Site Overhaul

Site Overhaul — Skill Definition

Purpose

End-to-end audit + realignment of acridautomation.com. Catches drift, fixes it, scores it, and forces a mission-level decision when identity and offering contradict each other. This skill exists because the site drifts every time Acrid ships anything, and tidying by hand burns runway.

When to Run

• Operator says "the site is a mess" or "clean up the site"
• Drift-checker reports ≥5 mismatches in a single session
• After any structural change (new product, removed product, new page)
• At minimum: once every 14 days as part of the compounding system
• Autonomously, if Acrid notices numeric contradictions across ≥3 surfaces

Pre-Execution Checklist

1. Read this entire file
2. Read skills/site-overhaul/RUBRIC.md
3. Read skills/site-overhaul/LEARNINGS.md — apply every lesson
4. Read soul/SOUL.md + soul/IDENTITY.md — voice + mission canon
5. Read CLAUDE.md — current stated build state
6. Read site-config.json — current "source of truth" (may itself be drifted)
7. Read the three sub-agent definitions:

• .claude/agents/drift-checker.md
• .claude/agents/site-syncer.md
• .claude/agents/content-auditor.md

Input

• Scope (optional): If operator passes nothing, run full overhaul. $ARGUMENTS is available but v1 ignores it — full runs only.
• Prior audit folders: memory/site-audit-*/ — checked for mission-decision inheritance.

Output Directory

memory/site-audit-YYYY-MM-DD/
  00-mission-review.md       # A/B/C decision + rationale (FORCED)
  01-drift-sync.md           # numbers reconciled, source of truth
  02-live-crawl.json         # URL map + per-page metadata
  03-seo-audit.md            # meta / OG / schema / canonical / H1 / alt
  04-voice-audit.md          # per-page brand-voice score (delegated)
  05-flow-audit.md           # conversion path walkthroughs
  06-visual-coherence.md     # header/footer/CSS/OG parity
  07-known-issues.md         # queue: what Acrid couldn't/wouldn't auto-fix
  08-decisions.md            # strategic calls + rationale
  09-changes-applied.md      # diff ledger of every file edited
  REPORT.md                  # master report for operator
  RUBRIC.md                  # scorecard (filled in)
  RAW/                       # raw HTML fetches, artifacts

Dated folder per run. No collisions. Natural history. Crash-recoverable (artifacts are append-only).

---

Phases

Six phases. One commit at the end (not per-phase). The repo stays dirty across phases — crash recovery comes from the audit folder, not from git. If a crash happens mid-run, next run starts fresh; nothing is lost except the incomplete run's artifacts.

Phase 0 — Bootstrap & Snapshot

Purpose: Scaffold the audit folder, lock the starting git SHA, create placeholder files. Actions:

1. mkdir -p memory/site-audit-$(date +%Y-%m-%d)/RAW
2. Write REPORT.md stub with header + git rev-parse HEAD captured
3. Touch all phase files empty

Delegate-or-Inline: Inline. Output: empty folder structure, stub REPORT. Commit: None. Scaffolding only.

---

Phase 1 — Mission & Identity Review (FORCED DECISION)

Purpose: Resolve the one thing drift cleanup can't fix — identity vs. offering contradiction.

Inputs: soul/SOUL.md, soul/IDENTITY.md, CLAUDE.md current build state, prior audit folders (inheritance check), current services page HTML.

Actions:

1. Inheritance check: Find newest memory/site-audit-*/00-mission-review.md within 30 days. If it exists AND SHA-256 of soul/IDENTITY.md is unchanged since that audit, inherit the prior decision verbatim and skip the forced call. Note the inheritance in 00-mission-review.md. Continue to Phase 2.
2. If no inheritance: Pick exactly one of:

• A — KEEP services, reframe as "Bridge Revenue." Explicit disclaimer block, moved off main nav, renamed copy.
• B — CUT services entirely. Archive the page, 301 redirect, remove from products.
• C — REFRAME services as "Done-With-You." Productize so Acrid does 80%, operator last-mile.

1. Writing "defer to operator" is banned and fails the rubric. Force the call.
2. Document the decision in 00-mission-review.md: choice, 3-sentence rationale, 3 specific changes it implies.

Delegate-or-Inline: Inline. This is a judgment call only Acrid can make. Output: 00-mission-review.md with the call and its implications. Commit: None (batched at end).

---

Phase 2 — Drift & Number Reconciliation

Purpose: Site-config.json becomes ground truth. Count from disk. Propagate.

Inputs: Disk (actual files), site-config.json, CLAUDE.md, site/index.html, site/products/index.html, site/how-it-works/index.html, site/llms.txt.

