audit-orchestrator

Files (1)

• SKILL.md

SKILL.md (first lines)

# Audit Orchestrator

## Purpose

This skill is the control plane for the audit. It coordinates a tight, explicit skill stack and combines the outputs into one final evidence-based report.

## Tight skill stack

Use only these core skills in order:

1. site-page-classifier
2. crawl-access-audit
3. semantic-entity-audit
4. robots-scope-analyzer
5. evidence-grounded-scorer
6. recommendation-engine

Do not broaden into exploratory or generic agents. The goal is not more skills; it is tighter, more specific reasoning.

## Responsibilities

- accept a target URL and site context
- classify likely site type and page patterns
- run the required checks in the right order
- merge outputs from crawl, render, semantic, robots, scoring, and recommendation checks
- score the severity of each finding
- produce a clean report with evidence, impact, and required fixes

## Inputs

- target URL
- optional site classification hints
- optional allowlist/denylist based on the sandbox rules

## Outputs

- site summary
- findings list
- overall AI readiness score
- engagement score
- severity counts
- prioritized recommended actions

## Core workflow

1. Fetch site metadata and initial page snapshot
2. Run crawl/access audit
3. Validate robots.txt scope before treating any block as a failure
4. Run render/content audit
5. Run semantic/entity audit
6. Run evidence-grounded scoring
7. Map findings to recommendations
8. Aggregate all findings
9. Run severity engine
10. Produce final findings report

## Behavior constraints

- do not perform authenticated actions
- do not violate robots.txt
- keep runtime under the contest limits
- prefer deterministic checks before LLM interpretation
- make evidence explicit in every finding
- keep the agent trace to a strict iteration/action/observation format
- cap reasoning at 20 iterations and show the iteration number for each step
- avoid broad exploratory skills that reduce precision or create noisy outputs

## Final implementation status

The current implementation is the benchmark-validated version of this skill package. It includes:

- explicit evidence gating for insufficient-data cases
- dedicated PASS / FAIL / N/A / UNKNOWN rule-state semantics
- canonical robots metadata for relevant crawler restrictions
- audit-safe classification logic for non-commercial and documentation-heavy pages
- final benchmark outputs saved in the repository under the bench directory

This package is ready for marketplace handoff and final submission packaging.

From Haloninja2605/Adobe_Hackathon, MIT.

More from Haloninja2605/Adobe_Hackathon