| name | honing-agent-facing-docs |
| description | Use when reviewing, refactoring, rewriting, or streamlining AGENTS.md, repo-local agent skills, skill references, or agent policy docs |
Honing Agent-Facing Docs
Overview
Agent-facing docs should load at the moment they help and stay quiet otherwise.
Required background: Use writing-skills for agent-facing docs and writing-clearly-and-concisely for all prose. For docs that serve both humans and agents, use both honing-human-facing-docs and this skill; also use documentation-writer.
When To Use
Use for AGENTS.md, CLAUDE.md, repo-local or user-global skills, skill references, skill-owned scripts, agent policy, validation rules, and maintainer docs that steer future agent behavior.
Do not use for ordinary user guides unless agent behavior or policy placement is part of the task.
Placement Rule
Put each durable fact in the narrowest place that will load when needed.
| Destination | Belongs There |
|---|
AGENTS.md | Always-loaded law: repo role, hard rules, source boundaries, generated-artifact policy, routing, validation expectations, and pointers. |
| Triggered skill | Task-shaped workflow for recognizable work. |
| Skill reference | Bulky examples, command catalogs, matrices, failure signatures, and rationale. |
| Skill script | Repeatable checks, scaffolding, renderers, linters, and inspections. |
| Repo docs | Human-readable architecture, operations, package behavior, troubleshooting, and policy rationale. |
| Nowhere | Obsolete, duplicated, unactionable, private-host, reliably scoutable, or wrong content. |
Method
- Inventory docs, skills, references, scripts, commits, and generated artifacts that may hold durable knowledge.
- Triage each meaningful instruction by audience, trigger, durability, actionability, and source-of-truth ownership.
- Move useful content before deleting it. Preserve the last useful copy unless the fact is wrong or reliably derivable from source.
- Shrink always-loaded docs to operating law. Move recipes, matrices, incident stories, command catalogs, and long rationale behind triggered paths.
- Turn repeated mechanical instructions into scripts or tests where practical.
- Verify discovery: a future agent can find the rule, source file, command, and closeout check without reading the whole repo.
- For no-code doc hones, request review for placement, discoverability, lost information, stale claims, and workflow fit.
Routing Checks
- Keep onboarding material in always-loaded docs only when needed before an agent can pick a workflow.
- Keep incident history only when it explains a live constraint; otherwise state the current rule.
- Put long command catalogs in repo docs when humans use them, in skill references when only agents need them, and in scripts when repeatable.
- Keep exact error strings in troubleshooting references when they are useful search handles.
Closeout
- Always-loaded guidance is shorter and still sufficient.
- Triggered guidance has clear entry conditions and closeout checks.
- Bulky material moved to references, scripts, or repo docs.
- Human-facing material did not get buried inside agent-only surfaces.
- Links and commands needed for discovery work.
- Skill changes followed
writing-skills, including pressure-scenario testing.