repo-seed
Skill10 starsChanged 32 days agoMITView on GitHub
Seed, audit, adopt, and progressively evolve an agent-native repository with risk-triggered specs, durable decisions, resident governance skills, capability state, deterministic gates, and safe upgrades. Use when initializing or upgrading a repository, when an unmanaged repository grows complex, or when delivery, ownership, security, release, or incident signals suggest a governance capability is missing.
- Reads credentials
- Commits and pushes
In this repository6 other files it gives its agents.
- AGENTS.mdAGENTS.md
- docs/AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- repo-decisionsSkill
- repo-governanceSkill
- repo-reviewSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
17 rules. About 0 are common, 10 are distinctive, and 7 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈10
Distinctive · estimate
This author's own choices. Worth a look.
7
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 23
Run the five steps in order.
- Very specific: line 35
Never read .env files or other secrets.
- Very specific: line 37
When the repository already contains scripts/audit-governance.mjs, run it with --json.
- Distinctive: line 41
Ask only the questions whose answers are not detectable and that materially change the output.
- Distinctive: line 46
Test command, topology, and lint command — ask only for material facts that repository evidence cannot resolve.
- Very specific: line 55
Run the scaffold against the target directory.
- Very specific: line 67
Use --dry-run first and show the user the plan.
- Distinctive: line 87
After refining, re-record the manifest so it matches the shipped state, then run the gates and report.
- Distinctive: line 110
Write only the seeded paths.
- Distinctive: line 111
Never git commit or git push unless the user explicitly asks.
- Very specific: line 112
Never read .env files or secrets.
- Very specific: line 113
Never delete a file repo-seed did not create.
- Distinctive: line 114
Never overwrite a user-modified seeded file without asking (default: preserve).
- Distinctive: line 115
Never enable CI/CD, modify a remote platform, install a hook, or replace an external source of truth without explicit authority.
- Very specific: line 126
Use the legacy --extensions flag for file-backed compatibility capabilities, or explicit --capability-state assignments for the catalog …
- Distinctive: line 140
Use tool-agnostic wording for actions: "ask the user", "read the repository", "write files", "run a command" — never name a specific tool …
- Distinctive: line 142
Keep commands that are genuinely cross-tool (shell snippets, node scripts) as literal code blocks.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
repo-seed: seed an agent-native, self-governing repository
repo-seed turns a traditional repository — empty or existing, any technology stack — into an agent-native repository whose governance grows with its actual risk: resident instructions, risk-triggered Specs, durable MADR decisions, incident feedback, three in-repo skills, deterministic gates, and a capability manifest that preserves user ownership and external sources of truth.
repo-seed is a progressive governance system, not a static template. It may autonomously inspect, audit, and produce a dry-run. Writing the baseline, enabling a capability, installing a hook, changing policy/source-of-truth, or connecting an external system requires user authority. Re-running repo-seed is the only upgrade channel: untouched seeded files refresh, user-modified files are preserved, and user-created files are never deleted.
In a managed repository, the resident AGENTS router is authoritative. Ordinary implementation consumes its standing rules and linked project documents directly; review, durable decisions, and governance evolution use the resident repo-review, repo-decisions, and repo-governance skills. Invoke the global repo-seed skill only for seeding, adoption, governance-layer upgrades, or upstream repair.
Workflow
Run the five steps in order. Analysis, interview, and instantiation require model judgment; scaffold, record, audit, and verification mechanics are deterministic scripts.
1. Analyze
Inspect the target repository (read-only):
- Stack manifests:
package.json,pyproject.toml,go.mod,Cargo.toml,requirements.txt,Gemfile, etc. Detect test/lint commands plus actual unit, integration, end-to-end, fixture, and naming conventions when derivable. - Representative implementation: when code exists, sample two or three real entry points, modules, and public interfaces. Identify the repository's canonical domain nouns, state-transition verbs, naming idioms, and abstraction/placement conventions from repeated evidence; do not promote one outlier into policy.
- Existing files at seeded paths (see the file list in step 4). List every conflict: a file already exists where repo-seed wants to write.
- Existing governance equivalents: AGENTS/CLAUDE instructions, architecture/testing docs, ADR/RFC systems, postmortems, CI, release/deploy markers, hooks, CODEOWNERS, and external requirement links.
- Git state: is it a git repository? Is the working tree clean? What is the default branch?
- Repository size: for large repositories, sample (directory-depth limit and file-count cap); never enumerate the whole tree.
- Never read
.envfiles or other secrets.
When the repository already contains scripts/audit-governance.mjs, run it with --json. In an unmanaged repository, use the global skill's read-only detection and offer adoption only when the current task or repository facts show meaningful complexity. An audit never authorizes writes.
2. Interview
Ask only the questions whose answers are not detectable and that materially change the output. The question bank lives in references/interview.md. Every question has a default; never block on a question whose default is acceptable. Minimal set:
- Project one-liner (always; fills the AGENTS.md opening line).
