A community-driven registry for Claude, Cursor, Windsurf, Cline & more. Not affiliated with Anthropic.
Are you the author? Sign in to claim
The Atlas series — prompt skills that keep AI agents honest about the goal. atlas-contract freezes tasks into auditable
Prompt skills that keep a capable AI agent honest about the goal — so a strong model stays accountable to what you actually asked for, instead of quietly trading it for an easier or "more sensible" version along the way.
中文说明 → README.zh-CN.md · Examples → EXAMPLES.md
Atlas is not a crutch for weak models, and it is not about code quality. A strong model with a decent CLAUDE.md already writes good code — and models keep getting better, so that floor only rises.
What a stronger model does not fix — and arguably makes worse — is this: deep into a task, a capable agent makes a stream of decisions on your behalf. It substitutes an approach, narrows a scope, picks a default, "improves" something next door. The smarter the model, the more plausible those decisions look, which makes them easier to wave through without noticing — and harder to catch when one quietly diverges from what you wanted. You end up betting that it chooses right every single time, on every decision you never saw.
Atlas doesn't make the agent smarter. It makes those decisions, shortcuts, and unverified claims impossible to hide — so you can trust a strong model on long, high-stakes work where this kind of silent drift actually happens. See EXAMPLES.md for a real run where both versions wrote good code, but only one let the user stay in the loop on a decision that changed the original request.
It swapped in another approach without asking. This time it was a good one — but can it really pick right in every case? That's just probability; you're betting it chooses correctly on its own. The other one lets you choose — while still telling you what it intended to do.
它换了种方式,但没经过同意。如果是好的方式,确实没问题——但它真能在每种情况下都自主选对吗?其实就是个概率,去赌它自己选对。第二个让你自己选,同时还给出了它自己的打算。
A silent goal change rarely feels like betrayal from the inside. It feels like progress — like fixing the build, like a harmless simplification, like a sensible substitution. So asking the model "are you being honest?" doesn't help: that question goes to the same judgment that already drifted.
You don't fix this by making the model smarter. You fix it by changing the process around it: freeze the goal up front, surface the decisions that change it, and force every deviation to become an auditable event you can see and approve.
| Skill | Role | When it runs |
|---|---|---|
| atlas-contract | In-conversation governance. Freezes the request into an auditable Goal Contract, scales its footprint to task complexity (Light / Medium / Heavy), stops on the action that changes the goal (substitute, narrow, mock, delete, weaken a test) instead of asking the model to judge its own risk, and runs an adversarial Final Audit that marks anything unproven as Unverified — never "done". | Every non-trivial coding task. |
| atlas-ledger | Cross-session memory (compounding). When a drift is caught, it distills the lesson into a reusable project clause (WHEN / DON'T / INSTEAD) and, after you confirm, records it in Atlas.md. | Only after a caught drift. |
How they compound: atlas-contract enforces the goal in the moment. atlas-ledger turns each caught drift into a permanent clause in Atlas.md. The next time atlas-contract builds a contract, it loads the relevant clauses as guardrails — so the project's defense line thickens with every catch, instead of repeating the same mistake.
atlas-contract ──catches a drift──▶ atlas-ledger ──writes clause──▶ Atlas.md
▲ │
└──────────────── loads clauses into the next contract ◀────────────┘
On a non-trivial task, the agent freezes the goal and stops for your confirmation before touching code — so any decision that changes your request surfaces before it's acted on, not after:
Atlas Goal Contract
Goal: Add a live news headline per conversion, fetched from a real source.
Must Do:
- [M1] Fetch a real headline at request time (hard)
Must Not Do:
- [N1] No hardcoded, mocked, faked, or placeholder headlines (hard)
Preserve:
- [P1] Existing conversion / rate / history / delete must keep working (hard)
Summary: I'll fetch real headlines per currency and wire them in, without
breaking any existing behavior.
ATLAS_STOP: awaiting your confirmation before starting.
See EXAMPLES.md for the full before/after run.
Each skill is a single SKILL.md. Copy its contents into:
.md skill file in the workspace.cursorrulesNo installation, no dependencies. Start with atlas-contract; add atlas-ledger once you want the project to remember its caught mistakes.
atlas-contract/SKILL.md in-conversation goal governance (start here)
atlas-ledger/SKILL.md compounding memory of caught drift → Atlas.md
EXAMPLES.md real before/after runs
LICENSE MIT
These skills are enforced by the same model they govern. They raise the floor of goal-fidelity and make silent drift structurally harder — but a sufficiently drifted model can still produce a clean-looking audit over incomplete work, because the adversarial pass is also self-run. For high-stakes or long-running work, a code-layer mechanical gate — one that compares tool actions against the contract before they execute, without asking the model to judge — is the external backstop a prompt cannot provide by itself. That layer is atlas-agent. Treat the Atlas skills as necessary layers, not a complete solution.
A practical approach to managing multiple AI agents in Cursor through strict file-tree partitioning and domain boundarie
📄 Configuration files that enhance Cursor AI editor experience with custom rules and behaviors
Cursor AI 编程规则精选集 | 132+ 规则，覆盖前端/后端/AI/DevOps 等 32 个领域