the original video presumably shows on-screen file browsing and the actual SOUL.md text, none of which is recoverable from captions alone. - Improve: add a labeled screenshot of an example SOUL.md/MEMORY.md pair; a simple diagram showing which file feeds which part of the agent's context. #### 🔊 Aural learners - Present: this is native audio/video content — full narration is the primary format. - Improve: a 2-3 minute audio-only recap isolating just the four copy-paste-able snippets (persona, model routing, note-taker tip, backup prompt) would serve return listeners. #### 🗣️ Verbal learners - Present: analogy-driven explanation ("character sheet for the agent"); plain-language contrast between SOUL.md (identity) and MEMORY.md (recall). - Improve: a short glossary distinguishing SOUL.md / AGENTS.md / MEMORY.md / USER.md by one-line function would resolve the "so many different workspace files" confusion the speaker himself flags. #### 🧩 Kinesthetic learners - Present: literal copy-paste prompts given for persona text and the GitHub-backup instruction — inherently hands-on. - Improve: a companion checklist ("open your workspace → find SOUL.md → paste this → test with a brainstorm prompt") would convert the verbal walkthrough into a repeatable lab exercise. #### 🧠 Logical learners - Present: implicit decision rule for model routing (routine → cheap model; explicit deep-analysis request → powerful model). - Improve: a simple decision-tree diagram for "which file does this instruction belong in" and "which model tier for this task" would formalize rules currently only stated in prose. #### 🧑‍🤝‍🧑 Social learners - Present: informal credit to "Tina" for one tip, situating the advice within a practitioner community. - Improve: link to where "Tina" or other practitioners publish these patterns, and/or a discussion prompt inviting viewers to share their own SOUL.md snippets. #### 🧘‍♀️ Solitary learners - Present: content is structured as personal practice the viewer can adopt individually and privately. - Improve: a self-check prompt, e.g., "does your current SOUL.md say anything about push-back behavior or model cost? If not, add one line before your next session." #### 🌿 Naturalistic learners - Present: none — no sustainability, resource, or ecosystem framing in the transcript. - Improve: could tie the cheap-model-by-default routing rule to compute/energy-cost framing, but this would be an addition, not present in the source. ### 📊 Accessibility Assessment #### Screen reader semantics - Present: N/A at the source level (this is a transcript-only analysis, not the original page/video UI). - Improve: ensure the original video page provides a proper transcript link and heading structure for the surrounding article/description, not just captions. #### Keyboard navigation - Not applicable to this source — no interactive UI was fetched or reviewed, only the caption track. #### Contrast and readability - Not applicable to this source — no visual/UI artifact was reviewed. #### Text resizing and reflow - Not applicable to this source. #### Media and captions - Present: auto-generated English captions exist and were fetched successfully. - Gap found directly in this source: the auto-captions mis-transcribe the actual product name "OpenClaude" (confirmed via the video's own description text) as "Open Plot" and "OpenClaw" at multiple points — a concrete example of auto-caption unreliability for proper nouns/brand names in technical content. - Improve: this talk would benefit from human-reviewed captions or a corrected transcript, especially given it's instructional content where exact tool/command names (`.openplot/workspace` vs. the real path) matter for viewers following along. #### Forms and interactions - Not applicable to this source. #### Standards reference - WCAG 2.2 AA gap: 1.2.2 (Captions, Prerecorded) is technically met (captions exist) but caption *accuracy* for proper nouns falls short of what's needed for a viewer to reliably follow copy-paste instructions — a practical, if not strictly checklist, accessibility concern for instructional video. ### 🌍 DEIB Assessment #### Cultural breadth - Not applicable to this source — single presenter, no regional or multilingual framing present in the transcript. #### Inclusive language - Present: gender-neutral language throughout ("the agent," "it"); no exclusionary phrasing detected. #### Neurodiversity support - Present: content is chunked into short, topic-per-minute segments naturally (persona → routing → boundaries → memory → sessions → backup), which aids predictable structure even without explicit headings in the original video. - Improve: on-screen or spoken signposting ("next, let's talk about memory") would help — the transcript shows the speaker does this informally ("let's talk about the memory MD") but not consistently for every topic shift. #### Representation - Present: one named practitioner credit ("Tina") beyond the speaker; otherwise no representation signals present in the transcript to assess. #### Psychological safety - Present: not particularly relevant here — content is technical instruction, not discussion of sensitive topics; no content warnings needed. ### ✅ Logical Validity #### Strong reasoning - Causal/practical reasoning at 00:02:29–00:03:01: single long-running chat sessions resend the full conversation history each turn → token cost grows with session length → therefore `/new` (context reset) or `/compact` (summarization) reduce cost. This is a sound, verifiable mechanical claim about how LLM context windows work. - Practical inductive reasoning at 00:04:15: agentic CLI tools are new and rapidly updated → reinstalls/corruption are likely at some point → therefore back up configuration now, before it's needed. Reasonable risk-mitigation logic, though no base-rate evidence for "how often" corruption actually happens is offered. #### Fallacies - None significant identified. The talk is instructional/prescriptive rather than argumentative — claims are presented as practitioner recommendations ("I would highly recommend," "I stole this from Tina") rather than as proven facts, so there's little exposed surface for