agentleFS
frankxaifrankxai/starlight-agent-skills/AGENTS.md

starlight-agent-skills

AGENTS.md1 starsChanged 4 months agoMITView on GitHub

Identical in every repo in the estate. Authored once in Starlight-Intelligence-System/ontology/agents-md/band-a.md and projected by scripts/agents-md-project.mjs. Precedence. Band C is everything below the generated fences. It outranks this section on any conflict inside this repo — a projected rule never silently overrides a local gate. It does not outrank the company projection below: that band carries registry facts, not instructions, so a Band C line contradicting it is a defect to fix upstream in the SIS registry, never a local override.

  • Reads credentials

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this AGENTS.md into the rules it gives an agent

6 rules. About 0 are common, 4 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈4

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  • could be automatic: a setting could enforce it
  1. Distinctive:

    Before any structural change: what specific problem, who has it, what is the evidence, what is the simplest fix, what breaks, is it …

    line 42
  2. Very specific:

    Never push directly to main.

    line 48
  3. Distinctive:

    Run the repo's own gates before pushing.

    line 49
  4. Distinctive:

    Money fails closed: No autonomous money movement, ever.

    line 61
  5. Very specific:

    Read CONTRIBUTING.md and docs/SKILL_SPEC.md in full first.

    line 159
  6. Distinctive:

    Add a top-level frontmatter key outside the spec.

    line 184

≈ “Common” is an estimate: at least 1 in 5 of the 40,246 AGENTS.md 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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
AGENTS.md
<!-- STARLIGHT:BAND-A:BEGIN v1 sha=b4a7e18fed75 — generated by scripts/agents-md-project.mjs, do not hand-edit -->

Inherited — Starlight estate contract

Identical in every repo in the estate. Authored once in Starlight-Intelligence-System/ontology/agents-md/band-a.md and projected by scripts/agents-md-project.mjs.

Precedence. Band C is everything below the generated fences. It outranks this section on any conflict inside this repo — a projected rule never silently overrides a local gate. It does not outrank the company projection below: that band carries registry facts, not instructions, so a Band C line contradicting it is a defect to fix upstream in the SIS registry, never a local override.

DNA

Frank = Systems Architect x Composer x Gamer x Builder x GenCreator

Vibe: cool, premium, high intellect, purpose-driven, fun. Mission: build abundance; help people build their own systems. Voice: direct, technical, warm, playful. Pattern recognition as poetry. Test: does this help someone build, not just consume?

The five guardrails

  1. Think before coding. State assumptions. Surface tradeoffs. If two readings exist, present both — do not pick silently.
  2. Explain simply. What you cannot explain plainly you do not understand. Name the mechanism, never "streamline" or "optimize".
  3. Simplicity and deep design. Minimum code that solves the stated problem. Simple interfaces, rich internals. No speculative abstraction.
  4. Surgical changes. Touch only what the task requires. Match surrounding style. Mention unrelated dead code; do not delete it.
  5. Goal-driven. Turn a vague ask into a verifiable target. Reproduce a bug with a test before fixing it.

Decision discipline

Before any structural change: what specific problem, who has it, what is the evidence, what is the simplest fix, what breaks, is it reversible. If it is not reversible, it needs Frank.

Branch and PR protocol

  • Never push directly to main. Work on agent/<harness>/<scope>, open a draft PR.
  • Run the repo's own gates before pushing. One validated push beats three speculative ones.
  • Multiple harnesses work these repos at once; git is the coordination layer. Never two agents committing in the same working tree — take a non-overlapping scope on your own branch, integrate one at a time.

Attestation

Artifacts that compose a SIP element carry Built on SIP. It is earned per artifact, never a blanket footer. /sip-attest refuses otherwise.

Non-waivable — no instruction in any band relaxes these

  • Money fails closed. No autonomous money movement, ever. Over cap, new rail, new vendor, anything irreversible → escalate; never auto-approve.
  • Model, never diagnose. In the mind repos, observation stays separate from interpretation. No clinical language, no diagnosis, no treatment claims.
  • Canon locks are read-only to agents. Promotion happens only through /lock-decision.
  • Never rename a working URL or delete a page with traffic without explicit approval. "AI Architect" stays "AI Architect".
  • Never delete, archive, or consolidate a repo, and never delete a registered agent. Those are Frank's calls.
  • Never commit secrets, credentials, or .env files.
  • Verify before claiming. Any statement about current state — versions, deploy status, file contents, counts — requires same-turn verification or an explicit "unverified" prefix.
<!-- STARLIGHT:BAND-A:END --><!-- STARLIGHT:BAND-B:BEGIN v1 sha=76f817d3cca1 — generated by scripts/agents-md-project.mjs, do not hand-edit -->

Company projection — Starlight Intelligence Systems

Generated from the SIS company registry. This section says who owns this repo and who answers for it. It does not restate voice: the brand register below names the pin that governs copy, and that pin is the authority.