- License — default MIT; skip if a LICENSE exists.
- Branch convention — default
mainwith short-lived feature branches. - Test command, topology, and lint command — ask only for material facts that repository evidence cannot resolve; never invent a tier or directory.
- Existing-file conflict resolution — preserve-and-merge (default) / overwrite-with-backup / skip.
- Monorepo — root-only seed (default) or also subtree AGENTS.md per package.
- Review policy input — known pitfalls, review red lines, and invariants this project must not regress (default: none; only the universal repo-review core).
- Capability choices — show only applicable recommendations with benefit, ongoing cost, urgency, and detected equivalent systems. Do not ask again when a declined/deferred assessment hash still matches.
- Hook authorization — default
skip; installing or replacing local Git state is never implied by seeding.
3. Scaffold (deterministic)
Run the scaffold against the target directory. The scaffold creates the approved directory skeleton, writes seeded files from the template set, and records .repo-seed/manifest.json with ownership, capability, governance-path, external-source, and artifact-policy state.
node <path-to-repo-seed>/scripts/scaffold.mjs <target-dir> \
--templates <path-to-repo-seed>/references/templates \
--extensions ci,release \
--hooks skip \
--repo-seed-version 0.6.5Flags: --dry-run (report only), --no-interview (non-interactive; preserve user-modified files), --adopt (reuse detected governance paths in an unmanaged repository), --hooks install|skip (default skip), --values k=v (repeatable), --extensions <ids> (legacy compatibility entry), --capability-state id=enabled|external|deferred|declined plus reason/assessment hash, --governance-path kind=relative/path, --external-source kind=https://stable-link, --user-owned <path> (repeatable), and --record-only (recompute manifest state without touching governed files).
Use --dry-run first and show the user the plan. Optional capabilities are never enabled implicitly. The legacy --extensions spec flag is accepted as a deprecated no-op because Spec is Core. An enabled capability remains enabled and upgradeable when a later run omits its legacy extension flag.
4. Instantiate (model)
For every seeded file that contains fill-in tokens (__UPPERCASE__), replace the tokens with content derived from the analysis and interview: project one-liner, real test/lint commands, stack description, architecture content for docs/architecture.md, and a risk-driven testing policy tailored to the detected stack in docs/testing.md. Resolve every token — the placeholder gate fails on any survivor. Do not invent commands, tiers, or directories: if a command cannot be resolved, omit that line rather than fabricate it.
Three files need composition, not string replacement:
docs/architecture.md— write the target's real module map and seams; never ship the placeholder skeleton.docs/testing.md— preserve the Core mission and evidence rules, then state the project's actual risk-to-layer choices, test roots, colocated naming conventions, fixtures, commands, and absent tiers from repository evidence..agents/skills/repo-review/SKILL.md— compose__REVIEW_PROJECT_BLOCKING__from AGENTS.md hard rules, interview Q8 red lines, and architecture seams, and__REVIEW_PROJECT_CHECKS__from the stack risk catalog plus repeated vocabulary and abstraction evidence in references/review-standard.md. Procedure skills (repo-decisions) ship static; review policy is instantiated per project. Keep the universal core verbatim, never invent a naming rule or glossary, and replace an empty token with "None beyond the universal requirements/checks."
Compose __DECISION_INDEX__ and __POSTMORTEM_INDEX__ from the records that actually exist after the run. Use one relative Markdown link per record, or None yet. for an empty postmortem index. These tokens keep user-created records discoverable across upgrades; never reset an existing index to the seed examples.
Compose __TEST_TOPOLOGY__ as a short list of the repository's actual test homes. Name each existing tier's path or colocated filename convention, the root entry or runner configuration that selects it, and fixture location where present; link rather than duplicate a command already owned by AGENTS.md. If a tier does not exist, state that fact and the risk that would justify adding it; do not create a directory merely to satisfy the template. Follow the ecosystem's idiom: for example Go unit tests stay beside source as *_test.go, while Python commonly uses tiered roots under tests/.
When adoption registers a non-standard path, instantiate __ARCHITECTURE_PATH__, __TESTING_PATH__, __DECISIONS_PATH__, __SPECS_PATH__, and __POSTMORTEMS_PATH__ from the manifest. Instantiate __DECISIONS_RULE__ as the local MADR lifecycle for a repository-owned decision log or as a no-duplicate-authority rule for an external decision source. Do not copy facts from an external source of truth into a second authoritative document.
