Skip to content

feat: add /eval skill — AI output evaluator and grader#1

Open
HMAKT99 wants to merge 22 commits intomainfrom
arun/eval-skill
Open

feat: add /eval skill — AI output evaluator and grader#1
HMAKT99 wants to merge 22 commits intomainfrom
arun/eval-skill

Conversation

@HMAKT99
Copy link
Owner

@HMAKT99 HMAKT99 commented Mar 18, 2026

Your AI feature works on 5 examples. Does it work on 5,000?

You ship a prompt change. It looks great on the 3 examples you tested. In production, it hallucinates on edge cases, complies with injection attacks, and regresses on queries that used to work.

What /eval does

Discovers every prompt and LLM call in your codebase, generates comprehensive eval suites (happy path, edge cases, adversarial, safety, format), scores results with letter grades, and detects regressions when prompts change.

You:   /eval --audit

Claude: PROMPT QUALITY SCORECARD
        Prompt        Clarity  Safety  Specificity  Examples  Grade
        ai_chat       4/5      3/5     4/5          0/5       B
        summarize     5/5      4/5     5/5          3/5       A-

        ai_chat needs: few-shot examples + safety instruction

Only .tmpl committed — build generates the rest.

Test plan

  • .tmpl follows template pipeline
  • Registered in gen-skill-docs.ts, skill-check.ts, test files

garrytan and others added 22 commits March 16, 2026 10:59
@garrytan @claude
fix: dynamic base branch detection across all SKILL templates (v0.3.1…
1e06b6a 
…0) (garrytan#81)

* feat: add {{BASE_BRANCH_DETECT}} resolver to gen-skill-docs

DRY placeholder for dynamic base branch detection across PR-targeting
skills. Detects via gh pr view (existing PR base) → gh repo view
(repo default) → fallback to main.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: ship skill detects base branch instead of hardcoding main

Replaces ~14 hardcoded 'main' references with dynamic detection via
{{BASE_BRANCH_DETECT}}. Fixes stacked branches and Conductor workspaces
targeting non-main branches. Adds --base <base> to gh pr create.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: review, qa, plan-ceo-review detect base branch dynamically

Same pattern as ship: replaces hardcoded 'main' with {{BASE_BRANCH_DETECT}}.
Also cleans up qa bash-isms (REPORT_DIR variable, port chaining).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: retro detects default branch instead of hardcoding origin/main

Retro queries commit history (not PR targets), so uses simpler detection:
gh repo view defaultBranchRef. Replaces ~11 origin/main refs with
origin/<default>.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: add explicit cross-step references in gstack-upgrade template

Bash blocks are self-contained, but cross-block variable references
(INSTALL_DIR from Step 2) were implicit. Adds prose making them explicit.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs+test: SKILL authoring guidance + regression tests

Adds "Writing SKILL templates" section to CLAUDE.md explaining that
templates are prompts, not scripts. Adds validation test catching
hardcoded 'main' in git commands, and resolver content test.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: update ARCHITECTURE + CONTRIBUTING for new placeholders

Add {{BASE_BRANCH_DETECT}} to ARCHITECTURE.md placeholder list.
Cross-reference CLAUDE.md template authoring guidance from CONTRIBUTING.md.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* chore: bump version and changelog (v0.3.10)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: add missing blank line between resolver functions

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* test: add 3 E2E smoke tests for base branch detection

- /review: verifies Step 0 detection + git diff against detected base
- /ship: truncated dry-run (Steps 0-1 only, no push/PR), asserts no
  destructive actions
- /retro: verifies default branch detection for git log queries

Covers the {{BASE_BRANCH_DETECT}} resolver path (review), the ship
template's dual abort check, and retro's inline detection pattern.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* chore: bump version and changelog (v0.4.2)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@garrytan @claude
feat: await support in browse js/eval + contributor mode v2 (garrytan…
78e519e 
…#104)

* feat: support await in $B js and eval commands

Auto-wrap await expressions in async IIFE context so
$B js "await fetch(...)" works without SyntaxError.

