Both deep-research and team-update skills hardcoded
~/Documents/basalt-cortex/ as a knowledge-vault location. This leaked
internal Jezweb setup into a public skill — most users don't have that
directory. Basalt Cortex is also now retired internally (the shared-
knowledge role moved to .jez/knowledge/ per 2026-04-22 team decision).
Reworded to generic "personal knowledge vault if you keep one" framing:
- deep-research/SKILL.md Step 2: point at Obsidian/Documents/notes/
generically, keep the grep-for-KEYWORD pattern as the useful bit
- deep-research/SKILL.md Step 7: drop "Basalt Cortex project" proper
noun from the Cloudflare-capabilities anecdote, keep the substantive
"Workers AI toMarkdown, Vectorize metadata filtering, D1 FTS5" detail
- team-update/SKILL.md Knowledge row: generic personal-vault pattern
- team-update/references/discovery-patterns.md: same
Closes internal follow-up on basalt plugin removal (team-skills 07eafb0).
Raised per Jez greenlight in Dev Chat (msg nP6g6buoQSY).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three principles for audit runs that catch the most bugs:
- Drive the audit from the main session, not a sub-agent (cross-interaction
state lives in the watching session; sub-agents start cold and miss
second-order findings)
- Use the browser tool directly (Chrome MCP / Playwright MCP via main
session calls — not screenshots handed to a fresh agent)
- Loop to exhaustion with variations (different persona/workflow/input
each pass; stop only when a full pass produces no new findings)
Plus: optional 15-min /loop check-in for audits expected to run >30 min.
Not a work-doing loop — a supervisor prompt that journals findings,
confirms activity, and provides a natural termination signal.
New reference file long-running-check-in-pattern.md documents the
check-in pattern generically — when to set up, supervisor prompt shape,
when to tear down, and the structural parallel to specialist heartbeats
at nightly/hourly/15-min scales.
Pattern surfaced by Jez's usage discovery: loop-to-exhaustion-with-
variations catches logic errors, business-logic issues, and odd edge
cases that single-pass audits miss.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Chrome on macOS Retina captures at 2x logical viewport (1440→2880px),
bloating context and hitting vision-pipeline size limits. Both skills
now instruct the agent to run img-process batch after each capture
phase to downsize to 1440px. The optimise action is idempotent —
no-op on already-small files — so it's safe to run unconditionally.
Playwright MCP users can set deviceScaleFactor:1 and skip the step.
The agent-browser skill duplicates what Vercel ships as a native
Claude Code plugin at vercel-labs/agent-browser. Users should
install from the source repo which is actively maintained.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Rebuilt from v1 archive as a proper v2 skill that produces working test
infrastructure. Detects project type, generates vitest.config.ts, test
setup, utilities, sample tests, and package.json scripts. Covers mocking
patterns, coverage, workspace config, and Jest migration.
Prompted by issue #69 (Dispatch discoverability).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Fixes#54. Visual browsing misses API failures that the UI hides —
TanStack Query and similar libraries swallow HTTP errors and show
empty/loading states. A People picker showing "No users found" because
of a 403 route collision looks normal visually.
New section monitors network responses throughout the audit:
- Collects all non-2xx API responses with URL, status, method, page
- Severity mapping: 500=Critical, 403=High (route collision class),
404=Medium, 401=Low, CORS=High
- Catches: route collisions, silent endpoint failures, CORS issues,
auth middleware bugs, missing endpoints after refactoring
- Report template includes Network Errors table
Runs in thorough + exhaustive modes.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
app-docs rewritten with:
- Exhaustive depth level (150+ screenshots, publishable HTML suite)
- Per-page documentation template with What You'll See / What You Can Do / How To
- Supporting sections: getting started, navigation guide, keyboard shortcuts, troubleshooting, admin guide
- HTML output option with TOC sidebar and print CSS
- Workflow diagrams and annotated screenshot patterns
- Quality bar: "a new user can complete any task by following the guide"
- Mockup and diagram techniques alongside screenshots
product-showcase: added mockup + screenshot mix guidance with CSS phone
frame for mobile screenshots, guidance on when to use real screenshots
vs simplified mockups for different sections.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Competitor deep-dive now includes: documentation quality (what needs
most explanation = what users struggle with), help/support content
(most-asked questions), onboarding/tutorials (how they teach, what they
assume), API documentation (patterns, SDKs), and migration guides
(reveal advantages AND switching friction).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New fourth depth level: exhaustive (4-8+ hours)
- Inventories every interactive element on every page (buttons, links,
inputs, selects, toggles, tabs, accordions, modals, dropdowns, etc.)
