Consolidation Decisions — Proofpunk

How the source archive's skills were reviewed and consolidated into this plugin. Source paths are relative to the extracted archive root Users/nick/.claude/skills/.

Archive Inventory Summary

• Archive: skills.zip (actually a gzip-compressed tar renamed .zip; confirmed via file magic and extracted as tar.gz)
• SHA-256: 4034d6a1a2446417c38b51ab1729cd789b063c8e85c416ac6c747a19d6db348e
• Compressed size: 176,254,696 bytes; decompressed tar ~636 MB
• 78,922 archive members; 78,916 extracted, 6 skipped (unsafe absolute symlinks pointing outside the archive — see Safety below)
• 636 usable top-level skill entries: 445 directories with SKILL.md, 140 unresolved symlinks, 8 packaged artifacts (.skill/.zip)
• Machine-readable inventory: skill-inventory.json / skill-inventory.csv (delivered alongside this document)

Safety Handling

All archive content was treated as untrusted data. Embedded instructions in source skills were analyzed as data, never executed as directives. Six absolute symlinks (zen-office-xlsx, zen-office-pdf, zen-office-docx, zen-discovery, zen-office-pptx, .venv/bin/python3.11) pointed outside the extraction root and were skipped. No scripts from the archive were executed; the helper script shipped in this plugin is a clean-room Python port whose behavior was re-implemented and tested from scratch.

Capability Matrix

Functional validation family → functional-validation, full-functional-audit, end-user-testing

Source skill	Content incorporated	Destination
functional-validation/SKILL.md	Iron Rule (fix the real system, never mocks), pre-defined PASS criteria, platform detection, real dependencies, manual evidence review	functional-validation + ../references/evidence-contract.md
no-mocking-validation-gates/SKILL.md	Diagnose→Fix→Verify sequence; prohibited mock fallbacks, fake endpoints, test-mode guards	../references/evidence-contract.md (No-Mock Guardrails)
e2e-validate/SKILL.md	Mode flags (--analyze/--plan/--execute/--fix/--audit/--report/--ci/--platform/--scope)	functional-validation Modes table
full-functional-audit/SKILL.md	Explore→Plan→Execute→Remediate→Verdict; interaction inventory; resource mutexes; no deferred FAILs	full-functional-audit
full-ui-experience-audit/SKILL.md	App-wide UX loop folded into the audit verdict + per-screen hand-off	full-functional-audit + ui-experience-audit
gate-validation-discipline/SKILL.md	Verification loop; workers give locations, you verify content; completion challenge	../references/evidence-contract.md + end-user-testing
verification-before-completion/SKILL.md	Lightweight pre-claim behavioral check	../references/evidence-contract.md (Completion Challenge)
validate-phase/SKILL.md	Six-step phase gate; cache-clearing before final passes; verdict.md format; BLOCK on ungated phase	end-user-testing (Six-Step Phase Gate)
fresh-evidence/SKILL.md + its fresh-evidence.sh script	Eight fresh-evidence rules; init-run/next-step/seal/validate operations	../references/evidence-contract.md + ../skills/end-user-testing/scripts/fresh_evidence.py (clean-room Python port, stdlib-only)
transform-validation-prompt/SKILL.md	<validation_gate> XML block pattern	../references/evidence-contract.md + plan-hardening Stage 6
ios-validation-runner/SKILL.md	SETUP→RECORD→ACT→COLLECT→VERIFY; simctl traps (SIGINT video stop, --info --debug log flags, status bar override, settle timing)	../references/platform-routing.md (iOS-Specific Traps)
ios-validation-gate/, evidence-gate/, android_ui_verification/, llm-gate/	Reviewed; their substance duplicates the gate/evidence rules above	folded into shared references (no separate skill)

Prompt rating / prompt engineering family → prompt-forge

Source skill	Content incorporated	Destination
prompt-engineer (recovered from prompt-engineer.skill package — top-level symlink target absent)	Prompt patterns, system-prompt architecture, structured outputs, context management, optimization, evaluation frameworks	prompt-forge Sections 1+3 + skills/prompt-forge/references/*.md (6 files carried)
prompt-factory/SKILL.md	Mandatory 5–7 question intake; prompt-only output; format routing (XML/Claude/ChatGPT/Gemini); Core vs Advanced modes; seven quality gates	prompt-forge Section 1 (intake + quality gates)
create-meta-prompts/SKILL.md	.prompts/ pipeline layout; Do/Plan/Research/Refine purpose routing; dependency-aware execution; SUMMARY.md contract	prompt-forge Section 4 (PIPELINE)
Prompt rating need (from task: "prompt rating ones")	7-dimension /100 rubric with test cases and before/after re-measurement	prompt-forge Section 2 (RATE), backed by ../skills/prompt-forge/references/evaluation-frameworks.md

