| name | agents-md-improver |
| description | Use to audit or improve AGENTS.md / CLAUDE.md / GEMINI.md / WORKFLOW.md against the canonical fleet entry-doc pattern. Fires when authoring or revising entry docs, when audit-entry-docs.sh flags drift, or when a new repo joins the fleet and needs entry docs scaffolded. Do not use to write product documentation — entry docs are agent-control content per OPERATING_MODEL Source Of Truth. |
AGENTS.md Improver
Use this skill when authoring, auditing, or repairing the entry-doc
set in any fleet repo. The skill encodes the canonical pattern
declared in
file://OPERATING_MODEL.md §"Source Of Truth" and §"Public
OSS posture", and is the policy reference that
file://scripts/audit-entry-docs.sh enforces mechanically.
When to use
- Creating entry docs for a new repo joining the fleet.
- Revising
AGENTS.md or CLAUDE.md in any existing repo.
- An
audit-entry-docs.sh run reported drift.
- A propagation cycle introduced changes that touch entry docs.
Canonical pattern (internal repos)
INTERNAL repos (active child product repos plus the source
policy repo) MUST have:
AGENTS.md (the canonical, agent-agnostic entry doc — read by
the agent runner, Cursor, Aider, Copilot, Jules, and other AGENTS-aware
tools per url://https://agents.md). REQUIRED sections:
## What this repo is — a one-paragraph orientation.
## Engineering standards — the studio rules (Conventional
Commits; explicit staging; no AI attribution; no verify-bypass;
no push to the protected branch). Reference enforcement hooks
by name where applicable.
- The literal phrase "Fleet Rule Origination" referencing
file://OPERATING_MODEL.md §"Fleet Rule Origination".
WORKFLOW.md (the autonomous-dispatch runner-compatible dispatch contract). MUST be
composed of three parts: per-repo YAML config (tracker /
polling / workspace / hooks / agent / codex / bes); a one-section
per-repo intro naming the repo and its canonical verify command;
the fleet-baseline prompt body verbatim from
<adopter-policy-repo>/agents/templates/WORKFLOW.body.md (fleet-baseline reference; bes-fleet-policy-layout-specific) (or
.workpads/WORKFLOW.body.md in a child repo).
Per-repo content lives ABOVE the body; the body is fleet-uniform.
Drift between a repo's WORKFLOW.md body and the fleet baseline
is a blocking failure surfaced by audit-entry-docs.sh.
CLAUDE.md (OPTIONAL but recommended) — Claude Code reads this
first per url://https://code.claude.com/docs/en/memory. It MUST
start with @AGENTS.md to import the canonical content; only add
Claude-specific extensions on top (slash-command list, hook
inventory, posture).
GEMINI.md (OPTIONAL) — same pattern as CLAUDE.md but for
Gemini CLI.
Canonical pattern (public OSS repos)
Public OSS repos (per file://OPERATING_MODEL.md §"Public
OSS posture") MUST NOT have:
- Root-level
AGENTS.md, CLAUDE.md, GEMINI.md, or WORKFLOW.md.
- Tracked
.agents/ or .claude/ directories.
.agents/ and .claude/ content lays in the working tree (gitignored)
so any AI assistant landing locally has the engagement model
available; nothing reaches the public surface. See
file://OPERATING_MODEL.md §"Public OSS posture" for the
full rationale.
Hard rules
- AGENTS.md is canonical. CLAUDE.md and GEMINI.md MUST NOT carry
policy that the agent-agnostic file lacks. If a rule applies to
any agent, it belongs in AGENTS.md.
- Agent-specific extensions are import-only at the top
(
@AGENTS.md) plus tool-specific additions (slash-commands,
statusline, hooks). Don't duplicate AGENTS.md content; reference it.
- Public OSS repos are protected: agent infrastructure MUST stay
out of the GitHub surface. The auditor flags any tracked
AGENTS.md/CLAUDE.md/GEMINI.md/WORKFLOW.md in those repos
as a blocking failure.
- Fleet rule origination per
file://OPERATING_MODEL.md §"Fleet Rule Origination":
changes to AGENTS.md and CLAUDE.md in child repos that contradict
the canonical pattern are repo-local drift, not fleet changes.
Amend the canonical doc in your-policy-repo first, then
propagate.
Enforcement
file://scripts/audit-entry-docs.sh checks all of the above
mechanically. Run it from the studio root (validates all 7 repos)
or from any repo root (validates that one repo). Exit codes:
0 — all checks pass.
1 — at least one blocking criterion failed; per-repo PASS/FAIL
written to stderr with file:line citations.
Quality Gate Handoff
This skill does not approve or block work. It documents the policy.
Mechanical enforcement is audit-entry-docs.sh. Substantive
review of entry-doc changes goes through the v1 SPEC procedure:
amend canonical docs in your-policy-repo via IDEA → SPEC →
review → propagate.