- hasAwait() strips comments before detection
- js: expression wrapping (async()=>(expr))()
- eval: smart wrapping — single-line=expression, multi-line=block
- 6 new unit tests covering async, false-positive, and return semantics

* feat: redesign contributor mode — periodic reflection with 0-10 rating

Replace passive "report when things break" with active reflection:
- Rate gstack experience 0-10 at workflow step boundaries
- Historical calibration example (await bug) anchors the reporting bar
- "What would make this a 10" field focuses on actionable improvements
- Removed category lists in favor of judgment-based assessment

* test: add deterministic contributor mode preamble validation

40 new skill-validation tests (4 checks × 10 skills) verify:
- 0-10 rating scale present
- Calibration example present
- "What would make this a 10" field present
- Periodic reflection (not per-command)

Update existing E2E contributor eval for new report format.

* chore: bump version and changelog (v0.4.2)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: improve contributor mode + qa-quick E2E reliability

Contributor mode:
- Add "do not truncate" directive to template — agent was stopping
  after "My rating" without completing Steps/Raw output/What would
  make this a 10 sections
- Restore assertions for Steps to reproduce and Date footer

QA quick:
- Make test server URL prominent: top of prompt, explicit "already
  running" and "do NOT discover ports" instructions
- Bump session timeout 180s→240s and test timeout 240s→300s
- Set B= at top of prompt (was buried in prose)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: use flexible assertions for contributor mode E2E

Agent writes thorough reports with creative section names
("Repro Steps" vs "Steps to reproduce"). Match intent not formatting:
- /repro|steps to reproduce/ for reproduction steps
- /date.*2026/ for date footer presence

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* docs: add E2E eval failure blame protocol

"Not related to our changes" is an extraordinary claim that requires
extraordinary proof. When evals fail during /ship:

1. Run the same eval on main — prove it fails there too
2. If it passes on main, it IS your change — trace the blame
3. If you can't verify, say "unverified" not "pre-existing"

Added to CLAUDE.md and as a comment in skill-e2e.test.ts.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* docs: update CONTRIBUTING.md and BROWSER.md for v0.4.2

CONTRIBUTING.md: update contributor mode description — now describes
periodic 0-10 reflection loop instead of passive friction detection.

BROWSER.md: add js/eval async documentation — await expressions are
auto-wrapped in async context, single-line eval returns values directly.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: restore v0.4.2 changelog entries lost during cherry-pick conflict

The base branch detection entries from main were dropped when resolving
the CHANGELOG conflict — should have merged both sets, not replaced.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
@garrytan @claude
feat: always-on ELI16 + branch detection (v0.4.3) (garrytan#108)
276d0cc 
* feat: always-on ELI16 + branch detection in preamble

- Add _BRANCH detection to preamble bash block (git branch --show-current)
- Merge ELI16 rules into default AskUserQuestion format (always-on)
- Remove _SESSIONS >= 3 conditional — better questions always
- Add simplification rules: plain English, no jargon, no raw function names
- Update tests for branch detection and simplification regression guard

* chore: bump version and changelog (v0.4.3)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
@garrytan @claude
feat: /document-release skill — post-ship doc updates (v0.4.3) (garry…
a68244a 
…tan#109)

* docs: update project documentation for v0.4.2

- README: skill count 9→10, added /document-release to skills table,
  install/uninstall sections, and dedicated section with example
- CHANGELOG: added /document-release bullet to v0.4.2 entry

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat: add /document-release skill with smart VERSION handling

New skill runs after /ship but before PR merge. Reads every doc file,
cross-references the diff, auto-updates factual changes, asks about
risky edits. CHANGELOG clobber protection: never uses Write tool on
CHANGELOG.md, only Edit with exact old_string matches.

