audit

name	audit
description	Evaluate the IP constellation — dispatches baseline checks, then 9 lenses in parallel (Millman, Bierut, Appleton, Shaw, Peers, Victore, The Walk, The Walls). Coordinator skill.
user_invocable	true

Constellation Audit — Coordinator Skill

Purpose

Evaluate Peter's IP constellation (hub site + spoke repos) in layers: visual inspection, mechanical baseline checks, then nine lenses dispatched in parallel. Reports only — never auto-fixes.

Quick Reference

/audit run                    # Full audit: visual + baseline + all 9 lenses
/audit run baseline           # Baseline checks only (dispatches /baseline)
/audit run [lens]             # Single lens (millman, bierut, appleton, shaw, peers, victore, the-walk, the-walls)
/audit check [page-or-repo]   # All checks scoped to one target
/audit rubric                 # Show/create rubric file
/audit report                 # Show last saved report

Discovering the Constellation

• Hub: The current project (petersalvato.com)
• Published pages: All files in _governance/, _infrastructure/, _output/ WITHOUT published: false
• Blog posts: All files in _blog/ WITHOUT published: false
• Standalone pages: bio/index.md, contact.md, colophon.md, thinking.md, vocabulary.md
• Homepage: _data/index.json + _layouts/systemworks.html
• Spokes: Scan ~/homelab/projects/active/ for repos containing a CNAME file
• Navigation: _data/navigation.json

Context

Read

• NORTHSTAR.md — master vision ("the site is a world, not a résumé"). The audit evaluates against this intent.
• docs/ARCHITECTURE.md — intended IA structure (three tiers, user journeys). Baseline checks mechanical health against this.
• docs/aetherwright-identity-decisions.md — persona layer (who Peter is)
• All published pages (frontmatter + body)
• Screenshots from .audit/screenshots/[today's date]/ — if full-pass passed a SCREENSHOT_DIR, use that path
• docs/visual-reference-index.md → active reference set
• .audit/rubric.md if it exists

Invoke

• /baseline → mechanical health → pass/fail per check
• /millman → personhood → STRONG/HOLDS/WEAK/BROKEN + direction
• /bierut → design thinking → STRONG/HOLDS/WEAK/BROKEN + direction
• /appleton → living knowledge → STRONG/HOLDS/WEAK/BROKEN + direction
• /shaw → world feel → STRONG/HOLDS/WEAK/BROKEN + direction
• /peers → peer benchmark → STRONG/HOLDS/WEAK/BROKEN + direction
• /victore → bravery → Brave/Safe/Performed/Sanitized + direction
• /the-walk → example positioning → ONE WORLD/HOLDS/SPLIT/WRONG ROOM + direction
• /the-walls → ambient frame → COHERENT/HOLDS/SPLIT/NOISE + direction
• /copy-verify → voice spot-check → pass/fail per item
• /knowledge → ideation grounding → source-attributed search results

Lens Criteria (summary)

M1-M5: Real person, vulnerability, voice, life-arc, courage B1-B5: Form/content, problem-solving, meaning, content-driven, intellectual A1-A5: Connections, growth, epistemic disclosure, contextual, evolving S1-S5: World immersion, total environment, stranger orientation PB1-PB5: Structure, visual evidence, public output, social proof, distribution Victore: Fierce self vs performing W1-W5: Room per page, sidebar entry, secondary identity leading, governance examples, one world F1-F5: Persistent frame identity, marginalia signal, frame/content agreement, stranger/initiate gradient, ambient density distribution

Dispatch: /audit run

Layer 0: Visual Evaluation (MANDATORY)

Visual inspection is required. The audit cannot evaluate design lenses by reading markdown source alone.

0.0 Start Dev Server

1. Check if localhost:4000 is reachable
2. If not: bundle exec jekyll serve --host 0.0.0.0 &
3. Poll until 200 response
4. If server can't start, note the error and proceed with source-only evaluation — flag as limitation

0.1 Screenshots Take screenshots of every published page at 3 viewports (375px, 768px, 1440px) using Playwright or equivalent. Save to .audit/screenshots/YYYY-MM-DD/. Capture all pages, not a sample.

0.2 Visual Assessment Read every screenshot as an image. Evaluate: typography hierarchy, whitespace, visual weight, layout coherence across viewports, visual-content alignment. Note ALL user-facing elements (images, codex strings, glyphs, altitude markers, faculty dots). Compare against previous screenshots if available.

0.3 First Visitor Walkthrough

1. Read homepage manifesto first 3 paragraphs
2. Report: what does a first-time visitor understand in 30 seconds? What confuses them?
3. Navigate 2 more pages as a new visitor would
4. Report the emotional arc: trust, curiosity, or confusion?

