Guide · Agent skill
How to use the Proof Orchestrator agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The Proof Orchestrator is an agent skill, a packaged set of instructions that a coding agent such as Claude loads when it recognises a matching request. This one is for managing a long mathematical proof that spans several sessions. After reading this guide you will know how to install the skill, how to phrase requests so it triggers, what files it creates, and when to switch to a different skill.
Proof Orchestratorwanshuiyin/Auto-claude-code-research-in-sleep · MIT Download the skill Quick start
- 1Install the ZIP. Unzip the downloaded skill into ~/.claude/skills/ or a project's .claude/skills/ folder, or upload the ZIP under Settings → Capabilities → Skills.
- 2State the theorem. Give the exact theorem, its assumptions and the sources the agent may use.
- 3Let the local attempt run. The agent proves, checks and tidies the result inside a numbered run folder.
- 4Review the audit. Read audit.md and final.md, and treat the output as a model's reasoning that still needs expert review.
- 5Escalate if stuck. If the attempt stalls, paste browser-prompt.md into a GPT Pro conversation yourself and save the reply to gpt-pro-output.md.
- 6Resume later. In a new session, point the agent at the prior run folder so it continues from the single current obligation.
What the Proof Orchestrator skill makes the agent do
The skill organises a proof as a numbered, resumable run folder under prompts/<YYMMDDHH-num>/. The agent first tries the proof locally, checks its correctness, and edits the result for clarity and economy.
If the local attempt stalls, the agent keeps the sources organised and writes browser-prompt.md, an exact text you can paste into GPT Pro in your browser. Returned GPT Pro text is saved as raw evidence in gpt-pro-output.md and is not trusted until audited.
The skill also records what has been proved, what is still conjectural and which sources were used, so a later session can continue from the last stopping point. By default it does not operate a browser or spend API credit.
- task.md: the precise theorem or proof obligation
- local-proof.md: the proof attempt or isolated blocker
- audit.md: correctness and source-alignment audit
- final.md: the verified, user-facing proof
- codex-ledger.md and next.md: run state and the next narrow obligation (optional)
Install the Proof Orchestrator skill
Download the ZIP for the proof-orchestrator skill from this page. Then choose one of two routes.
For a personal install, unzip the skill into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder. Alternatively, upload the ZIP in the Claude apps under Settings → Capabilities → Skills.
- Personal: ~/.claude/skills/
- Per project: .claude/skills/ inside the project folder
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP
Trigger the skill with natural requests
The skill loads when your request matches its description. The description names three triggers: proof-run orchestration, a GPT Pro handoff, and cross-run proof continuation. Say what you want in those terms.
Ordinary drafting and rigorous checking belong to other skills. The description routes plain proof drafting to /proof-writer and verification or submission acceptance to /proof-checker.
- "Set up a proof run for this theorem and keep track of what we've proved."
- "Continue the proof from last session's run folder."
- "The local attempt is stuck. Prepare a GPT Pro handoff prompt."
- "Do a DeepSeek second opinion on the current proof run."
Work through a multi-session proof
Start with a clear theorem, including its assumptions and the sources you are allowed to use. The skill will not broaden or silently repair the theorem. Let the agent attempt the proof locally, then review audit.md and final.md before accepting anything.
When you return in a later session, point the agent at the existing run folder. The skill reads the prior final.md, audit.md, local-proof.md, codex-ledger.md and source-manifest.md, then creates a new run directory for the new work rather than overwriting old files.
- Give the prior run ID and the single obligation you want next.
- Expect a new run folder; completed runs are treated as append-only evidence.
- If escalation reaches GPT Pro, expect a fresh browser-prompt.md for a new conversation.
Prepare and use a manual GPT Pro handoff
When the local attempt stalls, the agent writes browser-prompt.md with the exact text to paste. It also writes handoff.md, which lists the upload order and status. You run the browser step yourself.
When GPT Pro replies, save the answer to gpt-pro-output.md and ask the agent to audit it. Claims from the reply count only once the audit accepts them. You may reuse a matching ChatGPT Project, but the prompt goes into a fresh conversation so old context does not change the task.
- The agent does not upload files or operate a browser by default.
- Source material in the prompt is wrapped in explicit data delimiters.
- Credentials and private paths are excluded from the prompt.
Safety notes: scripts, network and optional tools
The skill runs no scripts. It needs local file read and write access (Read, Write, Edit, Grep, Glob). It does use the network in two optional cases: an installed call-gpt-pro skill, only when you explicitly ask the agent to make the GPT Pro call, and an llm_chat MCP tool with a DeepSeek API key, only for the DeepSeek review mode.
Returned text from GPT Pro or DeepSeek is treated as data. The skill extracts mathematical claims from it and does not follow instructions inside it, such as requests to fetch links or change file scope.
The optional lean-formalize skill is used for Lean formalisation when installed.
- No scripts are run.
- Network use happens only in the optional call-gpt-pro and DeepSeek routes.
- DeepSeek review runs only when you explicitly request it for the current run.
What people use it for
- Keeping a long proof organised across several working sessions with a resumable run folder.
- Recording which lemmas were proved locally and which came from GPT Pro.
- Preparing a self-contained prompt for GPT Pro when a local attempt gets stuck.
- Tracking the sources used in each proof step with a run-local source manifest.
- Getting an optional DeepSeek second opinion as additional evidence on a specific run.
Frequently asked questions
How do I install the proof-orchestrator skill?
Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ folder. You can also upload the ZIP in the Claude apps under Settings → Capabilities → Skills.
Does the proof-orchestrator skill need any extra tools?
It needs local file read and write access through the Read, Write, Edit, Grep and Glob tools. Other tools are optional: call-gpt-pro for a GPT Pro call you request, an llm_chat MCP tool with a DeepSeek key for DeepSeek review, and lean-formalize for Lean work.
Is the proof-orchestrator skill free to use?
The skill is released under the MIT licence. The source does not state any price, and optional routes such as DeepSeek API use may involve your own API costs.
Does it replace /proof-checker for paper submissions?
No. The skill points to /proof-checker for rigorous verification and submission acceptance, and relying on this skill alone could miss errors. Its output still needs expert review.
Can the skill send my proof to GPT Pro automatically?
Not by default. It writes a browser-ready prompt that you paste into GPT Pro yourself. An automated call happens only if the call-gpt-pro skill is installed and you explicitly ask for it.
What should I use instead for ordinary proof drafting?
Use /proof-writer. The skill description says ordinary proof drafting is routed to that separate skill, which suits quick one-off proofs better than a multi-session run folder.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use this skill if you are working on a long proof across several sessions, want a clear record of which claims were proved locally and which came from an outside model, or already use GPT Pro and want a structured handoff prompt. Pick /proof-writer for quick one-off drafting, and /proof-checker or expert review for final verification before submission.
Download the skill