Smart VERSION logic: instead of silently skipping already-bumped
versions, compares CHANGELOG entry scope against full diff and asks
if significant uncovered changes exist.

Also fixes gstack-upgrade/SKILL.md missing from skill-check.ts
SKILL_FILES array (existing inconsistency with gen-skill-docs.ts).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat: /review Step 5.6 — documentation staleness check

Review skill now cross-references code changes against doc files.
If a doc describes a feature that changed but the doc wasn't updated,
flags it as INFORMATIONAL with a pointer to /document-release.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* test: /document-release E2E with CHANGELOG clobber guard

E2E test creates a repo with existing CHANGELOG entries, runs
/document-release, and asserts original entries survive. Critical
guardrail against the incident where an agent replaced CHANGELOG
entries during conflict resolution.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* chore: bump to v0.4.3 — /document-release skill

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* chore: regenerate SKILL.md files after merge

* chore: regenerate SKILL.md files after merge

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@garrytan @claude
fix: update check cache — 60min UP_TO_DATE TTL + --force flag (v0.4.4) (
c86faa7 
garrytan#110)

* fix: split update check cache TTL + add --force flag

UP_TO_DATE cache now expires after 60 min (was 720 min / 12 hours).
UPGRADE_AVAILABLE keeps 720 min TTL to keep nagging.

--force flag deletes cache before checking, used by /gstack-upgrade
standalone invocation to always get a fresh result from GitHub.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat: /gstack-upgrade standalone uses --force for fresh check

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* chore: bump version and changelog (v0.4.4)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@garrytan @claude
fix: js statement wrapping + click auto-routes option to selectOption…
318ffdb 
… (v0.4.5) (garrytan#117)

* fix: js statement wrapping + click auto-routes option to selectOption

Bug 1: js command wrapped all code as expressions — const, semicolons,
and multi-line code broke with SyntaxError. Added needsBlockWrapper()
and wrapForEvaluate() helpers (shared with eval) to detect statements
and use block wrapper {…} instead of expression wrapper (…).

Bug 2: clicking <option> refs hung forever because Playwright can't
.click() native select UI. Click handler now checks ARIA role + DOM
tagName and auto-routes to selectOption() via parent <select>.

Bug 3: click timeouts on <option> elements gave no guidance. Now
throws helpful error: "Use browse select instead of click."

* chore: bump version and changelog (v0.4.5)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
@garrytan @claude
feat: Fix-First Review — auto-fix obvious issues, ask about hard ones…
a30f707 
… (v0.4.5) (garrytan#116)

* feat: Fix-First Review — auto-fix obvious issues, ask about hard ones

Replace the CRITICAL-only AskUserQuestion flow with Fix-First:
- Every finding gets action (not just critical ones)
- AUTO-FIX items (dead code, N+1, stale comments) applied directly
- ASK items (security, race conditions, design decisions) batched
  into at most one AskUserQuestion
- Fix-First Heuristic in checklist.md (single source of truth)
- Gate Classification → Severity Classification rename

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* chore: bump version and changelog (v0.4.5)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: polish CHANGELOG v0.4.5 voice — lead with user benefit

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@garrytan @claude
feat: /plan-design-review + /qa-design-review skills (v0.5.0) (garryt…
4a77cc2 
…an#102)

* feat: add {{DESIGN_METHODOLOGY}} resolver and register design review skills

Add generateDesignMethodology() to gen-skill-docs.ts with 10-category, 80-item
design audit checklist. Register plan-design-review and qa-design-review templates
in findTemplates(). Add both skills to skill-check.ts SKILL_FILES. Add command
and snapshot flag validation tests for both skills in skill-validation.test.ts.

* feat: add /plan-design-review and /qa-design-review skills

/plan-design-review: report-only designer audit with letter grades, AI slop
scoring, structured first impression, design system extraction, DESIGN.md
inference and export offer. Never modifies code.

