sync-agents-md
Syncs AGENTS.md with actual repository state. Prevents stale documentation by grounding facts in evidence.
Install
mkdir -p .claude/skills/sync-agents-md && curl -L -o skill.zip "https://agentskills.codes/api/skills/download/17287" && unzip -o skill.zip -d .claude/skills/sync-agents-md && rm skill.zipInstalls to .claude/skills/sync-agents-md
Activation
This is the description your AI agent reads to decide when to run this skill — the better it matches your request, the more reliably it fires.
Use when re-syncing AGENTS.md after the project has drifted — triggers: \"update AGENTS.md\", \"resync AGENTS.md\", \"is AGENTS.md still accurate\", \"audit AGENTS.md for drift\". Corrects AGENTS.md from repo evidence, requiring explicit approval before writing. Initial generation is init-project's job; constitution changes go through amend-constitution.Key capabilities
- →Re-sync `AGENTS.md` against the current repository state.
- →Diff existing `AGENTS.md` content against reality.
- →Propose corrections based on repository evidence.
- →Preserve human-authored prose during resync.
- →Identify commands and tech stack from build files and lockfiles.
- →Detect project structure from `git ls-files`.
How it works
The skill analyzes the repository for structural changes, dependencies, and CI workflows, then compares these findings against `AGENTS.md` to propose factual corrections while preserving human-authored prose.
Inputs & outputs
When to use sync-agents-md
- →Updating documentation after architecture shifts
- →Verifying AGENTS.md accuracy
- →Grounding agent guidelines in repository evidence
About this skill
Sync AGENTS.md
Re-syncs AGENTS.md against the current repo state.
Diffs existing file against reality, proposes corrections, preserves human prose.
Nothing written until approved.
Why this skill exists
As commands, deps, and structure shift, AGENTS.md goes stale silently. Stale is worse than empty — every session trusts it. Re-grounds it in evidence.
Behavioral guardrails (active for the entire session)
<!-- GUARDRAILS:skill -->- No guessing. Where input leaves something unspecified, write
[NEEDS CLARIFICATION: specific question]and surface it — never silently invent an assumption. - Investigate before claiming. Never make statements about the codebase without first reading the relevant files. If a claim requires looking at code, look first.
- Conservative by default. Recommend before you write; stop and ask before anything irreversible (deleting files, force-pushing, dropping tables, external service calls).
- Evidence or nothing. Every line written into AGENTS.md must trace to a
file you actually read (a build file, lockfile, CI workflow, the directory
tree, an existing doc). If you cannot ground a fact, do not write a
plausible-sounding one — write
[NEEDS VERIFICATION: what to check and where]. This is the same no-fabrication rule the constitution applies to cross-repo signatures. - Respect AGENTS.md's own ground rule. Every line should be something an agent could NOT infer from training. Do not add generic advice ("write clean code," "follow best practices") — if a line would be true of any repo, leave it out. Prefer deleting a section to filling it with filler.
- Never thicken the pointer files. Project facts go in AGENTS.md only.
Never write conventions into
CLAUDE.mdor.github/copilot-instructions.md— the pre-commit hook and CI block those from growing past a pointer. - Stay in lane. Universal principles →
amend-constitution, not here. Feature requirements →spec.md, not here. Path-scoped subtree rules →.github/instructions/*.instructions.md, not AGENTS.md. - Preserve human prose on resync. Treat existing filled-in text as authoritative unless the repo contradicts it. Touch only the lines that have actually drifted; never rewrite a section wholesale to "tidy" it.
Before starting
- Confirm
AGENTS.mdis filled in, not the stub — if stub, tell user to runinit-projectfirst. - Read
memory/constitution.md(or the path AGENTS.md references) — don't duplicate constitution material.
Evidence sources (read before writing anything)
Detect the ecosystem — don't assume it.
- Commands & tech stack:
package.jsonscripts,pom.xml/build.gradle,pyproject.toml/Makefile/tox.ini,go.mod,Cargo.toml,Gemfile, etc. Take the exact build/test/lint/run commands with their flags — AGENTS.md's Commands section is referenced constantly and must be exactly right. - Versions & key libraries: lockfiles and manifest version pins. Record specific versions, not categories ("React 18 + Vite + Tailwind," not "React project").