- Clicks/interacts with every one, screenshots each state
- Tests every form with valid, invalid, and empty data
- Tests every filter value, sort option, tab, combination
- Dark mode + light mode for every page and dialog
- Three viewport widths for every page and dialog
- Keyboard navigation through every page
- Context menus, edge states, concurrent tabs, every error path
- Per-page element inventory with pass/fail tracking
- Writes findings incrementally to report (doesn't hold in memory)
Depth progression: quick (5 min) → standard (30 min) → thorough (3 hrs) → exhaustive (8+ hrs)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
stripe-payments: Checkout Sessions, Payment Intents, subscriptions,
webhooks (with constructEventAsync for Workers), customer portal,
pricing pages. Decision table for which API to use. Cloudflare Workers
gotchas. No MCP server needed.
Argument hint cleanup: removed [url] and [file-path] from hints —
replaced with meaningful choices (depth levels, modes, purposes).
The URL/file is auto-detected, the hint shows what the user decides.
product-showcase: added depth levels (quick/standard/thorough).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Claude's training data is always behind on platform features. This section
forces active research of the FULL platform catalogue — not just the
services you already use. Includes changelog/blog fetching, service-by-
service evaluation checklist (Cloudflare: 25+ services across compute,
storage, AI, networking, security, media, email, observability), and
capture template.
The Basalt Cortex project exists because of capabilities (Workers AI
toMarkdown, Vectorize metadata filtering, D1 FTS5) that weren't obvious
without actively looking. Most developers use 20% of their platform
because they never looked at the other 80%.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
BREAKING: dev-session skill removed. Its useful bits are now in roadmap.
Roadmap skill now has four modes:
- plan: generate the roadmap (was the only mode before)
- start: begin executing from phase 1, don't stop until done or stuck
- resume: pick up from the last session, keep going
- status: show progress across phases
Execution rules: commit at phase boundaries, deploy after each phase,
quick audit between phases, thorough audit + onboarding-ux + project-docs
after the final phase. Only stop for blockers that need human input.
Progress tracked directly in ROADMAP.md with checked tasks and completed
phase markers — no separate SESSION.md needed.
dev-session was a 200K-era survival tool. At 1M context, the roadmap
itself is the session state.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
dev-session: removed "context getting full" trigger, "survive compaction"
framing, compaction-survival.md reference. Reframed as cross-session
handoff tool, not a context survival mechanism.
ux-audit: "track progress so work survives context limits" → "track
progress across the audit"
deep-research: "don't try to hold everything in context" → "the artifact
is the deliverable"
At 1M context, most features complete in a single session. These skills
are still valuable for genuinely multi-session work, but the urgency
framing was outdated.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Produces a ROADMAP.md detailed enough that "build phase N" is a complete
instruction for Claude Code. Includes: vision, stack decisions, complete
data model upfront, 4-8 phases with task checklists grouped by layer,
schema evolution map, API surface map, build order table, and
"deliberately not building" scope control.
Designed for all-day autonomous execution — each phase is independently
deployable with a definition of done. Pairs with deep-research (input)
and fork-discipline/ux-audit/onboarding-ux (post-phase checks).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
cloudflare-api: Hit the Cloudflare REST API directly for operations that
wrangler and MCP can't handle well. Covers bulk DNS, custom hostnames,
email routing, cache purge, WAF rules, redirect rules, zone settings,
Worker routes, and GraphQL analytics. Produces curl commands or scripts.