/qa-design-review: same audit, then iterative fix loop with style(design):
commits, CSS-safe WTF heuristic, before/after screenshots, final re-audit.

* chore: bump version and changelog (v0.5.0)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* docs: update README, ARCHITECTURE for design review skills (v0.5.0)

- Update skill count to 11, add /plan-design-review and /qa-design-review
  to skill table, install/uninstall commands, and demo walkthrough
- Add narrative sections: "senior designer mode" and "designer who codes mode"
  with compelling examples showing AI Slop detection and design system inference
- Add {{DESIGN_METHODOLOGY}} to ARCHITECTURE.md placeholder table
- Extend demo to show full plan→eng→review→ship→qa→design-review pipeline

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* chore: regenerate design review SKILL.md files after merge from main

Picks up BASE_BRANCH_DETECT resolver and updated contributor mode from main.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat: add /design-consultation skill — design consultant that creates DESIGN.md

6-phase consultant flow: product context → competitive research (WebSearch) →
complete coherent proposal → drill-downs on demand → font+color preview page →
write DESIGN.md + update CLAUDE.md. Opinionated recommendations grounded in
product context, not menu-driven forms.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* test: add E2E tests for design skill family (7 tests + LLM quality judge)

Tests 1-4: /design-consultation (core flow, research integration, existing
DESIGN.md handling, font+color preview generation).
Tests 5-6: /plan-design-review (audit report, DESIGN.md export).
Test 7: /qa-design-review (audit + fix loop).
LLM judge validates font blacklist compliance, coherence, and AI slop avoidance.
Also adds plan-design-review + qa-design-review to ALL_SKILLS test array.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* chore: mark /design-consultation as shipped in TODOS.md

Renamed from /setup-design-md to reflect the consultant approach.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
@garrytan @claude
docs: add /design-consultation skill to README (garrytan#127)
c8c2cbb 
The skill was fully implemented but completely absent from the README.
Add it to the skill table, write a detailed section with usage example,
and include it in install/uninstall instructions.

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@garrytan @claude
feat: show screenshots to user during QA and browse sessions (v0.5.0.…
5f41cd9 
…1) (garrytan#129)

Add rule 11 to QA and Design methodologies in gen-skill-docs.ts
instructing Claude to Read screenshot PNGs after taking them.
This makes screenshots visible as clickable elements in Conductor
and other Claude Code UIs. Also added to browse and gstack SKILL
templates.

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
@garrytan @claude
feat: Review Readiness Dashboard + gstack-slug helper (v0.5.1) (garry…
73b00b4 
…tan#130)

* feat: add bin/gstack-slug helper + migrate all inline SLUG computation

Extract the opaque SLUG sed pipeline into a shared 5-line shell script.
Replace 8 inline copies across templates with eval $(gstack-slug).
Sanitizes branch names (/ → -) to prevent subdirectory creation.

* feat: review readiness dashboard — track CEO/Eng/Design reviews per branch

Each review skill logs its result to JSONL. A shared {{REVIEW_DASHBOARD}}
placeholder displays run counts, timestamps, and a CLEARED TO SHIP verdict.
/ship pre-flight reads the dashboard and prompts when reviews are missing.

* chore: bump version and changelog (v0.5.1)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
@garrytan @claude
feat: /design-consultation — risk-taking, visual research, ambitious …
c99757b 
…preview (v0.5.2) (garrytan#131)

* feat: /design-consultation — risk-taking thesis, visual research, ambitious preview

Add SAFE/RISK breakdown to design proposals so users see which choices
match category conventions vs. which are deliberate creative departures.

Wire browse binary for visual competitive research — agent browses
competitor sites, takes screenshots, and analyzes fonts/colors/spacing
with graceful degradation to WebSearch-only or built-in knowledge.

Upgrade preview page instructions to render realistic product mockups
(dashboards, marketing pages, settings forms) instead of just swatches.

Rewrite README section with the thesis: "coherence is table stakes —
the real question is where you take risks."

