math-proof is a Claude Code plugin for hard, research-level mathematics problems, and it has two skills. `solo` has Claude work the problem alone in one long turn. `siege` runs hours of rounds: a judge writes questions, independent workers answer them, and the judge keeps a ledger. Both end in a self-contained proof.md, and both are written for people who want an unattended, auditable attempt at a single problem.
Why read it
Both the workers and the solo skill are told that each message has a limited output budget, which also covers private reasoning. Early in every message they must write down their plan and the exact statement they are attempting, then add each intermediate result to a file as soon as they have it. If a response is cut off at the output limit, only the reasoning that was never written is lost.
Each worker must finish its answer file with an exact line, `=== END OF ANSWER {Q} ===`. A script checks for that line and sorts each answer as answered, partial or no answer, so the orchestrator never has to read the content to judge it. Partial files are renamed and passed to the re-run worker as leads it may use but must not trust.
The top-level session only runs the protocol. It may not read worker answers, judge them, or summarize the mathematics in its own words. It also may not add encouragement or hints to the briefs. This keeps the reasoning inside fresh sub-agents and stops paraphrase drift from creeping in between steps.
Workers see only their query file plus the problem statement. The judge therefore has to put every earlier result a query depends on inside the query itself. It copies that text in with cat or sed rather than retyping it, which avoids transcription slips.
Claims go into a numbered ledger marked PROVED, REFUTED or OPEN. A claim is withdrawn only by adding an explicit RETRACT line, never by editing. The goal is tagged [GOAL], and the run can conclude only after two [AUDIT] lines from separate verify queries certify it. A Retrieval rule makes the judge go back to the original answer files instead of relying on the summary's paraphrase.
From round ESC_ROUND on, the summary must contain a chain that reduces the goal to exactly ONE unproved statement. Most of each round's queries then attack that statement, three of them as identical independent copies. The skill also defines strictly what counts as progress, so the run cannot drift into easier special cases.
proof.md must stand on its own for a referee who cannot open any other file. After finalizing, the orchestrator greps proof.md for references to run files. If it finds any, it relaunches the finalize step once with the matches included.
Both skills spell out how to tell settings like MAX_ROUNDS=8 apart from mathematics like "n=3", and how to tell a file path from a problem written as text. If a run stops, giving the same command again resumes it from the state it saved and never redoes a step whose output already exists.
A model read each skill to find these. Top to bottom is the usual order. Select a box to read the skill, or a line to see the text behind it.
| SKILL | KIND | STARTED BY | DESCRIPTION | LINKS |
|---|---|---|---|---|
| math-proof-judge | agent | The agent | The math-proof judge: it plans each round's questions, reads the answers, keeps the summary and ledger, and writes up, revises and finalizes proof.md. It does one step per launch and is launched only by the math-proof plugin's siege skill. | 2 |
| math-proof-worker | agent | The agent | A math-proof worker: it answers one self-contained question from a file, writing its answer to a file as it reasons. It has no memory between questions and is launched only by the math-proof plugin's siege skill. | 1 |
| math-proof-worker-deep | agent | The agent | The higher-effort math-proof worker, used from the escalation round on (round 4, unless the siege setting ESC_ROUND says otherwise). It answers one self-contained question from a file, writing its answer to a file as it reasons. It has no memory between questions and is launched only by the math-proof plugin's siege skill. | 2 |
| siege | skill | Only you | Work on one hard mathematics problem in rounds: each round a judge writes a few self-contained questions, independent workers answer them, and the judge keeps a ledger of what is proved, refuted and open, until the judge concludes or the rounds run out, and then a proof.md says plainly what is and is not proved. A run takes hours and dozens of worker runs. Usage: /math-proof:siege [NAME=value settings] <the problem, stated in full, or the path of a file holding it>. | 3 |
| solo | skill | Only you | Work on one hard mathematics problem in this session yourself, with no sub-agents: reason in stages, record each settled step in a notes file so that nothing written is lost if a response is cut off, and end with a self-contained proof.md. Usage: /math-proof:solo <the problem, stated in full, or the path of a file holding it>. | 1 |
From anthropics/claude-plugins-official at a7ba5b9, 2026-10-30. Skills read by DeepSeek V4.1 Flash on 2026-11-04. Summary and techniques by Claude Opus 5.5 on 2026-11-06.