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>
Revised the skill to incorporate gpt-image-2 capabilities that change
when/how to reach for GPT vs Gemini:
- Text rendering actually works (multi-script) — now the primary reason
to choose GPT over Gemini, superseding GPT Image 1.5 for this use
case
- 10-variation batch generation — one prompt, 10 images sharing
composition and palette, good for style exploration + A/B options
- Multi-reference compositing — product + lifestyle scene with correct
lighting, scale, perspective (enables "product in context" workflows
without multi-turn editing)
- Instant vs Thinking modes (Thinking = Plus/Pro/Business, plans layout
before drawing — fewer re-rolls on complex compositions)
- Aspect ratios 3:1 to 1:3 (wider than Gemini)
- 2K standard, 4K beta
Critical constraint documented: GPT Image 2 cannot do transparency.
Fall back to gpt-image-1.5 for transparent icons and logos only.
Generation time up to 2 min → build async UX with 180s timeout.
Updates:
- Model Selection table refined with gpt-image-2 routing + rule of
thumb (text → GPT 2, otherwise Gemini)
- Model IDs table adds gpt-image-2 and chatgpt-image-latest
- New "GPT Image 2 Specifics" section covering modes, batching,
multi-reference, aspect ratios, resolution, generation time,
constraints, pricing
- Determine Image Need table adds text-heavy and batch rows, plus
splits hero/OG by "with copy" vs "no text" for clear routing
- New code example for 10-variation batch with rendered text (pricing
card with headline, price, feature lines)
- Common Mistakes adds four new gotchas: gpt-image-2 no transparency,
gpt-image-1.5 unreliable text (use 2 instead), blocking on 2min
request, plus the existing Gemini transparency note
SKILL.md 389 lines (was 273), lint passes. 5-part framework, multi-
turn editing, API setup all retained unchanged.
Surfaced from Jez's summary in AI Everything Chat space 2026-04-22
09:11 AEST. Pre-empted Thursday's caretaker-sweep platform watch by
six hours.
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>
Update product-showcase and image-processing SKILL.md to use capture-screenshots and img-process CLI as primary approach. Custom scripts become the fallback for complex operations. Closes#70.
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>
Inlined CSV format reference and GraphQL mutations. 2 refs removed.
SKILL.md: 249 -> 462 lines.
Part of #65 (6 of 14 skills). Delegated to sub-agent worker.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Inlined shade generation, semantic mapping, dark mode palette, contrast
checking. All 4 refs removed. SKILL.md: 119 -> 412 lines.
Part of #65 (4 of 14 skills). Delegated to sub-agent worker.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Inlined ModeToggle component, 18 gotchas, architecture deep-dive,
color variable flow diagrams, dark mode switching flow. 3 refs removed,
migration-guide.md kept as optional. SKILL.md: 173 -> 419 lines.
Part of #65 (3 of 14 skills). Delegated to sub-agent worker.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Agents skip reference files and improvise. All critical-path content inlined into SKILL.md. 7 reference files removed, only integration-patterns.md kept as optional. SKILL.md: 182 -> 446 lines. Part of #65.
Deployment: Cloudflare Workers with Static Assets (recommended), Pages
drag-and-drop, or any static host. Custom domain guidance.
Enquiry forms: three patterns — Formspree (no backend), Worker endpoint
(SMTP2Go/Resend), Turnstile spam protection. Form field recommendations
and post-submission UX.
Additional pages: pricing, about, changelog, docs link, blog, privacy/
terms, status page — with guidance on when to include each.
SEO: per-page meta tags, OG images, canonical URLs, JSON-LD reference.
New quality rules: forms must submit somewhere, every page needs nav,
deploy to a real URL (not local files).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Major rewrite — the skill now produces multi-page marketing websites,
not just single-page marketing:
Standard mode: Home + Features + How It Works pages with real screenshots
Thorough mode: Per-feature deep-dive pages, animated GIF walkthroughs,
use case scenario pages ("A day in the life of..."), comparison page
New: Animated GIF capture — record 2-4 second workflow sequences as GIFs
using Playwright frame capture + PIL. Shows the app actually working,
not just static screenshots.
New: Agentic app marketing patterns — before/after, timeline, result
showcase, side-by-side, magic moment GIFs. Copy tips for invisible AI
value ("Know every client's history" not "AI-powered CRM").
New: Per-feature pages with hero screenshot, how it works, why it
matters, power-user tips. Use case pages with story-driven scenarios.
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>
Basalt Cortex depends on Gmail and Google Chat MCP servers, making it
team-specific rather than publicly shareable. Fully removed from
claude-skills (not just a stub). Install from jezweb/team-skills instead.
10 plugins, 59 skills.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Basalt Cortex skill moved to jezweb/team-skills (private) since it
depends on Gmail and Google Chat MCP servers. This stub remains for
backwards compatibility and directs users to install from team-skills.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Contact type now supports multi-identifier lookup (emails, phones,
Google Chat IDs, websites) for identity resolution
- Field catalog updated with google_chat_id, additional_emails, website fields
- Communication format updated with proportional summaries and freeform fields
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>
Depth levels: quick (hero + features), standard (full showcase), thorough
(every section including dark mode, shortcuts, before/after, integrations).
New optional sections: before & after comparison, dark mode showcase,
keyboard shortcuts / power features, integrations ecosystem, open source.
Premium techniques: screenshot enhancement tips (clean data, consistent
size, highlight the action), visual storytelling (lead with dashboard,
show results not forms, progressive reveal), CSS-only animations (fade-in,
hero float), gradient accents from app's brand colour.
Value proposition extraction: benefit-focused copy ("Find anything in
seconds") not feature descriptions ("Search functionality").
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>
Comprehensive setup skill covering Docker cgroup fixes, OpenShell installation,
sandbox creation, persistent port forwarding with health checks, Cloudflare
Tunnel remote access, workspace configuration, and workarounds for known bugs
(device auth, token mismatch, rate limiting). Based on real-world setup experience
on NVIDIA Grace Hopper GX10.
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>
Searches Gmail for dispatch emails, identifies carriers from tracking
number formats, builds direct tracking links for AusPost, StarTrack,
Sendle, DHL, FedEx, UPS, and provides paste-and-track links for Aramex,
CouriersPlease, Toll, Team Global Express, Hunter Express, and more.
Works with any available tools (Gmail MCP, web scraper, browser automation)
or degrades to providing the correct tracking link from a user-supplied number.
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>
Jeremy Dawes