Layer 1: Baseline (sequential — lenses depend on this)

Dispatch /baseline atomic skill. This runs all mechanical checks:

• 1.1 Links resolve
• 1.2 Images exist
• 1.3 Cross-references hold
• 1.4 Attribution chain
• 1.5 Voice violations
• 1.6 SEO metadata
• 1.7 Vocabulary tooltips functional

Baseline results inform lens evaluations (e.g., Appleton A5 needs tooltip results, Bierut B4 needs vocabulary visibility).

Layer 1.5: Ideation History Verification

Use /knowledge skill to verify copy grounding:

• Search each project by name against ideation history
• Flag projects with no ideation history as potentially ungrounded
• Compare Peter's conversational voice to published voice (for voice source check)

Layer 2: Lenses (all 9 in parallel)

After baseline completes, dispatch all 9 lenses simultaneously using the Agent tool:

1. /millman — Is this person real? (M1-M5)
2. /bierut — Does the design think? (B1-B5)
3. /appleton — Is the knowledge alive? (A1-A5)
4. /shaw — Does this feel like a world? (S1-S5)
5. /peers — Peer benchmark (PB1-PB5)
6. /victore — Is Peter being fiercely himself?
7. /the-walk — Front door to tattoo chair (W1-W5)
8. /the-walls — What room does the frame build? (F1-F5)
9. /copy-verify — Spot-check voice on a sample of published pages (optional)

Each lens reads the same published material and the rubric independently. Each returns its own verdict: STRONG / HOLDS / WEAK / BROKEN (except Victore which returns Brave / Safe / Performed / Sanitized).

Screenshot forwarding: If a SCREENSHOT_DIR was provided (from full-pass), include screenshot paths in the /shaw and /the-walls agent prompts. Shaw evaluates content and visual world-feel. The Walls evaluates whether ambient frame elements (sidebar, glyphs, codex, typography) build the right room — screenshots are essential for this. Example paths to include: SCREENSHOT_DIR/laptop/index.png, SCREENSHOT_DIR/laptop/encore.png, SCREENSHOT_DIR/mobile/index.png.

Layer 3: Synthesis

After all lenses return:

1. Collect verdicts and key findings
2. Cross-reference: do lenses agree or contradict?
3. Generate the report with all verdicts and evidence
4. Save to .audit/reports/YYYY-MM-DD.md
5. Print full report in conversation

Report Format

# Constellation Audit — [DATE]

## Baseline
[NUMBER] checks passed. [NUMBER] failed.
[List each failure with file path and specific issue]

## Millman — [STRONG/HOLDS/WEAK/BROKEN]
[Key findings from /millman — M1-M5 verdicts with evidence]

## Bierut — [STRONG/HOLDS/WEAK/BROKEN]
[Key findings from /bierut — B1-B5]

## Appleton — [STRONG/HOLDS/WEAK/BROKEN]
[Key findings from /appleton — A1-A5]

## Shaw — [STRONG/HOLDS/WEAK/BROKEN]
[Key findings from /shaw — S1-S5]

## Peer Benchmark — [STRONG/HOLDS/WEAK/BROKEN]
[Key findings from /peers — PB1-PB5]

## Victore — [Brave/Safe/Performed/Sanitized]
[Findings from /victore]

## The Walk — [ONE WORLD/HOLDS/SPLIT/WRONG ROOM]
[Key findings from /the-walk — W1-W5: room per page, sidebar entry, secondary identities, governance examples, one world]

## The Walls — [COHERENT/HOLDS/SPLIT/NOISE]
[Key findings from /the-walls — F1-F5: persistent frame, marginalia signal, frame/content agreement, stranger/initiate gradient, ambient density]

## Limitations
- Visual evaluation is Claude's reading as a multimodal viewer, not expert design critique
- First-visitor walkthrough is a proxy, not user testing
- Voice checking catches pattern-matchable violations only
- These are Claude's interpretations of the frameworks, not the original authors' evaluations
- Peer benchmark is based on publicly visible content, not private work or reputation

/audit run [lens]

Dispatches a single lens directly. No baseline, no synthesis — just the lens output.

/audit check [target]

Same checks, scoped to one page or spoke repo. Target can be:

• A file path: _output/aiden-jae.md
• A spoke name: Savepoint.Protocol
• A URL path: /infrastructure/encore

/audit rubric

Read and display .audit/rubric.md. If it doesn't exist, create it with starter content. Peter's rubric annotations override default evaluation criteria.

/audit report

Read the most recent file from .audit/reports/. If no reports exist, say so.

Notes

• Reports only. Never auto-fix.
• The rubric is Peter's. His annotations override defaults.
• Verdicts, not scores. STRONG/HOLDS/WEAK/BROKEN forces judgment.
• Always state limitations.