Actions:

1. Count from disk — this is truth:

• ls site/blog/ | wc -l → blogpostcount
• ls site/learn/ | wc -l → learnarticlecount (exclude _template.html + index.html)
• ls skills/ | wc -l → skill_count
• Manual count of products from site-config.json products.* arrays → product_count

1. Delegate drift-checker sub-agent — read-only audit of all surfaces. Capture its report into 01-drift-sync.md.
2. Update site-config.json first with disk-truth counts.
3. Delegate site-syncer sub-agent with instruction: "site-config.json is now canonical. Propagate counts/products/affiliates to every surface it touches."
4. Hand-patch any remaining surfaces site-syncer misses (CLAUDE.md product table, hardcoded counts in HTML like "40+ questions").

Delegate-or-Inline: Delegate drift-checker and site-syncer. Inline for the hand-patches and file-system counting. Output: 01-drift-sync.md with before/after table. Commit: None (batched).

---

Phase 3 — Live Crawl + SEO Audit

Purpose: Catch what's actually deployed (not just what's in git). Fix mechanical SEO defects.

Inputs: site-config.json pages[] array, site/sitemap.xml, live site at acridautomation.com.

Actions:

1. Seed URL list: union of site-config.json pages[], sitemap.xml entries, and filesystem-derived URLs (every site/*/index.html + /blog/* + /learn/*). Dedupe.
2. Fetch each page using WebFetch. (If Firecrawl MCP becomes available later, swap the fetch step — isolated by design.) Save raw markdown/text to RAW/.md. Record fetch status + final URL in 02-live-crawl.json.
3. Per-page SEO audit. For each page record: title length, meta description length, H1 count, canonical URL, og:title/description/image/url, twitter card, JSON-LD schema presence, alt attribute coverage (count of ` without alt`), Plausible script presence.
4. Auto-fix mechanical defects in the source HTML files (not live — local):

• Missing canonical → add
• Missing og:image → add using site default
• Missing twitter:card → add summary_large_image
• `` without alt → add alt derived from filename or surrounding H2
• Missing Plausible script → inject
• Stale year in footer copyright → replace
• Missing meta description → flag in known-issues (do NOT auto-write — this is judgment, not mechanical)

1. Blocklist (operation-scoped, NOT path-scoped):

• `` / meta / OG / canonical / alt / Plausible → FIXABLE on all pages, including blog + learn
• Body narrative / H1 / H2 content on existing blog + learn → READ-ONLY
• soul/*.md → read-only always
• memory/kaizen-log.md → append-only by /kaizen only
• .git/, node_modules/, *.bak → never touched

1. Queue strategic issues (thin content, duplicate H2s, weak keyword targeting) into 07-known-issues.md.

Delegate-or-Inline: Inline. Fetch + mechanical edits are scripted, not creative. Output: 02-live-crawl.json, 03-seo-audit.md (table: page × check × status × action). Commit: None (batched).

---

Phase 4 — Voice, Flow, Visual

Purpose: Score the non-mechanical stuff. Apply top-priority rewrites within a hard cap.

Inputs: crawled content from Phase 3, soul/IDENTITY.md voice canon.

Actions:

1. Delegate content-auditor with explicit scoring prompt: score every public page 1-5 on identity alignment, autonomy-claim consistency, tone (punchy/specific/no corporate filler), prohibited phrases. Write to 04-voice-audit.md. Pages ≤3 enter rewrite queue.
2. Flow walkthrough (inline). Walk three canonical paths on crawled HTML:

• Home → $17 Full Workspace Builder → Gumroad checkout
• Home → Learn article → in-article CTA → product
• Home → /roast → $99 GEO Audit upsell → intake

At every hop: does the CTA exist? does the link resolve? does the narrative match? Write to 05-flow-audit.md.

1. Visual coherence (inline). Diff ` / / nav / stylesheet sets across crawled pages. Verify favicon, verify og-image returns 200. Write to 06-visual-coherence.md`.
2. Apply strategic fixes. Cap: 10 edits per run. Priority: mission-decision changes (Phase 1) > flow fixes > visual parity > voice rewrites on score ≤2 pages. Everything beyond the cap → 07-known-issues.md.
3. Never rewrite site/about/, site/index.html hero, or any blog/learn body unless a specific voice score ≤2 defect is cited with page + line + reason in 04-voice-audit.md.
4. Log every edit in 09-changes-applied.md with before/after snippets.

Delegate-or-Inline: Delegate voice scoring to content-auditor. Flow + visual + edits inline. Output: 04-voice-audit.md, 05-flow-audit.md, 06-visual-coherence.md, 09-changes-applied.md (partial). Commit: None (batched).

---

Phase 5 — Rubric + Report + Session Close

Purpose: Score the run, write the operator-facing report, close state, commit, push.