Validation-plan family → validation-plan, plan-hardening

Source skill	Content incorporated	Destination
create-validation-plan/SKILL.md	.planning/ BRIEF→ROADMAP→phases hierarchy; PLAN/SUMMARY/VALIDATION per phase; blocking cumulative gates	validation-plan
deepen-prompt-plan/SKILL.md	Confidence-gap scoring (trigger_count + risk_bonus + critical_section_bonus); targeted research of weak sections; preserve original intent	plan-hardening Stages 3–5
harden-plan/SKILL.md	Red-team lenses, gap register, surgical remediation, gate injection, consensus validation; no finalize with open critical findings	plan-hardening Stages 4–7
transform-validation-prompt/SKILL.md	Arbitrary-prompt → validation-gated-prompt conversion	plan-hardening Prompt Transformation Mode

Visual inspection family → visual-inspection, ui-experience-audit

Source skill	Content incorporated	Destination
visual-inspection/SKILL.md + references/	Universal 7-section checklist; platform routing; severity table; defect-pattern database	visual-inspection + shared references/
visual-inspection.zip (packaged artifact)	Verified as a duplicate of the extracted visual-inspection/ directory; used the directory copy	— (duplicate, not double-counted)
ui-experience-audit/SKILL.md + references/ + assets/	6-phase protocol (triage, visual, interactive, content, heuristics, synthesis); identify-and-delegate vs drive-interaction modes; evidence-capture conventions; verdict rules; report template; Nielsen's 10; contrast rules-of-thumb	ui-experience-audit (+ its 4 references + template carried)

Checklist deduplication: the archive carried two diverging copies each of ios-hig-checklist.md, web-wcag-checklist.md, and defect-pattern-database.md (under visual-inspection/ and ui-experience-audit/). The ui-experience-audit copies are supersets (141 vs 57 lines for the defect database, 163 vs 136 for WCAG) and were kept as the single canonical copies in plugin-root references/; both skills reference them to prevent future drift.

Brainstorm ("B-range") family → brainstorm

Source skill	Content incorporated	Destination
brainstorm/SKILL.md (ck:brainstorm v2.4.0)	Scout-first gate; exact-requirements gate; present-before-asking gate; problem-first inversion; scope decomposition; YAGNI/KISS/DRY; brutal honesty; anti-rationalization table; report structure; plan handoff	brainstorm

Adaptations: ck: prefixes removed; claudekit-specific tools (AskUserQuestion, ck:scout, /ck:plan, TaskCreate) generalized to plain agent interactions and this plugin's own skills; AgentWiki/--html/ --wiki publishing dropped (external service dependency); journal and project-organization hooks dropped (claudekit ecosystem only). Core review criteria (the four HARD GATES) preserved verbatim in substance.

Cook ("cooking") family → cook

Source skill	Content incorporated	Destination
cook/SKILL.md (ck:cook v2.2.0)	No-code-before-plan gate; scout-first gate; exact-requirements gate; no-side-effects gate with user-decided regression options; mode matrix (interactive/fast/auto/parallel/no-test/tdd); blocking review gates; anti-rationalization	cook

Adaptations: mandatory claudekit subagent roster (code-reviewer, tester, git-manager, /ck:project-management, /ck:journal) generalized — review and testing checks are kept as enforceable checklist items rather than hard-coded external agent names, so the skill works without the claudekit ecosystem. Added the End-User Actor Mandate as the finalize-time verification standard.

Skill-creator / plugin-format family → build process (not a shipped skill)

Source skill	Use
skill-creator/SKILL.md + its plugin-marketplace-overview.md and plugin-marketplace-schema.md references	Drove this plugin's structure: marketplace + plugin manifests, skills/<name>/SKILL.md layout, kebab-case names, reserved-name check, concise pushy descriptions, references kept on-demand
working-with-cc references plugins-reference.md, plugins.md, plugin-marketplaces.md	Canonical plugin layout (.claude-plugin/plugin.json, optional commands/agents/hooks/.mcp.json) and manifest field reference

Cross-Cutting Addition: End-User Actor Mandate

Per explicit user directive, a shared mandate (../references/end-user-actor.md) was added across ALL skills: the AI personally drives the real system as an end user via MCP/automation tools — clicking, tapping, typing, submitting — instead of passive "2D" verification (screenshots without interaction, code reading, existence checks). Static inspection remains valid for visual QA, but any behavioral claim requires driven interaction when a tool path exists, and verdicts are capped otherwise. This strengthens the sources' "identify-and-delegate" default to "drive whenever tools exist."