5. Record and verify
After refining, re-record the manifest so it matches the shipped state, then run the gates and report. Mark the instantiated repo-review policy user-owned in the same step so a later re-run never refreshes it from the structure-only template.
node <path-to-repo-seed>/scripts/scaffold.mjs <target-dir> --record-only \
--user-owned .agents/skills/repo-review/SKILL.md
node scripts/run-gates.mjsThe gates run from the target directory. If and only if the user authorized hooks, pass --hooks install to the scaffold/apply invocation in step 3 and verify the managed hook calls run-gates.mjs --staged; --record-only does not install hooks. Preserve any custom hook. Then tell the user to review and commit. Never commit or push unless the user explicitly asks.
What gets seeded
The capability catalog is the single source of truth for the file contributions returned by seededFiles(). Core includes:
AGENTS.md(management identity, resident skill router, governance loop, and security rules),CLAUDE.md(@AGENTS.mdimport — no symlink, Windows-safe).docs/AGENTS.md(documentation standard), architecture/development docs, a risk-driven testing policy with instantiated topology,docs/specs/README.md, anddocs/postmortems/README.md.docs/decisions/— the unified MADR decision log with four seed records and an index..agents/skills/repo-review,.agents/skills/repo-decisions, and.agents/skills/repo-governance— resident review, rationale, and capability-evolution procedures.scripts/— capability/audit/configuration modules, the manifest-driven runner, deterministic verifiers, and the separately authorized hook installer.CONTRIBUTING.md,LICENSE,.editorconfig,.gitattributes,.github/(PR + issue templates),.repo-seed/update-strategy.md.
Security and ownership rules
- Write only the seeded paths. Never touch anything outside them without asking.
- Never
git commitorgit pushunless the user explicitly asks. - Never read
.envfiles or secrets. - Never delete a file repo-seed did not create.
- Never overwrite a user-modified seeded file without asking (default: preserve).
- Never enable CI/CD, modify a remote platform, install a hook, or replace an external source of truth without explicit authority.
- The authority for update semantics is references/update-strategy.md.
Progressive capabilities
The canonical catalog is CAPABILITIES in scripts/capabilities.mjs. Each CapabilityDefinition includes tier, files, prerequisites, applicability signals, equivalent-system signals, benefit, ongoing cost, urgency, AGENTS/review/gate contributions, and enable/upgrade/re-evaluation conditions. extensionPacks() and --extensions remain compatibility projections over the catalog.
Core includes resident instructions, architecture/testing docs, risk-triggered Spec, durable decisions, review/governance skills, deterministic gates, postmortems, manifest state, and the upgrade channel. Optional capabilities include ci, release, community, codeowners, ai-disclosure, monorepo subtree instructions, and local hooks.
Assessment timing follows urgency. Security, release, migration, or ownership boundaries that affect the current task are raised before implementation. Ordinary CI/community/maturity advice is raised at handoff. Record enabled, external, deferred, or declined; store the normalized assessment hash for deferred/declined advice so unchanged facts do not trigger another prompt.
Use the legacy --extensions flag for file-backed compatibility capabilities, or explicit --capability-state assignments for the catalog state. Omission never disables an enabled capability; removal is a separate, explicitly authorized operation.
CODEOWNERS handle: when the codeowners pack is enabled, ask the user for the GitHub handle or team to own the seeded paths (see interview Q10). Never invent a handle. If the user provides none, the scaffold generates @TODO-OWNER (the CODEOWNER_HANDLE default) and you must tell the user to replace it before enabling branch protection that requires owner review.
Unmanaged repository adoption
When the manifest is absent, the skill may perform a read-only audit and show a dry-run. Do not write until the user approves adoption. --adopt discovers existing AGENTS, architecture/testing docs, ADR/RFC, postmortems, CI, and hook systems; compatible non-standard paths are recorded in governance.paths, while external sources of truth are recorded in governance.externalSources and capability state. Existing files are user-owned. Fill only missing Core surfaces and never manufacture a second authoritative copy.
Jira, Notion, GitHub Issues, and other requirement systems use generic stable links first; connector setup is a separate future capability and always requires authorization. The adoption order is read-only audit → dry-run → user confirmation → incremental write → run-gates verification.
Authoring rules
These rules govern repo-seed's own instruction files (SKILL.md, templates, in-repo skill templates) and keep the skill portable across agent tools:
- Use tool-agnostic wording for actions: "ask the user", "read the repository", "write files", "run a command" — never name a specific tool or function (e.g. no "use the question tool", "call read_file").
- Describe intent and outcome, not the mechanism; each agent maps the wording to its own tools.
- Keep commands that are genuinely cross-tool (shell snippets, node scripts) as literal code blocks; they are tool-independent.
- The
allowed-toolsfrontmatter field is a declarative permission list per the agentskills.io spec, not an instruction; keep it minimal.
References
- references/interview.md — question bank.
- references/update-strategy.md — ownership, conflicts, preservation, manifest schema.
- references/decision-standard.md — the MADR + Class extension standard.
- references/review-standard.md — how to derive the project-specific repo-review policy.
- references/doc-standard.md — the documentation standard the seed writes.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

