gtm-product-page-audit

Files (1)

• SKILL.md

SKILL.md (first lines)

# Product-page audit

Produce one auditable pre-traffic decision for one public e-commerce product
page. Every run creates three separate outputs:

1. a plain-language report for the marketer or founder who needs to make the
   decision;
2. a structured audit for a downstream agent that needs evidence, limits, and
   implementation-ready changes.
3. a full score report with the category scores, criterion ratings, and
   evidence notes.

The three files are different deliverables. The human report explains the
result without machine-shaped tables. The agent audit preserves the evidence
contract. The score report keeps the detailed scoring separate so a person can
read the decision while another agent can recalculate the score.
The human report may summarize the audit, but it may not add a claim that is
absent from it. When the observations are sufficient, the agent audit also
includes a page-quality score covering conversion, purchase confidence, SEO,
GEO, and UX/accessibility/performance. The score describes the page evidence;
it never predicts sales, rankings, or AI citations.

Resolve one explicit `Target URL:` from the invocation prompt or the input. If
both are present, they must identify the same page; a mismatch is missing input,
not a reason to guess. In snapshot mode, use the public snapshots in the input
as the primary evidence. Do not browse or web-search unless the prompt
explicitly allows a check of that exact URL. A URL alone never authorizes
browsing. Do not follow links or inspect another URL. This keeps the result
reproducible under the Skillathon time limit.

## Required input

In snapshot mode, the input names one `Target URL`, access date, product or
merchant context, and a public snapshot with exact excerpts or precise fields.
In live URL mode, the prompt may provide the `Target URL` directly; if an input
is also supplied, its target must match. The input may include one competitor
page, one shopper query, and a small Google guidance snapshot. Public data only.
A source URL proves where an observation came from; a seller claim is not
independent proof of the promised outcome.

The default is snapshot mode: use the supplied snapshot and do not fetch the
web. Live URL mode is allowed only when the prompt explicitly permits checking
the exact `Target URL`; use that page only and keep all other evidence rules.

If the target URL is missing, or neither a usable page snapshot nor an explicit
exact-URL check is available, write all three required outputs with
`NEEDS_INPUT`, name the missing field, and stop without guessing. Use
`INSUFFICIENT_EVIDENCE` when a page is available but the evidence cannot support
the requested decision.

## Procedure

1. Resolve the output paths before doing research. The path named by the prompt
   is the agent-audit path. If no path is supplied, use
   `demo/output/product-page-audit.md`. Unless the prompt names a separate
   human-report path, create `product-page-report.md` beside the agent audit.
   Unless the prompt names a separate score-report path, create
   `product-page-score.md` beside the agent audit. A prompt that names only one
   or two outputs is still completed with all three files.
   If the invocation says to write "both" named outputs, treat the score report
   as the additional skill-required artifact and write it before stopping.
2. Resolve the `Target URL` from the prompt and input, reject a mismatch as
   `NEEDS_INPUT`, then lock the scope to one primary product page. Keep the
   competitor and query optional.
3. Extract page observations from the supplied snapshot or the permitted
   exact-URL check: product name, price and currency, availability, variants,
   brand, identifiers, images, shipping, returns, ratings or reviews, and
   Product JSON-LD when the input provides it. Mark absent fields as `unknown`,
   never as a likely value.
4. Build a short claim ledger. For each material claim, include the exact
   excerpt, source URL, source role, and one status: `verified`, `unsupported`,
   `contradicted`, or `unknown`. Keep first-party claims separate from facts
   supported by independent evidence.
5. Check product-data coverage for title, description, image, link, price,
   availability, condition, brand, identifiers, shipping, returns, and product
   markup. Use the supplied Google guidance as the rubric. Treat structured data
   as an eligibility aid, never as a ranking or display guarantee.
6. Score the observed page with the rubric below. Put the overall score or
   `UNSCORABLE`, audit coverage, provisional or unscorable label, and critical
   blockers in the agent audit. Put the category scores, category coverage,
   criterion ratings, evidence notes in the score report. Keep
   unknown criteria out of quality points, but count them in coverage.
7. Return at most three prioritized changes. A change includes the issue,
   recommendation, evidence, observable success signal, and effort or
   dependency. If the evidence cannot support three changes, return fewer and
   use `INSUFFICIENT_EVIDENCE`.
8. Draft the human report from the finalized agent audit. Do not do a second,
   looser research pass for it. Carry over the decision, strongest evidence,
   fixes, and limits without adding claims.
9. Write the agent audit first, then the score report, then the human report.
   Finish only after all three files exist and the human report points to the
   exact agent-audit path.

## Human report

Apply this built-in unslop pass on every run. The skill must work even when no
separate writing skill is installed.

- Write in the language of the user's request. Keep source excerpts in their
  original wording when precision matters.
- Lead with the decision in plain language. Say what you would do and why.
- Use concrete facts, short paragraphs, and a natural rhythm. Prefer "is" and
  "has" over inflated phrasing.
- Remove hype, generic filler, chatbot phrases, decorative emojis, em dashes,
  title-case headings, curly quotes in ordinary prose, and claims about impact
  that the evidence cannot support. Keep a source quote exact only when its
  wording is the evidence.
- Separate what the page says from what you conclude. Mark interpretations and
  hypotheses in ordinary language instead of presenting them as facts.
- Keep the report concise enough to read in one minute. Include only the
  strongest facts, the score and what it means when available, up to three
  fixes, and the important unknowns.
- If a numeric score is available, use the same rounded overall and category
  values as the finalized audit and score report. If it is `UNSCORABLE`, say
  that plainly.

Use this shape, adapting the wording to the evidence:

```markdown

From teopaleface/gtm-os, MIT.

More from teopaleface/gtm-os