- Project structure: the real top-level directory tree (
git ls-files, not a guess). Describe what actually lives in each path. - Git / PR workflow: branch names and commit-message shape from
git log, plusCONTRIBUTING*, PR templates, and CODEOWNERS if present. - CI / required checks:
.github/workflows/*(and this kit'sagent-harness.yml) for what blocks a merge. - Boundaries & brownfield areas: directories with no tests, generated code, vendored dirs — candidates for the ⚠️ Ask-first / 🚫 Never tiers and for characterization-test flags. Note every fact's source so the draft can cite it.
Resync
- Re-derive facts from the evidence sources above.
- Build a drift report: section · current claim · actual · proposed edit.
- Flag — don't auto-change — anything ambiguous or that may reflect a deliberate human decision. Ask the user.
- Edit only drifted lines — never rewrite a section wholesale.
Approval gate
Stop. Present the drift report and ask for explicit approval, resolving any
[NEEDS VERIFICATION] markers. Call out every judgement call. Write only after
approval.
Verify before finishing
After writing:
- No leftover
[bracketed placeholders]in committed AGENTS.md (except accepted[NEEDS VERIFICATION]markers). - Each command in Commands exists in its source. Where cheap,
--helpto confirm. Don't run build/test suites. - No project facts in
CLAUDE.mdor.github/copilot-instructions.md.
What this skill does not do
- Doesn't amend the constitution (
amend-constitutiondoes that). - Doesn't populate
spec.md/plan.md/tasks.md. - Doesn't fabricate facts — unverifiable lines are marked, not guessed.
Keeping it current
Offer to schedule a resync cadence or CI warning after completing successfully.
When not to use it
- →When `AGENTS.md` is a stub and `init-project` has not been run.
- →When amending the constitution, which is handled by `amend-constitution`.
- →When populating `spec.md`, `plan.md`, or `tasks.md`.
Limitations
- →The skill requires explicit approval before writing any changes to `AGENTS.md`.
- →It does not amend the constitution; that is handled by `amend-constitution`.
- →It will not fabricate facts, marking unverifiable lines as `[NEEDS VERIFICATION]`.
How it compares
This skill systematically re-grounds `AGENTS.md` in repository evidence, ensuring its accuracy and preventing staleness, which is more reliable than manual updates or assumptions about project state.
Compared to similar skills
sync-agents-md side by side with the closest alternatives in the catalog.
| Skill | Installs | Updated | Safety | Difficulty |
|---|---|---|---|---|
| sync-agents-md (this skill) | 0 | 19d | No flags | Intermediate |
| skill-development | 17 | 8mo | Review | Intermediate |
| skill-writer | 27 | 9mo | Review | Beginner |
| microsoft-skill-creator | 6 | 4mo | Review | Advanced |
Try saying
Example prompts that trigger this skill in your AI assistant.
More by saptarshibasu
View all by saptarshibasu →You might also like
skill-development
anthropics
This skill should be used when the user wants to "create a skill", "add a skill to plugin", "write a new skill", "improve skill description", "organize skill content", or needs guidance on skill structure, progressive disclosure, or skill development best practices for Claude Code plugins.
skill-writer
pytorch
Guide users through creating Agent Skills for Claude Code. Use when the user wants to create, write, author, or design a new Skill, or needs help with SKILL.md files, frontmatter, or skill structure.
microsoft-skill-creator
MicrosoftDocs
Create agent skills for Microsoft technologies using Learn MCP tools. Use when users want to create a skill that teaches agents about any Microsoft technology, library, framework, or service (Azure, .NET, M365, VS Code, Bicep, etc.). Investigates topics deeply, then generates a hybrid skill storing essential knowledge locally while enabling dynamic deeper investigation.
cartographer
kingbootoshi
Maps and documents codebases of any size by orchestrating parallel subagents. Creates docs/CODEBASE_MAP.md with architecture, file purposes, dependencies, and navigation guides. Updates CLAUDE.md with a summary. Use when user says "map this codebase", "cartographer", "/cartographer", "create codebase map", "document the architecture", "understand this codebase", or when onboarding to a new project. Automatically detects if map exists and updates only changed sections.
creating-skills
letta-ai
Guide for creating effective skills. This skill should be used when users want to create a new skill (or update an existing skill) that extends Letta Code's capabilities with specialized knowledge, workflows, or tool integrations.
add-pattern
langroid
Use this skill when you learn one or more design pattern(s) in the Langroid (multi) agent framework, and want to make a note for future reference for yourself. Use this either autonomously, or when asked by the user to record a new pattern.