Agent work documentation policy

# Agent work documentation policy

AI Summary

Purpose:

• Records the user's standing policy for how AI agents should document work, synchronize the LLM Wiki, and handle portfolio/resume writing.

Key points:

• Meaningful conversations, decisions, and work results should be documented in ~/workspace/repos/llm-wiki.
• llm-wiki should be pushed to the remote at least hourly when there are local changes.
• Code changes should happen in the target development repo or a new git repo, not as loose untracked files in the wiki workspace.
• Resume and portfolio writing should use https://github.com/epoko77-ai/im-not-ai plus humanizer-style review to remove Korean AI-writing tells before publishing.

Relevant when:

• Starting any agent session with durable decisions or work.
• Finishing coding, deployment, portfolio, resume, report, or research tasks.
• Preparing Korean human-facing output under human/**.

Do not read full document unless:

• You are deciding where to store work results or how to publish human-facing Korean text.

Linked documents:

• AGENTS.md
• ai/index.md
• ai/worklog/index.md
• ai/repo-notes/index.md
• ai/wiki/concepts/llm-wiki.md
• ai/repo-notes/hyunwook-tech.md

Open Questions

• Whether every single chat turn should be archived verbatim is unresolved. Current interpretation: store meaningful decisions, work summaries, commands, outcomes, and reusable context rather than raw chat logs unless the user explicitly asks for transcript archival.

Details

Standing policy

The user wants durable work memory to live in this repository. For future sessions, agents should treat ~/workspace/repos/llm-wiki as the canonical place for documenting:

• user decisions and preferences that affect future work;
• task plans and final outcomes;
• deployment steps, verification results, rollback locations, and backup paths;
• repo-specific lessons learned;
• portfolio/resume content decisions;
• reusable procedures or pitfalls.

Do not leave important context only in chat. At the end of meaningful work, update the appropriate layer:

• ai/worklog/YYYY/YYYY-Www.md for time-based work history;
• ai/repo-notes/<repo>.md for durable repo-specific knowledge;
• ai/wiki/** for stable concepts, projects, people, and long-lived facts;
• human/** only for polished Korean outputs generated from ai/**.

Git synchronization

The wiki should be pushed to its git remote after completed edits. In addition, an hourly automated sync should commit and push local llm-wiki changes when present. The sync should be conservative:

1. Work from /home/khw-bot/workspace/repos/llm-wiki.
2. Pull with rebase before committing.
3. Add and commit local changes only if the working tree is dirty.
4. Pull with rebase again before push.
5. Push to the configured remote.
6. Fail loudly rather than resolving rebase conflicts automatically.

Code work location

Code changes should not be mixed into llm-wiki unless the code is part of the wiki itself. For other projects:

• locate or create the relevant git repository;
• work inside that repository;
• commit/push there according to the project's workflow;
• return to llm-wiki only to document the work and durable lessons.

Korean human-facing writing policy

For resume, portfolio, brief, report, and other human-facing Korean text, use both:

• https://github.com/epoko77-ai/im-not-ai, a Korean AI-writing humanization tool/skill focused on preserving meaning while removing Korean AI tells;
• Hermes humanizer skill rules for anti-AI prose review.

Use this especially for:

• human/portfolio/**;
• resume-style portfolio content in hyunwook.tech;
• public-facing career summaries;
• Korean reports, briefs, and slides.

The practical review checklist:

• preserve facts, numbers, names, URLs, dates, and scope;
• remove generic AI phrases, translationese, mechanical parallelism, over-polished marketing language, excessive bullets/bold/emojis, and uniform sentence rhythm;
• keep Korean natural, specific, and grounded in the actual operational problem;
• do not invent anecdotes, metrics, warmth, or ownership;
• mark unverified facts as Needs confirmation rather than smoothing over uncertainty.