# Don't Ignore Your SOUL.md with Erik Hanchett β€” Erik Hanchett ## πŸ—‚οΈ Triage β€” annotated highlights surfaced > [!note]- Cross Reference > - **SOUL.md defines an agent's personality and behavioral boundaries** β€” a config file most users ignore is functionally the agent's identity system prompt. #xref > - **Frame the agent as collaborator, not compliant assistant** β€” a one-line persona change that shifts the agent toward pushback and clarifying questions. #xref > - **Route routine tasks to the cheapest capable model** β€” a cost-control policy that lives right in the persona file. #xref ## 🧭 Bib capsule Erik Hanchett (O'Reilly AI Superstream), "Don't Ignore Your SOUL.md," YouTube, 2026 β€” https://www.youtube.com/watch?v=NOXLNvfG2IU ## 🧠 Author's thesis (Adler Q1+Q2) SOUL.md β€” a persona-definition file in AI coding-agent workspaces (OpenClaude, per the description; auto-captions garble it as "Open Plot"/"OpenClaw") β€” is the most-skipped and most-valuable config file, and treating it as a living persona spec rather than one-time onboarding boilerplate meaningfully changes how the agent behaves and what it costs. In detail: SOUL.md is who the agent is (collaborator-not-assistant, pushback, model-routing rules); MEMORY.md is what it remembers (standing rules injected every session); session-hygiene commands (`/new`, `/compact`, `/btw`) control token bloat; and auto-committing the whole workspace to a private GitHub repo hedges against the inevitable corruption and bad updates of fast-moving CLI tooling. ## 🧩 Key arguments - Config files are living infrastructure β€” persona, memory, and backup β€” not disposable boilerplate; write more persona detail than the default ships with. - Model-routing-by-task-type in the persona file is a low-effort lever on LLM spend: cheapest capable model by default, escalate to a stronger model (Claude named) only when explicitly asked for deep analysis. - Cross-file duplication (SOUL.md / AGENTS.md / MEMORY.md) is treated as acceptable β€” "that happens" β€” but the talk offers no reconciliation strategy, leaving a real drift/conflict risk. ## πŸ—’οΈ Notes β€” Literature atoms ### SOUL.md defines an agent's personality and behavioral boundaries > [!quote] > 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. - **Where**: 00:00:00 - **Why**: establishes the core reframe β€” a config file most users ignore is functionally the agent's system prompt for identity and permissions ### Frame the agent as collaborator, not compliant assistant > [!quote] > I told it it's your collaborator, not an assistant, push back when I'm wrong, challenge ideas during brainstorming, clarifying questions. - **Where**: 00:00:32 - **Why**: the concrete, copy-paste-able persona instruction β€” the aphoristic center of the talk ### Route routine tasks to the cheapest capable model > [!quote] > Only use a more powerful model when explicitly asked for deep analysis. For routine tasks, use the cheapest capable model. - **Where**: 00:00:44 - **Why**: a default cost-control policy meant to live in the persona file, paired with a fallback-model setup ## πŸ”– Permanent candidates - Agent persona should default to collaborator, not assistant β€” defining an AI agent's persona explicitly as a pushback-capable collaborator (paired with explicit permission to disagree) is a one-line change that reliably shifts it toward more critique and clarifying questions. ## Links - [[AI Agent Safety MOC]] β€” home; SOUL.md/MEMORY.md as steering and boundary-setting is agent-safety hygiene β€” persona, memory, and versioned backup treated as living infrastructure - [[Personal Knowledge Management MOC]] β€” the daily-log β†’ MEMORY.md pipeline and private-repo backup are PKM practices applied to an agent's own memory - [[Setting Up a Prompt Engineering Workflow]] β€” the persona file is a persistent system prompt; this connects agent-config to prompt-engineering practice ## βœ… Processing checklist - [x] Claim atoms carried from deep-read analysis - [ ] Permanent candidates reviewed for promotion - [ ] Version History Diff check