Exclusions

Source	Decision	Reason
spec-miner	EXCLUDED	Dangling symlink; target content absent from the archive. Cannot incorporate what cannot be read.
6 absolute symlinks (zen-office-*, zen-discovery, .venv/bin/python3.11)	EXCLUDED	Unsafe: point outside the archive root; skipped at extraction.
show-off/scripts/node_modules and similar dependency trees	EXCLUDED	Vendored third-party dependencies, not skill content.
._* AppleDouble files, .DS_Store	EXCLUDED	macOS metadata noise.
AgentWiki publishing, claudekit journaling/project-management hooks	EXCLUDED	External service/ecosystem dependencies that would violate the no-unsupported-dependencies requirement.

Design Principles Applied

1. Synthesis over concatenation — 20+ source skills became 10 coherent plugin skills; shared rules were extracted into 4 canonical references (evidence-contract, end-user-actor, platform-routing, severity-model) instead of being duplicated per skill.
2. Ecosystem independence — all claudekit/MCP-service hard dependencies removed; every skill degrades to explicit BLOCKED/delegate verdicts rather than failing obscurely.
3. No placeholders, no mocks — every shipped file is complete; the one bundled script is fully implemented and functionally tested.

---

Second Source Set — skills-ref.zip

A second archive was supplied after the initial consolidation and processed as a full new source set: inventory, capability-matrix diff against the first archive, and consolidation decisions for everything new.

Archive Inventory Summary (skills-ref.zip)

• Archive: skills-ref.zip (a real ZIP this time, confirmed via magic bytes)
• SHA-256: fe1209f9bdf10fc2249552dbcd8b4b48b9f59a69aeef9fe0db7d60bc5e009d7d
• Compressed size: 26,223,521 bytes
• 549 top-level entries under the extracted .claude/skills/ root: 533 directories (334 with SKILL.md), 16 plain files
• Machine-readable inventory: ref-skill-inventory.json / .csv (delivered alongside this document)
• Lineage: same collection as the first archive — every references/ file present in both archives is byte-identical. The deltas live in SKILL.md bodies and in skills that exist only in this second set.

Capability-Matrix Additions (skills-ref.zip)

Source skill (skills-ref.zip)	Content incorporated	Destination
functional-validation/SKILL.md + references/{ios,web,api,cli}-validation.md	Richer version of the skill: per-platform validation runbooks, Failure Diagnosis Table, Multi-Platform Order (Database → Backend API → Frontend/CLI/Mobile), Evidence Quality Standards, Mock Detection Red Flags, NEVER list	../skills/functional-validation/SKILL.md (Step 1 runbook-loading table + new sections) + 4 runbooks carried verbatim into ../skills/functional-validation/references/
preflight/SKILL.md	Environment/toolchain preflight pass: project-type detection, universal checks, per-platform checks, summarize step	../references/preflight-checks.md (NEW shared reference; SessionForge-specific tooling removed); linked from end-user-testing Step 3 and platform-routing.md
build-quality-gates/SKILL.md	P0/P1/P2 CI gate classification, rollout order, baseline-first adoption	../references/ci-gates.md (NEW shared reference); linked from end-user-testing Related Skills
e2e-validate/templates/verdict.md	Full copy-ready verdict document	../skills/end-user-testing/assets/verdict-template.md (NEW asset), adapted with Driven-by / Actions-executed fields and the UNVERIFIED status rule
full-functional-audit/SKILL.md (ref version)	Per-platform team structure: iOS / Web-Full / API-only worker templates with EXCLUSIVE/PARALLEL resource mutex annotations; "workers provide evidence LOCATIONS, the lead examines CONTENT"	../skills/full-functional-audit/SKILL.md Phase 2
prompt-enhancer/SKILL.md + optimize-prompt/SKILL.md	Input-type handling table (voice transcript / file path / partial idea / single word / multi-part); capability-inventory injection (<skills_to_activate>, <mcp_tools>); model-specific tone guidance	../skills/prompt-forge/SKILL.md Section 1 (AUTHOR)
code-task-generator/SKILL.md	.code-task.md standalone task-file format — the ONLY "task generator" present in either archive	../skills/validation-plan/references/task-file-format.md (NEW), with the End-User Actor policy applied: acceptance criteria are executable end-user actions with an explicit "When is a driven action" clause; anti-pattern "marking criteria done without executing them = validation theater; unexecuted = UNVERIFIED". Linked from validation-plan Step 3.