* chore: bump version and changelog (v0.5.2)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: restore SKILL.md files to match main

Prior commit included SKILL.md files regenerated from stale templates.
Restore to match origin/main content.

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
@garrytan @claude
feat: SELECTIVE EXPANSION + smarter ship gates (v0.5.3) (garrytan#134)
5e9f0e7 
* feat: SELECTIVE EXPANSION mode + user control for CEO review

Add 4th mode to /plan-ceo-review: SELECTIVE EXPANSION holds current scope
as baseline but surfaces expansion opportunities one by one for cherry-picking.
All modes now present every scope-expanding idea as individual AskUserQuestion
calls — user opts in or out of each one. EXPANSION recommends enthusiastically,
SELECTIVE recommends neutrally. CEO plan persistence writes decisions to disk.

* feat: review dashboard — eng required, CEO/design optional

Only Eng Review gates shipping. CEO Review recommended for big product
changes, Design Review for UI work — both informational only. Adds
skip_eng_review global config to disable the gate entirely.

* chore: bump version and changelog (v0.5.3)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
@garrytan @claude
feat: always-full eng review + ship review gate persistence (v0.5.4) (g…
b65a464 
…arrytan#135)

Remove SMALL/BIG CHANGE menu from /plan-eng-review — every plan gets the
full interactive review. Scope reduction is now proactive (only when
complexity check triggers) rather than a menu item.

Add review gate override persistence to /ship — when the user says "ship
anyway" or "not relevant", that decision is saved to the branch's
reviews.jsonl so subsequent /ship runs don't re-ask.

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
@garrytan @claude
feat: Test Bootstrap + Regression Tests + Coverage Audit (v0.6.0) (ga…
a2d756f 
…rrytan#136)

* feat: test bootstrap, regression tests, coverage audit, retro test health

- Add {{TEST_BOOTSTRAP}} resolver to gen-skill-docs.ts
- Add Phase 8e.5 regression test generation to /qa and /qa-design-review
- Add Step 3.4 test coverage audit with quality scoring to /ship
- Add test health tracking to /retro
- Add 2 E2E evals (bootstrap + coverage audit)
- Add 26 validation tests
- Update ARCHITECTURE.md placeholder table
- Add 2 P3 TODOs (CI/CD non-GitHub, auto-upgrade weak tests)

* chore: bump version and changelog (v0.6.0)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* feat: make coverage audit trace actual codepaths, not just syntax patterns

Step 3.4 now instructs Claude to read full files, trace data flow through
every branch, diagram the execution, and check each branch against tests.
Phase 8e.5 regression tests now trace the bug's codepath before writing
the test, catching adjacent edge cases.

* feat: coverage audit now maps user flows, interactions, and error states

Step 3.4 now covers the full picture: code branches AND user-facing behavior.
Maps user flows (complete journey through the feature), interaction edge cases
(double-click, back button, stale state, slow connection), error states
(what does the user actually see?), and boundary states (zero results,
10k results, max-length input). Coverage diagram splits into Code Path
Coverage and User Flow Coverage sections with separate percentages.

* fix: raise test gen cap to 20, add validation tests for user flow coverage

- Raise Step 3.4 test generation cap from 10 to 20 (code + user flow combined)
- Add 3 validation tests: codepath tracing, user flow mapping, diagram sections

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
@garrytan @claude
feat: /gstack-upgrade detects and syncs stale vendored copies (v0.5.4…
1f3b691 
….1) (garrytan#137)

When the global gstack is already up to date, standalone /gstack-upgrade
now checks if the local vendored copy in the current project is at a
different version and syncs it automatically. Also adds rollback on
setup failure and update-check fallback matching the preamble pattern.

