Install
npx skillscat add marin-community/marin/fix-docs Install via the SkillsCat registry.
SKILL.md
Fix the markdown docs within lib/iris, lib/zephyr and lib/fray so they
comply with the principles below. Do NOT touch docs outside those directories.
Output: dispatch sub-agents that (1) parse the code and docs and (2) make the
documentation changes locally. Commit the changes into a single local commit,
inform the user, and summarize what changed. Never push without explicit user
approval.
Principles
- Agents are the primary consumers of documentation. Coding agents write
most Marin code; humans use agents as "documentation brokers". Segregate docs
by audience:- Human docs live in
docs/— high-level, "getting started", concise. - Agent docs live in the directory tree next to the code they describe —
token-efficient.
- Human docs live in
- Agentic docs follow standards so agents know where to look. Entry points:
**/AGENTS.md— targeted agent guidance for the subtree, plus a recursive
ToC/index of subtree content.**/OPS.md— field guide for gathering telemetry, running commands, and
debugging the systems in the subtree..agents/skills/*/SKILL.md— prompts for common agentic workflows.
- Avoid rot. Rot confuses agents and wastes context. When a doc is out of
date with the code: update it (if recent and the change is small) or archive
it to.agents/projects/<YYYYMMDD>_<filename>.md(if historical-only; date from
the first commit viagit). The code is the source of truth. - Context quality and quantity are paramount. Model performance degrades on
a full context window. Agent docs should focus on:- "Negative guidance": assume agents are smart; document sharp edges based on
observed suboptimal agent behavior. - Using markdown as indexes/ToC for context efficiency ("for X, read file Y").
- "Negative guidance": assume agents are smart; document sharp edges based on
Common issues
- Stale design docs that guided initial system design — move to
.agents/projects/<YYYYMMDD>_<filename>.md, dated from the first commit
(git).