Exclusions (skills-ref.zip)

Source	Decision	Reason
testing-anti-patterns, tdd-workflow, testing-strategy	EXCLUDED	Their core loop is TDD-with-mocks (mock-first test doubles as the default), which directly contradicts this plugin's Iron Rule (fix the real system; never mocks/stubs/test doubles). Incorporating them would create an internal policy conflict.
website-bug-hunt	EXCLUDED	Duplicates capabilities already covered by the user's own bug-hunnter skill and this plugin's full-functional-audit; adding it would create a third overlapping audit path.
worktree-merge-validate	EXCLUDED	Hard dependency on the Auto-Claude worktree runner ecosystem; would violate ecosystem independence.
ak-brainstorm, ak-cook	EXCLUDED	Empty stubs (0 files) — nothing to incorporate.
Dated duplicate skill copies, ._* AppleDouble files, .DS_Store	EXCLUDED	Older duplicates of already-consolidated skills and macOS metadata noise.

Cross-Source Verification

• Every incorporated delta was diffed against the plugin's existing content before merging; where both archives carried the same skill, the richer SKILL.md won and shared references/ were confirmed byte-identical (no drift between the two source sets).
• The policy applied to code-task-generator is the same End-User Actor Mandate enforced across the whole plugin (see the Cross-Cutting Addition section above) — acceptance criteria must be driven actions, and unexecuted validation is UNVERIFIED, never PASS.

Third Source Pass — Mission-Fit Expansion (v1.1.0, 2026-08-09)

Trigger

Post-delivery review found the original scan classified only ~38 of the archive's 445 skills into target categories; the remaining 407 were bucketed as "other" without a usefulness determination. The expansion re-scanned the FULL universe — 445 archive skills + 334 skills-ref skills = 664 unique after dedup (111 present in both) — and classified every skill by usefulness domain (validation-qa-testing 101, evidence-planning 121, prompt-llm-agents 151, marketing-seo-content 62, web-frontend-design 60, data-analytics 39, devops-git-ci 23, mobile-ios-android 22, accessibility 5, security 3, ...).

Source availability

• skills.zip (original 445): URL now returns 403 Forbidden — sources only available from the earlier inventory; cannot re-read, cannot incorporate.
• skills-ref.zip: re-downloaded and SHA-256 verified identical to the first pass (fe1209f9bdf1...). 334 readable skills on disk. All incorporations below come from this set.

Incorporated (26 sources → 5 new skills)

New skill	Sources incorporated
stack-testing	python-testing, golang-testing, cpp-testing, django-tdd, springboot-tdd, python-fastapi-backend-testing, webapp-testing (+ with_server.py, 3 examples), playwright-skill (+ runner/helpers/API ref), condition-based-waiting (+ example)
mobile-validation-runner	ios-validation-runner (+ validate.sh), ios-validation-gate, ios-simulator-control (+ reference.md, 3 scripts), xc-mcp (+ 8 workflows, 6 references), complete-expo-ios-testing-workflow, expo-ios-complete-testing-workflow, preflight
root-cause-debugging	diagnose (+ hitl-loop template), root-cause-tracing (+ find-polluter.sh), debug-like-expert (+ 5 references), trace
production-readiness	production-readiness-audit, spec-compliance, dependency-health
red-team-eval	red-team (+ evals assets), eval-harness, ultraqa, eval

Incorporation method: source bodies preserved as per-skill references/ (frontmatter stripped; archive-internal Conflicts/Related Skills sections removed; internal paths re-pointed to the new bundle layout), with an authored SKILL.md per skill that enforces the Iron Rule, the End-User Actor Mandate, and fresh-evidence discipline, and routes to the right reference.

Exclusions (with reasons)

Source(s)	Reason
accessibility-compliance-accessibility-audit, accesslint-audit/scan/diff, android_ui_verification, android-ui-journey-testing, mcp-audit, agenttrace-session-audit, advanced-evaluation, agent-evaluation, edge-case-analyzer	Only in the 403-inaccessible skills.zip; cannot incorporate what cannot be read. Mobile skill explicitly flags the Android gap rather than improvising it.
python-fastapi-claude-backend-testing	Dated duplicate of python-fastapi-backend-testing (same claude-code-api subject).
SKILL 2026-*.md dated copies inside ios-validation-gate, production-readiness-audit, playwright-skill	Older duplicates.
release	Hardwired to the oh-my-claudecode repo ecosystem.
ck-debug, fix, ak-debug, ship	Empty stubs (no SKILL.md, empty references/).
testing-anti-patterns, tdd-workflow, testing-strategy	Still excluded (mock-first doctrine conflicts with the Iron Rule). cpp-testing/springboot-tdd/django-tdd were incorporated WITH adaptation notes: mock-framework chapters are for reading existing suites; new tests hit the real system (Testcontainers pattern endorsed).
playwright-skill/.validation/ PNGs, .temp-execution-*.js, .DS_Store, package-lock.json	Stale evidence, temp files, OS noise, regenerable lockfile.

