audit-orchestrator

Files (4)

• SKILL.md
• composition.md · references/
• report-schema.md · references/
• orchestrator.py · scripts/

SKILL.md (first lines)

# audit-orchestrator

## Purpose
This is the **only entrypoint** in the marketplace. It receives an audit
request (a URL) and returns the final structured report. It never performs
analysis itself beyond aggregation — all domain reasoning lives in the four
specialist skills it composes.

## When to use
Invoke this skill whenever a user asks to audit a website's AI discoverability
and/or on-site engagement. Do not invoke the specialist skills directly for a
full audit; they are designed to be composed by this orchestrator, though each
is independently runnable for isolated testing (see `references/composition.md`).

## How it works
1. Crawl the target origin once, bounded by `--max-pages`/`--max-depth`,
   respecting `robots.txt` at every fetch.
2. Pass the shared crawl result to all four specialists in-process (no
   re-crawling).
3. Merge findings, deduplicating same-root-cause manifestations reported by
   multiple specialists (see `references/composition.md`).
4. Apply confidence scoring, severity, and priority rules.
5. Classify the site/page type conservatively to scope which checks apply.
6. Generate proactive recommendations separate from defects.
7. Validate the report against the required minimum schema before returning it.

## Usage
```bash
python scripts/orchestrator.py https://example.com --max-pages 15 --max-depth 2 --out report.json
```

## Output
A single JSON report: `site`, `audited_at`, `summary` (counts by severity),
`audit_metadata` (crawl stats, site-type inference, limitations),
`findings` (each with evidence, root_cause, mechanism, suggested_action), and
`proactive_recommendations`. See `references/report-schema.md`.

## Safety
Read-only. Never submits forms, authenticates, or modifies the target site.
Always honors `robots.txt`.

From Adamkoda2306/Brand-AI-Readiness-Audit, MIT.

More from Adamkoda2306/Brand-AI-Readiness-Audit