Run a sequential revision experiment on a text and record what the Pangram detector reports for each version, as an observed outcome: frozen editorial plan, one submission per step, real evaluations, preservation checks, chart and report.
Install with the Skills CLI (verifies the checksum, previews the destination, never runs scripts)
skills install @spboucher/revision-detector-experiment@1.0.0 --client claude-codeAdd --project to install into .claude/skills/revision-detector-experiment instead of ~/.claude/skills/revision-detector-experiment.
1Or manually: download the release archive
curl -fsSLo revision-detector-experiment-1.0.0.tar.gz https://skills.spboucher.ai/api/v1/artifacts/spboucher/revision-detector-experiment/versions/1.0.0/download2Verify the SHA-256 checksum
echo "b38cf37c825e6c15514b06120234285bf3ddb0e73e70d432f230708cab9b6bcc revision-detector-experiment-1.0.0.tar.gz" | shasum -a 256 -c -3Inspect the contents before extracting
tar -tzf revision-detector-experiment-1.0.0.tar.gz4Extract into ~/.claude/skills/revision-detector-experiment
mkdir -p ~/.claude/skills/revision-detector-experiment && tar -xzf revision-detector-experiment-1.0.0.tar.gz -C ~/.claude/skills/revision-detector-experimentInstalling only copies or configures. Nothing is executed for you. The Skills CLI verifies the checksum before writing files.
An Agent Skill for running a sequential text-revision experiment in which the Pangram AI-text detector is an observed outcome, never a target.
The agent revises one text step by step, each step for an editorial objective fixed in advance. After every step Pangram evaluates the new version. The skill keeps the plan frozen, takes one revision per step, checks that facts survive, and produces a chart and a report that separate what the detector said from any claim about authorship.
It is a measuring procedure. It is not for making a text pass a detector, and it tells the agent to decline that.
SKILL.md — the rules and the procedure.references/method.md — why each rule exists, the files of an experiment, the assessment, the measures, the stop rule.references/pangram-api.md — what was verified about the Pangram API on 2026-11-10 and what was not.references/report-format.md — the five report files and how to present results.references/default-plan.md — the default ten-step editorial plan.scripts/experiment.mjs — the command-line tool (Node.js 20 or later, no dependencies).scripts/lib/ — the engine, identical to the one in @spboucher/pangram-experiment-mcp.Node.js 20 or later and a Pangram API key in PANGRAM_API_KEY. Each evaluation sends the text to Pangram, a third-party service, and is billed to the key. Do not submit confidential text. A default experiment makes 11 evaluations.
skills install @spboucher/revision-detector-experiment --client claude-code
node scripts/experiment.mjs doctor
node scripts/experiment.mjs init --file original.txt --max 3 --fact "The vote was 5 to 2."
node scripts/experiment.mjs baseline
node scripts/experiment.mjs next
node scripts/experiment.mjs submit --file step01.txt --assessment step01.json
node scripts/experiment.mjs status
node scripts/experiment.mjs report
One text, one run: no general or causal claim. A detector score is not evidence of who wrote a text. Preservation figures are word-overlap proxies; quality scores are the agent's own. The tool guarantees one submission per step; it cannot see how the revision was written. Pangram's error formats, rate limits and billing were not verified. Tested on macOS with Node 26.
MIT.
Select a file to preview it. Files are shown as plain text and never executed.
No advisory findings from static-1.0.0 on Nov 10, 2026. 6 informational notes.
Static checks are advisory. They never certify that a package is safe. Compatibility entries are publisher claims unless marked verified.
| Version | Published | Size | SHA-256 | State |
|---|---|---|---|---|
| 1.0.0latest | Nov 10, 2026 | 33 kB | b38cf37c825e… | published |
Peer-review protocol for LaTeX and Markdown manuscripts: deterministic audit script, five review roles, adversarial verification of findings, anchored rubric, referee report and decision letter templates.
Build, audit and grade Excel model workbooks: layout standard, dependency-free command-line auditor (errors, hard-coded numbers, broken patterns, circular references, controls, stale values) and rubric-based grading aid.
Verify citations against Crossref, OpenAlex, arXiv and doi.org: a DOI must resolve and its metadata must match. Audits .bib files, resolves references without DOI, flags retractions, guides literature searches.