Third edition. Skills Curator now ships for claude.ai (web + desktop) in addition to the two existing Claude Code editions (Lite and Python full). Same judgment model, same registry schema — re-architected for a runtime that has no persistent ~/.claude/ filesystem and no slash commands. The Claude Code editions are unchanged in behavior; this is purely additive.

Added

• skills/skills-curator-claudeai/ — new edition bundle for the claude.ai Skills feature. Slim SKILL.md (~250 lines) with progressive disclosure to four reference files. Designed to be zipped and uploaded via claude.ai → Settings → Capabilities → Skills.
• skills/skills-curator-claudeai/references/catalog.yaml — 20-entry curated catalog, mirrored from the Lite edition's embedded YAML. Loaded on demand during RECOMMEND to keep activation cost low.
• skills/skills-curator-claudeai/references/signals.md — framework + goal signal extraction tables and the Tag-rivals stack-mismatch detector. Loaded on demand during RECOMMEND and CUSTOMIZE.
• skills/skills-curator-claudeai/references/security-patterns.md — 15 severity-tagged regex patterns for pre-install scanning. Loaded on demand during EVALUATE and SCAN.
• skills/skills-curator-claudeai/references/persistence.md — deep dive on the three persistence modes (Project Knowledge / upload / Gist sync), schema field-by-field, migration rules, troubleshooting.
• skills/skills-curator-claudeai/INSTALL.md — end-user install guide: prerequisites, zip command (PowerShell + bash), upload steps, Project Knowledge setup for Mode A persistence, verification check.
• New verb on the claude.ai edition: SCAN — lightweight quick-scan that runs only the security patterns and returns a findings table, no full ADOPT/PARTIAL/SKIP verdict. Use when the user just wants a safety check without adoption advice.
• /mnt/user-data/outputs/ persistence flow — the claude.ai edition writes the updated registry to a downloadable file rather than pasting it back into chat. Eliminates the context-cost problem registries would otherwise cause.
• CUSTOMIZE outputs a zip — on the claude.ai edition, the headline CUSTOMIZE verb packages the fork as /mnt/user-data/outputs/<fork-name>.zip so the user can upload it back to claude.ai (or move to ~/.claude/skills/ if they're also on Claude Code) in one step.
• README "Editions" section replacing the old "Two tiers" section. Three-column comparison covering runtime, engine, install, persistence, verbs supported, and tradeoffs. Old "Two tiers" anchor preserved as a redirect via the new heading.

Changed

• Plugin description in .claude-plugin/plugin.json updated to describe three editions, not two. New claude-ai keyword added.
• Plugin version 4.4.5 → 4.5.0 in plugin.json, registry.py VERSION constant, and the Python tier's SKILL.md metadata. Lite tier's internal version (2.0.0) is unchanged — Lite tracks separately and didn't change in this release.
• README header status line rewritten to mention the third edition. Editions badge replaces the old Tiers badge.

Notes

• The claude.ai edition is not auto-installed by npx skills add captkernel/Skills_Curator — that command installs the Claude Code editions to ~/.claude/skills/. The claude.ai edition lives in the repo at skills/skills-curator-claudeai/ and is installed via the standalone zip + upload flow documented in its INSTALL.md. Adding it to plugin.json's skills[] array would incorrectly drop the claude.ai variant into Claude Code installations.
• Registry JSON schema is shared across all three editions (v3.0). A registry written by one edition can be read by another. Power users can use Gist sync (the Python tier's existing --sync feature, or the claude.ai edition's Mode C) to roam a single registry across all three runtimes.
• claude.ai isn't one of the engine's 55 supported platforms — those are all CLI/IDE agents. --migrate does not produce a claude.ai-compatible bundle. This was a manual port informed by claude.ai's runtime constraints (sandbox-per-conversation, no ~/.claude/, no slash commands, Project Knowledge as the only durable storage Claude can read on its own).
• No engine code changed. All 37 pytest cases continue to pass.

README banner removed. The repo now opens with the title + badges (no hero image above the title). Cleaner first impression that matches what most well-designed CLI-tool repos do (HTTPie, DVC, Bun, etc.) — the value-prop tagline carries the opening, with content-rich images (the Claude Code session screenshot, the --customize flow) reserved for the Demo and feature sections below.

Removed

• docs/images/hero.png no longer ships in the repo. It was the warm-cream editorial cover; useful as a separate brand asset but unnecessary as a banner. The README opens with the title and badges directly.

Notes

• docs/images/ is now six files: customize-workflow.png + the five screenshot-*.png. Every shipped image is content (a diagram or a real product screenshot), not pure branding.
• The local .share/images/hero.png is retained as a social asset (X tweet attachments, LinkedIn share, manual GitHub Settings "Social preview" upload), but it lives only in the local social-launch folder.

Image set tightened. README hero reverted to the editorial cover. Every diagram regenerated with real data — no AI-hallucinated platform names, no abstract block diagrams that don't carry information.