classic fallacies (no appeals to authority beyond personal practice, no strawmanning, no false dichotomies detected in the transcript). #### Tone of voice - Casual, practitioner-to-practitioner, first-person ("I give it an identity," "I stole this from Tina as well," "100% set this up"). Enthusiastic and prescriptive rather than hedged or academic. ### 🛡️ DARVO Analysis - Not applicable to this source — this is a solo instructional talk with no accusation, conflict, or interpersonal dispute present. No deny/attack/reverse-victim-offender pattern exists to analyze. ## 🎯 Conclusion - Hanchett's core claim: SOUL.md (and its companion files) is under-used leverage — a small amount of persona and policy text meaningfully changes how an agent behaves and how much it costs to run. - Practical next steps for viewers: (1) rewrite SOUL.md's identity section as an explicit collaborator-not-assistant persona with push-back instructions; (2) add an explicit cheap-model-default / escalate-on-request routing rule; (3) adopt `/new`/`/compact`/`/btw` deliberately rather than letting one session run forever; (4) set up automatic private-repo backups of the whole agent workspace before a reinstall forces the issue. ## 🔄 Refactor Appendix — Don't Ignore Your SOUL.md with Erik Hanchett ### 📚 Literature Atoms (H3) #### [H1] SOUL.md defines an agent's personality and behavioral boundaries > [!quote] ID:rw_NOXLNvfG2IU_h1 > your soul MD is like basically the character sheet for the agent. It defines the personality, voice, temperament, and what it's allowed to do. **Context & use**: - [00:00:00](https://www.youtube.com/watch?v=NOXLNvfG2IU&t=0) - Establishes the core reframe of the talk: a config file most users ignore is functionally the agent's system prompt for identity and permissions. **Tags**: #xref **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> #### [H2] Frame the agent as collaborator, not compliant assistant > [!quote] ID:rw_NOXLNvfG2IU_h2 > I told it it's your collaborator, not an assistant, push back when I'm wrong, challenge ideas during brainstorming, clarifying questions. **Context & use**: - [00:00:32](https://www.youtube.com/watch?v=NOXLNvfG2IU&t=32) - The concrete, copy-paste-able persona instruction the speaker offers as his own default — the aphoristic center of the talk. **Tags**: #aphorisms #xref **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> #### [H3] Route routine tasks to the cheapest capable model > [!quote] ID:rw_NOXLNvfG2IU_h3 > Only use a more powerful model when explicitly asked for deep analysis. For routine tasks, use the cheapest capable model. **Context & use**: - [00:00:44](https://www.youtube.com/watch?v=NOXLNvfG2IU&t=44) - A default cost-control policy meant to live in the same persona file, paired with a fallback-model setup. **Tags**: #xref **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> ### 🧠 Permanent Candidates (H4) #### [P1] Agent persona should default to collaborator, not assistant **Claim** (one sentence, Courtney's voice): Defining an AI agent's persona explicitly as a pushback-capable collaborator, rather than a compliant assistant, is a one-line change that reliably shifts the agent toward more critique and clarifying questions. **Evidence pointer**: [[#H2]] **Scope & conditions**: Holds for agentic coding CLIs with an editable persona/system-prompt file (SOUL.md, AGENTS.md, or equivalent); breaks down if the tool doesn't persist persona instructions across turns, or if "collaborator" framing isn't paired with explicit permission to disagree (framing alone without permission may not change behavior). **Links**: - [[AI Agent Configuration Files]] — because extends - [[Prompt Engineering Practices]] — because applies-in agentic-CLI persona design <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Confirm whether "OpenClaude" is a distinct named product or informal shorthand for a Claude Code-style workflow — auto-captions garble it as "Open Plot"/"OpenClaw"; video description is the only confirming source found — https://www.youtube.com/watch?v=NOXLNvfG2IU #xref #fleeting #inbox - [F2] Look into "Tina" (credited for the aggressive-note-taker tip) as a possible source for more agent-config patterns worth tracking — https://www.youtube.com/watch?v=NOXLNvfG2IU #xref #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Add a collaborator-not-assistant persona line and a cheap-model-default routing rule to any personal agentic-CLI config files — dev-tooling #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Don't Ignore Your SOUL.md With Erik Hanchett.md|Don't Ignore Your SOUL.md With Erik Hanchett by Erik Hanchett]] ## Highlights - Let's update your soul. So, your soul MD is like basically the character sheet for the agent. It defines the personality, voice, temperament, and what it's allowed to do. I would highly recommend everyone after you install Open Plot, just go into your agents under the files, go to soul, or if you like in the command line, you can go to .openplot/workspace. That's where all your really important files are. And then just give it a little bit more of a soul than maybe what you wrote in the onboarding. So, here's something I use. You can screenshot this if you like, but I give ([View Highlight](https://read.readwise.io/read/01kxnfqvb6aqtbb9g4gzfpz6yp)) - Note: Core reframe: SOUL.md as the agent's character sheet, not onboarding boilerplate. - Tags: atom/literature/h1 - it a an identity. I told it it's your collaborator, not an assistant, push back when I'm wrong, challenge ideas during brainstorming, clarifying questions. You might want to put that in there as well. Here's the defaults. Only use a more powerful model when explicitly asked for deep analysis. For routine tasks, use the cheapest capable model. So, if you have multiple models like I suggest you do with fallbacks, then you can maybe use that Claude model for the more expensive things and then use a cheaper model for everything else. ([View Highlight](https://read.readwise.io/read/01kxnfqzff7st75fkd3gzfzgj8)) - Note: Collaborator-not-assistant persona instruction + cheap-model-default routing rule. - Tags: aphorisms, atom/literature/h2