Not a docs mirror — workflow recipes for real operations.
deep-research: Updated argument hints to explain what focused/wide/deep
actually mean inline, so you don't have to remember.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Three modes:
- audit: map the boundary, find violations (hardcoded client checks,
config replacement, scattered client code, missing extension points,
migration conflicts), produce a health score and report
- document: generate FORK.md documenting the boundary and how to add clients
- refactor: concrete steps to move client code out of core, replace client
checks with feature flags, implement config merge, add extension points
Includes the "right time to run" guide: don't refactor at 1 client,
audit at 2, full refactor at 3+.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Audits a web app for onboarding gaps (empty states, missing guidance,
feature discoverability, contextual help), then generates the actual
content and code to fix them:
- Empty state components with copy and CTAs
- Welcome banner or onboarding checklist
- Feature tour configuration (react-joyride pattern)
- Tooltip content map for every complex field
- Contextual hints (dismissable, shown once)
- Onboarding-quality seed data
- Help/FAQ page content
Pairs with ux-audit: audit finds problems, onboarding-ux builds solutions.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds URL resolution that prefers production: checks wrangler.jsonc for
custom domains, then CLAUDE.md/README for deployed URLs, then falls back
to running local dev server, then asks. Live site has real data, real
latency, real CORS — that's what users experience.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Depth levels redefined by scope of ambition not time:
- focused: answer one specific question
- wide: understand the landscape (enough to write a spec)
- deep: leave no stone unturned (enough to drive weeks of coding)
Added library/component research section (deep mode): npm packages,
React libraries, headless UI primitives, hooks, platform-specific
packages. Research brief template now includes library evaluation table.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
BREAKING: knowledge-cortex plugin removed. Replace with basalt plugin.
- /basalt:basalt-cortex replaces /knowledge-cortex:cortex-mine and cortex-query
- Basalt format: Obsidian-compatible markdown with YAML frontmatter (not JSONL)
- Multi-source: Gmail, Google Chat, Slack, Drive, local files, MCP, web, calendar
- Pattern-based teaching (no shipped scripts)
- References: basalt-format.md, source-patterns.md, prefilter-patterns.md, sync-patterns.md
NEW: deep-research skill in dev-tools
- Deep research and discovery before building something new
- Three depths: focused (30 min), wide (1-2 hours), deep (3-6 hours)
- Local exploration, competitor analysis, ecosystem signals, future-casting
- Produces comprehensive research brief to .jez/artifacts/
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New reference: scenario-tests.md with six structured tests:
1. New hire onboarding — figure out the app with zero guidance
2. Interrupted workflow — close tab mid-task, come back, see what survived
3. Explain it to a colleague — write a 2-min guide, gaps = UX failures
4. Wrong turn recovery — go to wrong page, measure recovery cost
5. Day two — repeat tasks, measure efficiency gains and shortcuts
6. What changed? — log in after absence, can you tell what needs attention?
Thorough mode now runs all six after QA sweep. Report template includes
scenario test results with tables and the explain-it guides (which
double as user documentation).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New reference: workflow-comprehension.md — evaluates whether users can
follow the threads of their real work through the app. Covers workflow
threads, wayfinding, visual cues/nudges, page-to-page continuity,
mental model alignment, and productivity assessment.
Thorough mode now starts by identifying workflow threads before testing
individual pages. Per-screen evaluation adds comprehension, wayfinding,
and flow questions. Report template includes workflow findings section.
The shift: from "does this page work" to "does the user understand the
system and can they follow their work through it."
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Thorough mode now includes:
- Automated JS injection to detect overflow, clipped text, off-screen
elements, and invisible text on every page
- Responsive breakpoint sweep at 1280/1024/768/375px widths
- Console error sweep (JS errors, failed requests, framework warnings)
- Accessibility basics (heading hierarchy, alt text, contrast)
- Data edge cases (long text, special chars, empty, rapid clicks)
- Navigation/routing tests (deep links, back button, 404s, stale URLs)
Designed to run unattended overnight and produce a complete report.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds a comprehensive Visual & Layout Inspection section covering clipped
text, overlapping elements, overflow, broken grids, contrast, sidebar
misalignment, responsive breakpoints, image issues, and invisible elements.
Each issue type includes what to look for and common cause.
Layout check added to per-screen evaluation and Step 3 key questions.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Created 45 command files with descriptive argument-hints so every skill
has a /plugin:skill-name shortcut. Added compatibility: claude-code-only
to 17 skills that require Bash/filesystem/browser (d1-migration, db-seed,
project-docs, git-workflow, landing-page, shadcn-ui, tailwind-theme-builder,
ai-image-generator, color-palette, favicon-gen, icon-set-generator, etc).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
ux-audit: add quick/standard/thorough depth levels. Thorough mode runs
autonomously overnight — multiple personas, all pages, all states,
resilience testing, comprehensive report with screenshots.
app-docs: browse a running app, screenshot every screen, produce a
complete user guide with step-by-step instructions.
product-showcase: explore a running app, capture key screens, generate
a marketing site (single HTML) with real screenshots in browser-frame
mockups, feature grid, how-it-works, and CTAs.
