Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add chgagne/claude-skills-research --skill verifying-proofsgit clone --depth 1 https://github.com/chgagne/claude-skills-researchWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/chgagne/claude-skills-research/verifying-proofs)<a href="https://agentmods.dev/skills/chgagne/claude-skills-research/verifying-proofs"><img src="https://agentmods.dev/badge/skills/chgagne/claude-skills-research/verifying-proofs/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/chgagne/claude-skills-research/verifying-proofs"><img src="https://agentmods.dev/badge/skills/chgagne/claude-skills-research/verifying-proofs.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00088 | $0.01778 |
| Opus 5.5 | $0.00035 | $0.00711 |
| Sonnet 5.5 | $0.00018 | $0.00356 |
| Haiku 4.5 | $0.00009 | $0.00178 |
Grade A, and why
verifying-proofs scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured today.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 116 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Verifying Proofs
You are the checker. Measured against the same agent without this skill, careful reading found planted inequality, algebra and dropped-hypothesis errors as often as the skill did, and real slips in unplanted papers at the same rate. What this skill adds is a method for the reading, a hygiene script, and rules that keep reports honest.
Read every load-bearing proof
The proofs the contribution rests on, not all of them. For each:
- Every displayed chain, one relation at a time. Each
=,\le,<must follow from the line before it plus something stated. Middle relations of long chains are where slips hide: a lost factor, a constant halved, a square dropped,\lewhere only\geholds. - Every hypothesis of the statement: where is it used? Find the step that needs it. Then the
reverse: for every step that needs a condition (a sign, a range, a vanishing trace, a
positivity, a "for all large N"), find where it was assumed. A condition used but never stated,
or stated in the main text and lost in the appendix restatement, is the most common real
defect. Work
reference/structural-audit.md(hypothesis use, quantifier order, induction coverage, inequality direction). - Every cited result: are its hypotheses met where it is applied? A theorem applied at one
width when it needs all large widths, or at
\beta \ne 0when it needs\beta > 0.
Back each finding with a witness. A counterexample with concrete values satisfying every stated hypothesis, or the one-line derivation of the correct form. For polynomial arithmetic, a short Z3 or SymPy script settles it; write the values into the report.
Certify serious findings in Lean (default when Lean is available)
Check for Lean first: python3 ~/.claude/skills/verifying-proofs/assets/lean-certify.py --check
(the project is $LEAN_PROJECT, else ~/.lean-project). If it is unavailable, say so in the report
and go on; never install it on your own. If it is available, then for every finding of the form
"the step is false", up to five per paper, most consequential first:
What ships with it
36 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
- assets/lean-certify.py 7.0 KB runs code
- assets/proofcheck/__init__.py 0 B runs code
- assets/proofcheck/__main__.py 12 KB runs code
- assets/proofcheck/compose.py 19 KB runs code
- assets/proofcheck/engines/__init__.py 1.1 KB runs code
- assets/proofcheck/engines/gradient.py 4.9 KB runs code
- assets/proofcheck/engines/rational.py 7.7 KB runs code
- assets/proofcheck/engines/smt.py 6.2 KB runs code
- assets/proofcheck/engines/symbolic.py 7.2 KB runs code
- assets/proofcheck/ledger_io.py 5.7 KB runs code
- assets/proofcheck/report.py 8.0 KB runs code
- assets/proofcheck/sandbox.py 7.3 KB runs code
- assets/proofcheck/stubs.py 15 KB runs code
- assets/proofcheck/twotrans.py 7.8 KB runs code
- assets/run-proofcheck.py 655 B runs code
- assets/tests/test_adam_refutation.py 11 KB runs code
- assets/tests/test_compose.py 20 KB runs code
- assets/tests/test_gradient_on_a_real_claim.py 7.1 KB runs code
- assets/tests/test_gradient.py 3.8 KB runs code
- assets/tests/test_lean_certify.py 2.8 KB runs code
- assets/tests/test_rational.py 6.2 KB runs code
- assets/tests/test_report.py 7.6 KB runs code
- assets/tests/test_sandbox.py 6.0 KB runs code
- assets/tests/test_seeded_errors.py 13 KB runs code
- assets/tests/test_smt.py 5.9 KB runs code
- assets/tests/test_stdlib_only.py 3.3 KB runs code
- assets/tests/test_stubs.py 13 KB runs code
- assets/tests/test_supersession.py 7.9 KB runs code
- assets/tests/test_symbols_template.py 3.4 KB runs code
- assets/tests/test_twotrans.py 6.2 KB runs code
- reference/design-notes.md 25 KB
- reference/engines.md 5.2 KB
- reference/false-alarms.md 26 KB
- reference/lean.md 3.4 KB
- reference/step-ledger.md 5.3 KB
- reference/structural-audit.md 5.6 KB
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- today Changed · -333 lines 21a4d8c8e5b8
- 1mo ago First seen · 449 lines · 88 tokens per session scan A 9bf45ff47a8f
verifying-proofs is a skill published in the GitHub repository chgagne/claude-skills-research (5 stars, last pushed today), licensed MIT. It adds 88 tokens to every session and 1,778 once invoked, about $0.0004 per session on Opus 5.5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
paper-writer
Medical/scientific paper writing workflow skill. Manages the full pipeline from literature search to submission-ready manuscript. Creates and manages a project directory with IMRAD-format section files, literature matrix, reference management, and quality checklists. Supports both English and Japanese papers.…
food-research
Run a comprehensive, multi-source literature and evidence-synthesis workflow for food & nutrition science. Use when the user wants to research a food/nutrition topic in depth, do a literature review, build an evidence brief, screen and synthesize many sources, verify citations, or scope a systematic review.…
food-deep-research
General-purpose deep research that produces a fully written, source-validated literature review on any question: scope it, design the method, discover and screen sources by journal ranking, validate every source, extract and verify evidence, synthesize, stress-test, then write and format the review (APA 7.0 by…
food-figure
Comprehensive figure system for food & nutrition manuscripts: analyzes the user's data, recommends the best figure(s) to make, then produces submission-grade graphics in Python or R at the target journal's spec. Handles all common scientific figure types (bar/box/violin, line/kinetic, scatter/regression, Bland–Altman…
food-paper
Multi-subagent manuscript system for food & nutrition science covering the whole research process: understand the field, frame research questions, curate and analyze data, run statistics, build figures and tables, construct the discussion, draft, polish, and self-review — journal-aware throughout. Includes a…
food-pipeline
Master orchestrator for the whole food & nutrition research-to-publication workflow. Coordinates the specialist skills — each with its own subagent set — into one governed path: journal selection, research (food-research / food-deep-research), writing & analysis (food-paper), figures (food-figure), peer review…