Actions:

1. Fill RUBRIC.md — copy the template from skills/site-overhaul/RUBRIC.md, fill every gate + scored dimension. Rubric scoring is done by content-auditor (delegated) to avoid self-grading bias — send it the audit folder and the rubric template.
2. Write REPORT.md — 10-bullet "walking back into" summary + links to every phase file + rubric pass/fail + commit list.
3. Append to infrastructure/launch-cockpit.md — dated entry noting the audit ran, the mission decision, the count of fixes applied.
4. Append to skills/self-improvement/SITE-IMPROVEMENTS.md — any queued issues from 07-known-issues.md that are site-level.
5. Update skills/site-overhaul/LEARNINGS.md — dated entry (format below).
6. Session continuity:

• Append to memory/kaizen-log.md via /kaizen convention (append-only)
• Verify site-config.json reflects any new counts

1. Pre-commit gate:

• If any DITL post was touched, run ./scripts/validate-ditl.sh — refuse commit on fail, git restore , log to known-issues
• If any learn article was touched, run ./scripts/validate-learn.sh — same rule

1. Commit everything in one commit:

`` feat(site): overhaul YYYY-MM-DD — mission call [A/B/C], N fixes, N queued ``

1. git push once. Netlify deploys once. One green build.

Delegate-or-Inline: Delegate rubric scoring to content-auditor. Inline for everything else. Output: Full audit folder written, commit pushed, report ready for operator. Commit: Single final commit.

---

Hard Rules

1. Autonomous by default. No "ask operator." Document judgment in 08-decisions.md.
2. Capability truth. Don't claim fixes you didn't make. Don't invent metrics. If a check fails mid-run, log it and keep going.
3. One commit, one push. Crash recovery lives in the audit folder, not in git history. This keeps Netlify from deploy-storming and keeps git log clean.
4. Blocklist is operation-scoped: ``/meta/OG/canonical/alt = fixable anywhere. Body narrative on blog/learn = read-only. soul/ = read-only. memory/kaizen-log.md = append-only by /kaizen.
5. 10-edit cap on strategic changes per run. Everything else queues to 07-known-issues.md.
6. Mission decision inheritance: if prior 00-mission-review.md within 30 days + identity SHA unchanged, inherit.
7. Rubric scoring is delegated to content-auditor — the skill does not grade itself.
8. Never rewrite site/about/ or site/index.html hero unless a cited voice score ≤2 defect demands it.
9. Pre-commit validator gate. Touched DITL → validate-ditl. Touched learn → validate-learn. Fail → restore → queue.
10. If a strategic fix needs more than 10 edits, stop at 10 and queue the rest. Do not blow through the cap.

---

Failure Conditions

• Mission decision deferred to operator — automatic rubric failure
• Audit folder not created before Phase 1 starts
• Commit without running pre-commit validators on touched files
• site-config.json not updated to disk-truth counts
• Report does not include 10-bullet operator summary
• Rubric not filled in
• `` element edits that break HTML validity
• Any edit to soul/*.md
• More than one push per run

---

LEARNINGS.md Entry Format (post-run)

## YYYY-MM-DD — Site Overhaul Run

**Mission call:** [A/B/C — one sentence why]

**WHAT WORKED:**
-

**WHAT FELT WEAK:**
-

**ONE THING TO DO BETTER NEXT TIME:**
-

**Rubric score:** [overall pass/fail, 5 dimensions × 0-5]
**Fixes applied:** N mechanical, N strategic
**Queued:** N items to SITE-IMPROVEMENTS.md

---

Reused Infrastructure (do not duplicate)

| Asset | Path | Used In | |-------|------|---------| | drift-checker | .claude/agents/drift-checker.md | Phase 2 | | site-syncer | .claude/agents/site-syncer.md | Phase 2 | | content-auditor | .claude/agents/content-auditor.md | Phase 4, Phase 5 rubric | | validate-ditl.sh | scripts/validate-ditl.sh | Pre-commit gate | | validate-learn.sh | scripts/validate-learn.sh | Pre-commit gate | | site-config.json | repo root | Canonical source | | soul/IDENTITY.md | soul/ | Voice canon | | SITE-IMPROVEMENTS.md | skills/self-improvement/ | Queue merge target | | launch-cockpit.md | infrastructure/ | Operator dashboard |

---

Token Budget

Expected: 60-100k tokens per full run (includes 3 sub-agent delegations + live crawl + authoring report). Acceptable to run at full model. /fast not recommended — this skill's value is judgment, not speed.

---

The site drifts because Acrid ships. That's a good problem. This skill makes the cleanup cost near-zero.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: acrid-auto
• Source: acrid-auto/acrid-brain
• License: MIT

Install and usage instructions live in the source repository linked above.