49 → 51 skills.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New skills batch (44 → 49 skills, 10 → 11 plugins):
- project-docs: generate ARCHITECTURE.md, API_ENDPOINTS.md, DATABASE_SCHEMA.md from codebase
- landing-page: single-file deployable HTML with Tailwind CDN
- db-seed: realistic seed scripts for D1/Drizzle with FK ordering and batch limits
- ai-image-generator: Gemini/GPT image gen with 5-part prompting framework
- social-media-posts: platform-formatted posts for LinkedIn, FB, Instagram, Reddit
New plugin: social-media
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds gws CLI commands (auth, gmail, calendar, drive, sheets, docs, chat,
tasks, slides, forms, people, admin) to permission presets with explicit
subcommand entries to work around the wildcard matching bug.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
.claude/ is a protected system directory that triggers permission prompts.
Using .jez/ for artifacts, scripts, and screenshots avoids this.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
README, CLAUDE.md, plugin.json, marketplace.json all missing agent-browser.
Updated skill count from 24 to 36 and plugin count from 7 to 9.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace thin tone section with concrete rules: no emoji spam, no hedging,
no restating context the team already has. Includes bad/good examples.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Each consultation gets its own folder under .claude/artifacts/brains-trust/
with prompt and responses preserved for future reference.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When triggered without arguments, brains-trust now infers scope from session
context and defaults to consensus mode. Adds progress output rule for UX.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Multi-model consultation skill that queries leading LLMs (Gemini, GPT, Qwen,
DeepSeek, etc.) for second opinions on code, architecture, strategy, prompting.
Uses models.flared.au for runtime model discovery instead of hardcoded IDs.
Supports single opinion, consensus (parallel multi-model), and devil's advocate.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Wildcard Bash(sqlite3 *) may not match :memory: arguments due to
the known Claude Code permission matching bug. Add explicit entry.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Same workaround as gh — broad Bash(npx *) can miss subcommands.
Adds explicit npx entries for tsc, tsx, vite, vitest, drizzle-kit,
shadcn, better-auth, playwright, tailwindcss, create-cloudflare,
create-next-app, astro, lighthouse, repomix. Also adds npx prisma
to Database preset and npx wrangler@latest to Cloudflare preset.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Remove SEO plugin (moved to jez-skills where MCP-dependent skills belong)
- Add 20+ commonly used commands to permission presets (Linux, Mobile, etc.)
- Add ERRATA.md convention to skill anatomy for versioned corrections
- Fix dev-tools plugin description (removed stale "Skill creation" ref)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Skills now teach Claude what to do instead of shipping pre-built scripts.
Claude generates scripts adapted to the user's environment (OS, runtime,
installed packages) rather than running rigid pre-built ones.
- Remove skill-creator skill (use Anthropic's official one instead)
- Remove all 17 pre-built scripts across 7 skills
- Add 5 reference files with annotated implementation patterns
- Rewrite 6 SKILL.md files from "run this script" to pattern-based
- Update CLAUDE.md with "teach patterns" philosophy
- Net: -2,631 lines deleted, +109 added
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Permission Auditor now checks for missing file access patterns
(Read/Edit/Write for .claude/**, //tmp/**) in project settings and
home-relative paths (~/Documents/**, ~/.claude/**) in global settings.
Setup step notes that home paths belong in global settings only.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Pre-approve Read/Edit/Write for .claude/** (project scripts, artifacts,
screenshots), //tmp/**, and document home-relative paths for global
settings. Prevents "allow access to" prompts when skills read rules,
cross-project files, or write to .claude/ subdirectories.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Skill scripts (gemini-peer-review, gemini-image-gen, image-processing,
project-health agents) all use python3 — should be pre-approved in every
project, not just Python-language projects. Also add GEMINI_API_KEY=*
pattern for Gemini skill env var prefix.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
gemini-peer-review no longer requires gemini-coach or gemini CLI.
New scripts/gemini-review.py uses urllib to call the Gemini API
directly — zero dependencies beyond Python 3. Prompts written to
.claude/artifacts/ to avoid shell escaping issues. API key from Vault.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Jeremy Dawes