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
@garrytan @claude
feat: Completeness Principle — Boil the Lake (v0.6.1) (garrytan#140)
9d47619 
* feat: Completeness Principle — Boil the Lake (WIP, pre-merge)

Add Completeness Principle to all skill preambles, dual-time estimates,
compression table, anti-pattern gallery, Lake Score, and completeness
gaps review category. VERSION/CHANGELOG will be rebased after merge.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: update stale version reference in TODOS.md (v0.5.3 → v0.6.1)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: update CHANGELOG date + README for v0.6.1 features

- Add date to CHANGELOG 0.6.1 entry
- Add Completeness Principle to README intro
- Add SELECTIVE EXPANSION mode to CEO review section
- Add test bootstrap mention to /ship section
- Fix uninstall command missing design-consultation in project uninstall
- Add "recommends shortcuts" and "no tests" to Without gstack list

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: split README into lean intro + docs/ directory (gh CLI pattern)

README: 875 → 243 lines. Keeps intro, skill table, demo, install, and
troubleshooting. All per-skill deep dives, Greptile integration guide,
and contributor mode docs moved to docs/ directory.

- docs/skills.md — full philosophy and examples for all 13 skills
- docs/greptile.md — Greptile setup and triage workflow
- docs/contributor-mode.md — how to enable and use contributor mode
- README now links to docs/ via Documentation table
- Updated skill table entries with latest features (fix-first, regression
  tests, test health, completeness gaps)
- Updated demo transcript with AUTO-FIXED, coverage audit, regression test

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: remove "competitor" language, rewrite README in Garry's voice

Replace "browses competitors" with "knows the landscape" / "what's out
there" throughout all user-facing copy. Trim README from 243 to 167
lines — tighter, more opinionated, less listicle energy. Remove
Completeness Principle from README top (it lives in CLAUDE.md and the
skill preambles where Claude actually reads it).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: rewrite README in Garry's raw voice — AGI era, L8 factory, real stories

The README now sounds like Garry, not a product page. Leads with the
live experiment, the 16k LOC/day reality, the real-life coding stories
(Austin, hospital bedside). Highlights the newest unlocks (design at
the heart, /qa parallelism, smart review routing, test bootstrap).
Closes with an open invitation — free MIT, fork it, let's all ride
the wave together.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: add Garry's bonafides to README intro — Palantir, Posterous, YC, 600k LOC

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: add real /retro numbers — 140k lines, 362 commits across 3 projects

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: add "in the last 60 days" timeframe to 600k LOC claim

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: add GitHub contribution graphs — 2026 vs 2013 side by side

Same person, different era. 2013: 772 contributions building Bookface.
2026: 1,237 contributions and accelerating. The difference is the tooling.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: clarify /retro stats are from last 7 days

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: add designer/PM/eng manager roles to intro

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: remove Josh/L8 reference from README

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: move demo up, make it dramatically more impressive

Show the actual architecture diagram, auto-fixed issues, 100% coverage,
regression test generation. Punch line: "That is not a copilot. That is
a team."

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: remove "My journey" section — intro already covers it

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: prefix all skill commands with You: in demo transcript

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: collapse You/Claude lines in demo — no gap between command and response

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: clarify plan mode flow in demo — approve, exit, Claude implements

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: move /ship to end of demo — review → QA → ship is the real flow

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: add /plan-design-review to demo, tighten CEO response

Shorter CEO reply, compressed eng diagram, added design audit with
AI Slop score. Seven commands now: plan → eng → build → design →
review → QA → ship.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: move design review before implementation — it's part of planning

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: reorder demo — design before eng, after CEO

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: remove URL from /plan-design-review in demo

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: add [...] annotations showing what actually happens at each step

Each step now shows what the agent does under the hood: 8 expansion
proposals cherry-picked, 80-item design audit, ASCII diagrams for
every flow, 2400 lines written in 8 minutes, real browser QA, bug
found and fixed. Makes the demo feel real, not abstract.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: rename Contributor Mode to How to Contribute in docs table

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: add Coinbase, Instacart, Rippling to YC bonafides

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: add "one or two people in a garage" to founder story

