Starlight Intelligence
Explore

A place to begin

See where a mission can lead.

Follow a team’s completed repository review and inspect the diagnostics files. Then write your own brief, or bring the prompt below to your coding agent.

Explore a completed mission

Bring the plan into your own repository.

This prompt helps your coding agent draft a team and decide how to review its work. You choose what to keep and what to run.

Choose your coding agent

Paste into a new Claude Code session in the repo you own.

Paste into Claude Code
Build a Starlight open harness in this workspace. You are already a coding agent. Do not invent a new platform. Do not claim live authority, AGI, or a deployed swarm.

Outcome tonight: one human-owned council of three agents that can run one recurring workflow, with a memory limit and a stop condition.

Do this in order:

1. Read https://starlightintelligence.ai/start.json and https://starlightintelligence.academy/academy/departments/start.json
2. Write AGENTS.md: human coordinator, three named roles, veto, irreversible gates.
3. Write MEMORY_MAP.md: none / session / user / org. Private vault prohibited on the public web.
4. Write one SKILL.md for the smallest useful job in this repo. Inputs, outputs, stop conditions, one example, one test.
5. Draw a typed graph: routes_to, verifies, requires_human_gate, failure_routes_to. The maker cannot be the final verifier.
6. Write EVAL_RUBRIC.md: one fluent lie the system must not accept.
7. Run that eval on a synthetic case. Keep the first attempt. Then revise.
8. If a Starlight plugin is available, use design-ai-department then verify-ai-department-package. If not, stay on files in this repo.
9. Stop before send, deploy, spend, secrets, production, or any external effect.

This is self-study and a blueprint. It grants no credential and no runtime tools until a named human says so.

Paste the prompt. Optional plugin: starlightintelligence.academy/academy/install

  1. AGENTS.mdWho acts, who stops, who holds the veto.
  2. MEMORY_MAP.mdWhat persists. What expires. What never leaves the machine.
  3. SKILL.mdOne portable procedure with inputs, outputs, and a stop.
  4. EVAL_RUBRIC.mdThe fluent lie this system must not accept.
  5. RUN_RECEIPT.jsonWhat happened, who decided, what evidence remains.
  1. 01
    Name the human

    The coordinator is a person. Irreversible work returns to a named gate.

  2. 02
    Bound the memory

    Session, user, or org. Never a private vault on a public page.

  3. 03
    Install one skill

    The smallest job worth repeating. One example. One test.

  4. 04
    Compose three agents

    Maker, verifier, steward. Typed routes. Maker is never the final verifier.

  5. 05
    Run one eval

    A fluent answer that is false. Keep the first attempt. Then revise.

Waitlist · Beta · private source

Take a closer look at “done”.

Starlight Observatory is a beta command-line tool for checking an agent’s claims of “done” against its session evidence. Its source is private today. Leave your email for updates about its public release.

We store your email and any answers you give. Nothing is sent yet; when it is, it will be about Starlight Observatory only. Ask us any time to delete your record. What we store and why