Skip to content

Source Audit

Marketing source-audit Source

Install: /plugin install marketing-skills@claude-code-skills

Audit a finished draft before it is published: for every cited link, does the source actually support the claim the draft makes next to it?

Do only this. Typos, style, structure and tone belong to other skills; mixing them in hides the one failure that matters, a claim the source never made.

Before Starting

Ask only what the draft cannot answer: the draft path, the publish target (blog, newsletter, docs, social long-form; it sets how strict to be) and the scope (full audit, or only links the author flags). If a link cannot be opened from this environment, say so up front and plan to mark it unreachable.

Modes

  • Full audit (default): extract, fetch, compare, write the report.
  • Risk-first: only links that carry checkable facts (versions, prices, licenses, platform support, feature lists); list what was skipped.
  • Single link: one verdict and its quote, no report.

Workflow

  1. Extract links and context.
    python3 scripts/extract_claims.py draft.md --json -o claims.json
    
    Each item has url, link_text and the paragraph context. If link_count is 0 but links are visible, run --sample to confirm the extractor works. List the checkable claims per context; mark subjective ones opinion. Read one paragraph up when context is thin, but never attribute the whole article to one link.
  2. Fetch every source. Never decide from a search snippet or memory; projects change. Fetch rules per source type are in references/verdict-examples.md. Never issue a verdict on an unreachable source.
  3. Compare claim by claim. Every verdict needs a quoted passage that locates the evidence.
Verdict Meaning
supported The source explicitly backs every key part of the claim.
partial The source backs part of it, or the wording is too vague to confirm.
contradicted The source explicitly says something incompatible.
not-found The fact is absent from the source. Absence is not denial.
unreachable Fetch failed after a retry.
opinion Subjective; not audited.

Borderline cases (several links, combined facts, moved facts, summaries, vendor copy) follow the decision rules in the reference. 4. Write the report to <draft>-audit.md using the reference template, and give a summary table in the reply.

Example

Draft: "Acme CLI supports Windows, macOS and Linux, and runs on 40 languages."

Claim Source Evidence Verdict
Windows, macOS and Linux repo README "Windows 10 2004+, macOS 12+, Ubuntu 22.04+" supported
40 languages repo README no language section found not-found

The second row is the common defect: not wrong, simply unsourced. Fix it by removing the number or citing a page that states it.

Quality Bar

  • every supported and contradicted verdict has a quote;
  • every not-found names what was searched for;
  • every unreachable names the failure (status code or timeout);
  • opinion claims are listed as skipped, not silently removed;
  • the fix list can be applied without re-reading the audit.

Anti-Patterns

  • A verdict without a quote, or judged from memory or a search snippet.
  • Treating absence as contradiction.
  • Auditing the whole article against one link.
  • Editing the draft or copy-editing inside the audit; the deliverable is a report.
  • research-summarizer: starts from sources and summarises; not for auditing a finished draft.
  • meta-research: tests whether a hypothesis is true; not per-link verification.
  • copywriting / md-review: use after the audit for wording and structure.