Skill System Baseline

Parent Skill

algolia-search-audit is the full-pipeline orchestrator. It routes the audit through waves, spawns isolated worker agents, verifies file outputs, and blocks progression when gates fail.

Important properties:

• Orchestrator does no data collection or report writing.
• Worker skills run in isolated contexts.
• Communication is via files.
• Gates check file existence, byte size, screenshot count, JSON validity, and factcheck verdict.
• The skill system uses model tiers by task type: Haiku for programmatic collection, Sonnet for structured extraction/light synthesis, Opus for deep reading and creative generation.

Canonical Output Contract

AGENT-CONTEXT.md defines canonical JSON keys, CSS classes, typography tokens, renderer functions, path conventions, and validation rules.

The most important JSON top-level fields include:

meta, cover, score, company_snapshot, executives, intelligence_signals,
competitors, findings, gap_pairs, financials, traffic, tech_stack,
ae_fields, next_steps, methodology, bibliography, competitive_synthesis,
golden_angle, strategic_angles, hiring, icp_mapping, abx_sequence,
case_studies, demos, partner_intel, tab_subtitles, recommended_first_play,
industry_context

Finding objects have exact required keys:

id, title, severity, category, tested_query, expected_behavior,
actual_behavior, impact_stat, impact_stat_source, screenshot_file,
prospect_description, algolia_solution, algolia_case_study_url,
algolia_case_study_company, algolia_case_study_result

This is not just presentation schema. It is a product contract PRISM must preserve or consciously migrate.

Workflow

Wave 1 — Intelligence Collection

Runs independent modules in parallel:

1. algolia-intel-company
2. algolia-intel-techstack
3. algolia-intel-traffic
4. algolia-intel-competitors
5. algolia-intel-financial-public or algolia-intel-financial-private
6. algolia-intel-investor
7. algolia-intel-hiring
8. algolia-intel-social
9. algolia-intel-news
10. algolia-intel-partner
11. algolia-intel-industry

Wave 2 — Query Generation

algolia-intel-queries reads company context and traffic keywords to produce 05-test-queries.md.

Layer 2 — Browser Audit

algolia-audit-browser executes 20 browser tests and captures screenshots.

Layer 3 — Synthesis

Business case, sales plays, report rendering, JSON generation, and deterministic patching happen here.

Layer 3D — ABX

algolia-campaign-abx creates 10 campaign files.

Layer 4 — Factcheck

algolia-audit-factcheck verifies claims and writes the gate verdict.

Deterministic Scripts

The parent skill relies on scripts for repeatable collection and patching:

Script	Role
collect-company.py	Company context, source website metadata, Scout fill-null enrichment
collect-techstack.py	BuiltWith + SimilarWeb technology collection
collect-traffic.py	SimilarWeb traffic endpoints
collect-competitors.py	SimilarWeb competitor discovery
collect-financials.py	Yahoo Finance/yfinance financial profile
collect-hiring.py	ICP classification reference after LinkedIn scraping was removed
collect-social.py	Apify social posts with fallback rules
collect-news.py	Google News RSS and newsroom feeds
collect-investor.py	SEC/earnings-call-oriented investor intelligence
collect-industry.py	Tavily/WebSearch industry benchmarks and trends
audit-browser.js	Playwright stealth browser audit
calculate-score.py	Weighted 10-area scoring
calculate-roi.py	Revenue impact scenarios
generate-audit-data.py	Post-LLM deterministic JSON correction
validate-json-schema.py	Blocks render when data keys do not match renderer expectations
render-audit.ts	Generates SPA/report artifacts

Evidence Discipline

The skill system is evidence-first:

• No naked data points.
• Every stat needs source provenance.
• Factcheck tiers determine whether claims stay, warn, or drop.
• If a source cannot be cited at write time, the claim should not be written.