audit-orchestrator

Imported from kavyan256/adobe-group (round3-archive/submission-stale/brand_ai_readiness_audit_submission/brand-ai-readiness-audit/skills/audit-orchestrator/SKILL.md). Install upstream with npx skills add kavyan256/adobe-group --skill audit-orchestrator. Copyright stays with the author (Apache-2.0).

Audit Orchestrator (marketplace entrypoint)

When to use

When someone asks why ChatGPT, Perplexity or Claude doesn't mention their site or gets its facts wrong, why visitors arriving from AI answers don't stay, or simply asks for an AI-readiness audit. Never use it to change a site: everything here is read-only and recommend-only.

Inputs

Work in a directory outside the marketplace folder (WORK/ below).

Procedure

Steps 1 and 3 run scripts; 2 and 4 are yours. audit.py crawls once (scripts/bundle.py, robots.txt to RFC 9309, no JavaScript, truthful User-Agent), runs the four scripted skills as separate processes on the bundle, marks findings behind a total access block as latent, derives severity from scripts/model.py, ranks, and writes one report. The whole audit, agent review included, must finish inside five minutes.

1. Run the scripted audit and keep the bundle. The 100 s crawl deadline leaves the agent review its share of the five minutes.

   python3 skills/audit-orchestrator/scripts/audit.py https://example.com \
       --max-pages 12 --crawl-budget-s 100 --bundle WORK/bundle.json -o WORK/report.json

   A skill that fails or times out lands in checks_skipped[], never silently.
2. Read WORK/report.json: summary.headline, summary.run_status (partial, blocked or failed means findings are a sample, or that access is the whole story), coverage.pages_ok, and checks_skipped[].
3. Run the agent review. It is part of the audit, not an extra. Follow agent-review-audit/SKILL.md: it judges the scripted findings, reviews every candidate problem the scripts cannot measure, adds verified claims and site-specific fixes, and rewrites WORK/report.json as the final report. If it cannot run, say so first when presenting: agent_review.status is not_run and every finding is unjudged script output. The whole invocation of this skill (steps 1-3) should run under an external wall-clock timeout of about 280 s — the caller's own step budget, or wrapping the command in timeout 280 ... — rather than relying on the agent review pacing itself. agent-review-audit checkpoints WORK/report.json after every verdict and every candidate disposition, so a kill at any point during step 3 leaves the last checkpointed report in place: valid and schema-complete, just less refined than an uninterrupted run. Present whatever WORK/report.json holds when you stop.
4. Present the final report in this order: the headline; the top three prioritized_actions[] with their evidence; what latent means (real, but hidden behind an access block until that is fixed) and what confirm_intent means (confirm with the owner before acting); then what was not assessed: limits[], checks_skipped[], agent_review.unjudged_medium_or_above and agent_review.candidate_coverage.missing, so nobody reads "no finding" as "no problem".

Output

One JSON report: the required site, audited_at, summary and findings[] (id, title, severity, evidence, suggested_action), plus hurts, status, severity_rationale, source (scripted or agent_review), prioritized_actions[], proactive_recommendations[], already_in_place[], coverage, checks_skipped[], agent_review and limits[]. Findings are ordered by severity band, then priority; ids follow that order. summary.{critical,high,medium,low} count every finding and add up to total_findings; summary.active_by_severity counts active ones only.

References

• references/report-schema.md — the report, field by field
• references/severity-rationale.md — every base severity, and why
• references/jsonld-normalisation.md — the one parse every skill reads