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: add skill table to top of skills.md with anchor links

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: consolidate — roll contributor-mode into CONTRIBUTING, greptile into skills

- docs/contributor-mode.md → merged into CONTRIBUTING.md (session awareness section)
- docs/greptile.md → merged into docs/skills.md (Greptile integration section)
- Reordered docs table: Skills > Architecture > Browser > Contributing > Changelog

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@garrytan @claude
feat: diff-based test selection for E2E and LLM-judge evals (v0.6.1.0) (
17c1c06 
garrytan#139)

* feat: diff-based test selection for E2E and LLM-judge evals

Each test declares file dependencies in a TOUCHFILES map. The test runner
checks git diff against the base branch and only runs tests whose
dependencies were modified. Global touchfiles (session-runner, eval-store,
gen-skill-docs) trigger all tests.

New scripts: test:e2e:all, test:evals:all, eval:select

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* chore: bump version and changelog (v0.6.1.0)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: plan-design-review-audit eval — bump turns to 30, add efficiency hints

The test was flaky at 20 turns because the agent reads a 300-line SKILL.md,
navigates, extracts design data, and writes a report. Added hints to skip
preamble/batch commands/write early while still testing the real SKILL.md.
Now completes in ~13 turns consistently.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
@garrytan @claude
feat: cognitive patterns for plan-review skills (v0.6.2) (garrytan#141)
d8894b7 
* feat: cognitive patterns for plan-review skills — latent space activation

Enrich /plan-ceo-review, /plan-eng-review, and /plan-design-review with
researched cognitive patterns from Bezos, Grove, Munger, Horowitz, Altman,
Rams, Norman, Zhuo, Gebbia, Larson, McKinley, Brooks, Beck, and Majors.
Patterns are evocative activation keys, not checklists — they trigger the
LLM's deep knowledge of how these people actually think.

* chore: bump version and changelog (v0.6.2.0)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
@garrytan @claude
feat: design review lite in /review and /ship + gstack-diff-scope (v0…
28becb3 
….6.3) (garrytan#142)

* feat: gstack-diff-scope helper + design review checklist

bin/gstack-diff-scope categorizes branch changes into SCOPE_FRONTEND,
SCOPE_BACKEND, SCOPE_PROMPTS, SCOPE_TESTS, SCOPE_DOCS, SCOPE_CONFIG.

review/design-checklist.md is a 20-item code-level checklist with
HIGH/MEDIUM/LOW confidence tags for detecting design anti-patterns
from source code.

* feat: integrate design review lite into /review and /ship

Add generateDesignReviewLite() resolver, insert {{DESIGN_REVIEW_LITE}}
partial in review Step 4.5 and ship Step 3.5. Update dashboard to
recognize design-review-lite entries. Ship pre-flight uses
gstack-diff-scope for smarter design review recommendations.

* test: E2E eval for design review lite detection

Planted CSS/HTML fixtures with 7 design anti-patterns. E2E test
verifies /review catches >= 4 of 7 (Papyrus font, 14px body text,
outline:none, !important, purple gradient, generic hero copy,
3-column feature grid).

* chore: bump version and changelog (v0.6.3.0)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
@garrytan @claude
docs: restructure README for faster conversion (garrytan#146)
f91222f 
* docs: restructure README for faster conversion

Move install block to top (first scroll) for visitors who are already sold. Add "Who this is for" self-identification section and "Quick start: 10 minutes" funnel to guide first-time users. Condense install instructions to bold literal commands. Move "See it work" and "The team" below install. Relocate hiring box after "Come ride the wave" section. Strip redundant section dividers. Add governance framing to final CTA and new troubleshooting case for missing skills.

Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>

* docs: update "Who this is for" to lead with founders and new CC users

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Haiku 4.5 <noreply@anthropic.com>
@HMAKT99
feat: add /eval skill — AI output evaluator and grader
b087598
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@HMAKT99 @garrytan