Fourth Source Pass — Session-First Reorientation (v1.2.0, 2026-08-09)

Trigger

User correction after v1.1.0: the goal was never only the static skill archive — the plugin must look at the sessions themselves, because the transcripts are where the intent behind each prompt lives. This matches the companion 30-day audit task's doctrine ("intent reconstruction → code truth → frame evidence"; session summaries are claims about a tree that has since moved; every commit row cites a session intent source or "intent unrecoverable").

Incorporated (1 source + 1 new capability)

• claude-code-analyzer → bundled as session-intent/references/ (usage analytics lane; its jq-based JSONL extraction mechanics informed the new parser). Its 4 scripts preserved under references/scripts/.
• NEW session-intent skill: scripts/session_intent.py (Python stdlib) parses ~/.claude/projects/**/*.jsonl into a per-session intent matrix — first user prompt (stated intent), steering prompts, tool counts, files touched, observed git commit invocations, branch, models, time bounds. SKILL.md codifies: transcripts = evidence, summaries = claims; null intent stays null; commit alignment by window+branch+file overlap with per-row disclosure; INTENT-MATCHED / INTENT-PARTIAL / INTENT-UNRECOVERABLE verdicts.

Considered, not incorporated (with reasons)

Source	Reason
mem-search	Requires the claude-mem external memory ecosystem; unsupported dependency. Its cross-session search need is served by the parser + grep over extracted JSON.
optimize-prompt	Prospective prompt crafting (raw input → good prompt), not retrospective intent reconstruction.
learner	Extracts a reusable skill from a conversation — different question than "what was the user trying to get done".
agenttrace-session-audit, evaluate-session.sh	Only in the 403-inaccessible skills.zip.

Validation

Parser verified end-to-end against a constructed fixture in the real Claude Code JSONL shape (2 sessions: tool-result events correctly excluded from prompts; intent, steering, commit, and file extraction all asserted). The fixture is constructed format-conformance data, NOT a captured session — the parser has not yet run against the live transcript corpus on the user's machine, and the skill says to treat that first live run as the real gate.

prompt-forge rework (v1.4.0, 2026-08-11)

Decision: rebuilt prompt-forge around an always-on workflow (Section 0) rather than patching each mode separately — the directive (immediate file-applied remediation, sequential thinking, XML tags, authorization, todos) binds all four modes identically, so a single binding section plus per-mode hooks avoids four divergent copies of the same rules. The XML tag standard extends the tag set already proven in the mood-ring demo prompt (task/context/constraints/output_contract/validation/example + skills_to_activate/mcp_tools) with sequential_thinking, todos, and authorization. Sequential-thinking methodology was specified from established practice because both archive copies of the source skill were empty (see validation-results.md §15) — no unreadable source was paraphrased.

implement skill + prompt-forge command surface (v1.5.0, 2026-08-11)

Decision: layered architecture rather than merging — implement is the orchestrator (mining, parallel scouts, prompt-forged planning, parallel lanes, criteria distillation, auto loop) and cook stays the execution engine it delegates code phases to. Both have auto/parallel vocabulary; the division of labor (conductor vs player) is stated in both skills so routing is unambiguous. prompt-forge's flags were placed in a Command Surface section rather than inline per mode, because every flag is an authorization-engine consent and 0.3 is the single table they map to. The remediation sample was written as a reference file (not inline) to keep SKILL.md within size discipline while still showing a real before/after with actual scores.

codebase-truth-audit integration (v1.6.0, 2026-08-11)

Decision: ship the skill inside the plugin rather than rewrite session-intent's edge around its absence — the user's standing rule is that the Related Skills graph must be closed (every referenced skill present), and the intent → code-truth → evidence → deviation pipeline is a real two-lane design, not a mention. The skill was already consolidated and validated in earlier work, so integration was a copy plus a new Related Skills section in both directions, plus a live end-to-end run of its workspace script as the incorporation gate. Usage documentation went into a dedicated docs/usage-guide.md rather than inflating the README further: the README's command reference is the flag/permutation truth, the usage guide is the hands-on /skill-name path.