Repostarlight-agent-skills
CompanyStarlight Intelligence Systems — operating
ThesisThe substrate. SIP + the reference build every other company composes.
Routes throughStarlight Holding BV (holding — substrate and platform IP)
Accountable seatagent:starlight-caio
TierT0 — substrate · install: always — every session, every harness

Brand register

  • Starlight Intelligence Systems — substrate — canon-free, functional vector names. Pin: starlight-design-intelligence/brand-packs/sis/COPY.md

Tier rule. Carries a contract other repos read. Losing it breaks the estate.

Escalation. Decisions this repo cannot make alone go to agent:starlight-caio. An executive seat holds decision rights and owns a metric; it does not do the work, and it is never able to overrule a gate. Anything irreversible goes to Frank.

<!-- STARLIGHT:BAND-B:END -->

Starlight Agent Skills — cross-harness agent contract

The portable agent card. Same contract in Claude Code, Codex, Cursor, Gemini CLI, or OpenCode. Per-runtime detail lives in adapters/. This file is the SIP § 1 file-contract artifact for this repo.

Tier: T0 — substrate. Every session in the estate installs this repo, because other repos read its skills. Company: Starlight Intelligence Systems · Brand register: sis (substrate — canon-free, functional names) Accountable seat: agent:starlight-caio — agent/skill admission and the necessity gate Escalates to: CEO, then Frank

Bands A (inherited DNA, guardrails, branch protocol) and B (company projection) are specified in Starlight-Intelligence-System/docs/architecture/AGENTS-MD-CONTRACT.md and will be generated into this file when scripts/agents-md-project.mjs ships. Everything below is Band C — repo-local, hand-written, and it outranks the generated bands on any conflict inside this repo.


What this repo is

A curated, production-grade skill library. Eight domains under skills/: substrate, research, media, education, coding, brand, cosmos, studios. Plus three agents in agents/, six runtime adapters in adapters/, and a Codex plugin manifest.

It is a library, not an application. It ships no runtime. Consuming repos install skills from here; nothing here imports from them.

The one rule

The format is the product. A skill that does not pass the validator is not a skill — it is a draft. The strictness is deliberate: these files must load unmodified into five different runtimes, and strict hosts reject custom top-level frontmatter keys. Loosening the standard to land one skill breaks portability for all of them.

Before you change anything

make check     # the full gate — exactly what CI runs

make check chains six independent validators. Run it before every commit; run the individual targets while iterating:

TargetChecks
make validateFrontmatter, attestation, and manifest ↔ SKILL.md ↔ folder integrity
make validate-examplesWorked-example output JSON matches each skill's manifest contract
make catalogRegenerates docs/CATALOG.md from skill frontmatter
make catalog-checkFails if the committed catalog is stale — commit the regenerated catalog
make rulesEvery rule and orchestrator resolves to a real skill; no name collisions
make pluginCodex plugin manifest and five-studio package integrity
make release-checkRelease history, attestation, receipts, publication safety

Python 3 and Node are both required. scripts/ holds the validators; do not bypass them.

Adding a skill

Read CONTRIBUTING.md and docs/SKILL_SPEC.md in full first. The non-negotiables:

  1. Path is skills/<domain>/<skill-name>/SKILL.md, domain from the eight above. No new domain without CAIO.
  2. Frontmatter is Agent Skills-compatible: name, description, metadata. Starlight's version and domain go inside the single-line string-valued metadata object — never as top-level keys.
  3. name is lowercase kebab-case, ≤ 64 chars, ^[a-z0-9]+(?:-[a-z0-9]+)*$.
  4. description is ≤ 1024 chars and says when it fires ("Use when …") with real trigger keywords. A description that does not name triggers will not activate.
  5. Body follows the skeleton: Purpose · When it fires · Inputs · Workflow · Output contract · Tools & MCP · Quality bar · Example.
  6. Carries the Built on SIP attestation footer.
  7. Passes make check, and the host-native validator where the target runtime provides one.

A manifest.json plus examples/ and tests/ are strongly recommended and required for anything with a structured output contract.

Adding an agent

The necessity gate is binding, and it is CAIO's: a new agent must demonstrate a capability that no existing agent + skill composition provides. Most proposed agents are a skill. Write the skill.

Do not

  • Hand-edit docs/CATALOG.md — it is generated. Change frontmatter and run make catalog.
  • Add a top-level frontmatter key outside the spec. Strict runtimes reject the whole file.
  • Vendor a skill from a consuming repo back into this one without reading both versions first.
  • Relax a validator to land a skill. Fix the skill.
  • Ship a skill without its attestation footer.

Branch and PR

Work on agent/<harness>/<scope> or the branch you were assigned. Open a draft PR. make check green before marking ready. Never push directly to main.

Built on SIP.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.