Gh body check
Audit a drafted or filed GitHub issue / PR body against gh-body-conventions via a fresh-context subagent. Any unresolved ⚠ blocks the caller.From its SKILL.md
npx -y skills add ultimatile/development-skills --skill gh-body-checkAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
5.6 KB, ~1.3k tokens by cl100k_base, as published. Nobody here has run it
GH Body Check
Two checks: a mechanical math scan (Unicode-math glyphs, the GitHub-unsupported macro \operatorname, and inline math neutralized by an enclosing code span), and a cold-reader audit delegated to a fresh-context subagent.
Why a cold-reader subagent
The author has just drafted the text. They read what they meant, not what the text literally says. A fresh-context subagent with no access to the chat history, the plan, or the author's notes is the cleanest implementation of the operative definition of leakage: a token whose referent cannot be resolved from the target repo's public state. Whatever the cold reader cannot resolve is, by definition, leakage.
Hard-wrap and sub-clause line endings are out of scope here: gh-post's detect_hardwrap rejects hard-wrap at submission, and GitHub's renderer (plus gh-post's auto-format) collapses soft breaks to spaces, so source-side sub-clause shape has no wire or render consequence.
Procedure
1. Prepare the body and target metadata
Write the body to a temp file:
BODY_FILE=$(mktemp -t gh-body-check-XXXXXX.md)
cat > "$BODY_FILE" <<'EOF'
<the drafted body, exactly as it will be filed>
EOF
Determine: artifact kind (issue / pr), target repo (e.g., owner/repo — the cold reader needs this to reason about what counts as "public" for that repo), target language (English / Japanese / matches-repo).
2. Math scan
${CLAUDE_SKILL_DIR}/body-math-scan.sh "$BODY_FILE"
Exit 0 = clean, 1 = hits found (printed as line:match), 2 = usage / environment error. It flags the raw Unicode math glyphs, the \operatorname macro, and inline math $`...`$ neutralized by an enclosing code span — all forbidden by gh-body-conventions § Math. For a Unicode-glyph or \operatorname hit: any hit → ⚠; a hit inside a fenced code block, an inline code span, or prose that merely names the construct → ⊘ N/A, judged by main-context inspection. A code-span-neutralized inline-math hit is NOT auto-dismissed by that inline-code-span exemption — the enclosing code span IS the defect — so judge intent in main context: math a copied display form silently neutralized → ⚠ (fix); a legitimate literal $`...`$ shown as code or data → ⊘ N/A with a one-line justification.
3. Cold-reader audit (fresh-context subagent)
Invoke Agent with subagent_type: "general-purpose". Pass only the body, the target repo name, and the artifact kind. Do NOT pass chat history, the plan, the author's prior messages, or any context about why the body is being filed — the fresh context is the entire point.
Prompt template:
You are an external reader of <target-repo>. Your public knowledge
consists of: this repository's README, public issues and PRs,
public code, and well-known external standards (RFCs, arXiv,
language specs). You have no access to: chat history, private
notes, private workflows, local files, or the author's mental
model.
Read the following <issue|PR> body and list every place where a
referent cannot be resolved from that public knowledge. For each
hit, return:
- Quote (the phrase or line, verbatim).
- Why it cannot be resolved from public knowledge.
- Suggested public source to check (if any).
Out of scope: formatting, grammar, math notation, line width,
sub-clause line breaks. Focus purely on "can an external reader
resolve every referent here?"
Do NOT browse the repo or run tools. Judge from the body text alone.
--- body ---
<body content>
--- end body ---
Common leak shapes the cold reader will surface (illustrative, not exhaustive — the operative definition above is what governs):
- Local filesystem paths (
/Users/...,~/foo). - Private skill / workflow names (
/research-and-implement, etc.). - Phase / Step numbering from the working session, unless the artifact is itself a sub-issue / sub-PR of a public umbrella.
- Chat-tone scaffolding ("as we discussed", "following up on chat").
- Unresolved placeholders (
<TODO>,<owner>). - Private project nicknames or unsanctioned shorthand.
- Inline Japanese clauses in an otherwise-English body.
4. Merge and gate
Combine the rg hit (if any) and the cold-reader report into a single status. For each cold-reader ⚠, judge in main context — the finding-triage SSOT's actionable / false-positive split applied to a cold-reader concern:
- True positive (
actionable) — fix before proceeding. - False positive due to missing context — record explicitly why (e.g., the cold reader did not recognize a public external reference, or the term is a standard library identifier the reader was unfamiliar with). Per
finding-triage, false-positive classification is itself a triage step the user can challenge; do not silently override.
Any unresolved ⚠ blocks the caller's next step. Return the report; the caller revises the draft and re-runs gh-body-check. Iterate until clean, or each remaining ⚠ has an inline waiver with a one-line justification.
What this skill does NOT do
Does not draft or file the body (caller's job). Does not maintain the rule set (gh-body-conventions is SSOT — update it first, then add the corresponding check here if a new mechanical rule is needed). Does not check hard-wrap (delegated upstream to gh-post's detect_hardwrap) or reference anchoring (raw line numbers, broken issue refs — a separate concern that may live in a future reference-validity tool).
What ships with it: 2 files
14.0 KB alongside SKILL.md, 2 of them executable
- body-math-scan.shruns6.3 KB
- test-body-math-scan.shruns7.7 KB