Make the skill installable, and stop it giving a verdict the product retired

[peopleworks]

The skill funnel, finally — plus a defect the move uncovered.

The packaging

The skill has existed since July in skill/signs-of-ai/ and was never installable by any route people
actually use. The instructions were "paste the repo link into your AI harness, or copy the folder".
Meanwhile the ecosystem standardised on a root SKILL.md plus plugin manifests, which is how a
competitor with a tenth of the engine reaches four assistants with one command.

It now lives at the repository root, where every installer looks:

npx skills add peopleworks/SignsofAI -g     # Claude Code, Codex, Gemini CLI, Cursor…

/plugin marketplace add peopleworks/SignsofAI
/plugin install signs-of-ai

New: .claude-plugin/plugin.json, .claude-plugin/marketplace.json, .codex-plugin/plugin.json,
commands/gemini/signs-of-ai.toml. skill/ keeps its README (rewritten) and eval.md.

The defect

Reading it again to move it turned up something worse than the packaging. Its detect mode told the
assistant to give "a plain-language verdict (reads clean / mixed / heavily AI-flavored)" — a
three-band verdict on a scale nobody measured, which is exactly what #32 removed from the report
and the score card on 5 August.

The product stopped saying it. The instructions kept saying it, to an assistant that would repeat them
verbatim in a session none of us will ever see.

Detect mode is now examine mode, and it reports the tells with what they do and do not support.

What else went in

Six rules about what a finding may claim, mirroring the report's own:

1. A finding is a fact about the tool, not about the writer.
2. Finding nothing is not evidence a human wrote it.
3. If you quote the score, quote the error rate — 0 of 90, interval 0%–4.1%.
4. Only English and Spanish have a measured rate. Never borrow one.
5. The boundary was measured on ~3,100-word documents and never validated on a pasted paragraph (The verdict boundary has no length condition, and short passages cross it #59).
6. A tell is not a tally — human academic writing carries a median of seven.

Plus the two hand-offs that need their own warning: compare_to_baseline has no result meaning
"someone else wrote this", and check_originality must show the passages rather than a percentage.

The hand-off section was stale in two ways: six MCP tools listed when there are ten, and the
server configured by absolute path into a build output instead of dnx SignsOfAI.Mcp --yes.

The assistant's own turn is now the first family of tells — "I hope this helps", "As of my last
training update", "¡Por supuesto!" — since the engine learned to score it in #65. The vocabulary
section also now names the seven words it deliberately does not flag, and why.

Guarded, not trusted

SkillDocumentTests checks the numbers in SKILL.md against the calibration this build actually
ships, and that the refusals are still in the text. Re-running the calibration and forgetting the
skill fails the suite. 314 tests, 7 new.

Note

Touches README.md §5, as does #65 (a Credits paragraph). Different hunks, so they merge — but
verify with git merge-base --is-ancestor rather than the badge, as ever.

🤖 Generated with Claude Code

https://claude.ai/code/session_015PEbbiYSNPw7jE3LrPNhyF