Changed

• README hero is the editorial cover again (docs/images/hero.png). The Clair Obscur painterly direction lived briefly in v4.4.2 — pulled back in favor of the cleaner brand identity.
• docs/images/customize-workflow.png regenerated as a three-lane diagram. Source column shows the external skill as authored (6 sections). Plan column scores each section and tags it keep, keep-trim, rewrite, rewrite-fm, or drop — with drop sections visibly struck through and excluded. Final column ("Your Project") contains only the surviving sections; dropped sections are physically absent, with a footer noting their exclusion. Visually proves "infuse, don't invoke" in a single image.

Removed

• docs/images/brand-curator-portrait.png and docs/images/brand-gallery-hallway.png — the Clair Obscur brand imagery is no longer in the public repo. README banner uses the editorial cover.

Notes

• The local .share/ gallery (not shipped) was also reorganized: engine concept replaced with a live-state dashboard view, cross-agent migration replaced with a real-data platform matrix (55 platforms, organized by tier, every install path shown), decision history replaced with an actual evaluation-record artifact (pros / cons / conflicts / adoption plan). Local gallery at .share/gallery.html.

Visual identity refresh. No engine or behaviour changes — purely the README, brand imagery, and image manifest.

Added

• Clair Obscur brand direction. Two new oil-painted Belle Époque brand images named after the painterly aesthetic of Clair Obscur · Expedition 33 — appropriate for a project literally called "Curator." docs/images/brand-curator-portrait.png (a bearded curator in a dark coat and burgundy vest examining a small gilded frame, surrounded by floating gilded miniatures) replaces the editorial-cover hero in the README. docs/images/brand-gallery-hallway.png (a long museum hallway with salon-style hangings and a single figure at the far end) ships as a secondary banner option.
• Real product screenshots ship to the repo. Five styled-HTML terminal/IDE captures (Edge headless): screenshot-claude-session.png (the agent activating Skills Curator in a Claude Code chat), screenshot-customize.png (real --customize per-section plan output), screenshot-recommend.png, screenshot-symptoms.png, screenshot-check.png (split-panel: self-clean scan vs flagged community skill).
• New modern technical diagrams. customize-workflow.png is now a sleek isometric block diagram on midnight navy (replacing the hand-drawn version). All ship under docs/images/.

Changed

• README hero is now the Clair Obscur curator portrait — stronger first impression than the editorial cover, ties the visual identity to the project's naming.
• --customize section in README now embeds the real terminal screenshot of the per-section decomposition output, replacing the hand-drawn workflow diagram.
• Demo section in README opens with the Claude Code session screenshot — proves what the tool does in one image.

Notes

• Original hero.png editorial cover remains in the manifest as an alternate for users who prefer the cleaner brand option.
• deploy.py FILES_TO_PUSH expanded with the eight new image paths.

Quality + dogfooding pass. The biggest fix is that running --check on Skills Curator itself no longer returns "DO NOT INSTALL" — the scanner was matching its own pattern definitions. Plus a CSO-compliant frontmatter rewrite, SKILL.md trim, and a small additive scanner:ignore mechanism documentation can use to opt out of self-scans.

Fixed

• Scanner no longer flags itself. cmd_check against the Python tier returned 6 findings (3 CRITICAL + 2 HIGH + 1 MEDIUM); against Lite, 1 MEDIUM. Both were false positives caused by descriptions like "eval() — arbitrary code execution risk" echoing the literal trigger they detect, and by documentation tables that legitimately list patterns. Both tiers now return "✅ No risks detected" on self-scan.
• SKILL.md frontmatter description rewritten to remove the workflow-summary CSO trap (was: "Maintains a trust-rated catalog... recommends with pros/cons... persists every decision..."). The summary form caused agents to follow the description instead of reading the body. Now strict "Use when..." triggering conditions only. Pressure-tested via subagent: 5/5 true positives, 3/3 true negatives across 8 user prompts.
• deploy.py commit message hardcoded as "Release v4.0.0 — Skills Curator" since v4.0.0; every push since shipped with a stale title. Now reads the current VERSION from registry.py and pulls the matching one-paragraph summary from CHANGELOG.md. Tag-suggestion next-step also uses the dynamic version.
• registry.py docstring said "Skills Curator v4.0.0" while VERSION = "4.4.0". Replaced with "see VERSION constant" so it can't drift again.
• README + CLAUDE.md doc freshness — test count was 26/35 in different places, engine was "1944 LOC". Now consistent: 37 tests, ~2.3k LOC.

Added

• scanner:ignore line/block markers. Lines containing scanner:ignore (anywhere in the line) are skipped during --check. Blocks bounded by scanner:ignore-block-start / scanner:ignore-block-end (works as <!-- ... --> in markdown) are skipped wholesale. Used internally to mark SECURITY_RISK_PATTERNS and the pattern-listing tables in references/commands.md and skills-curator-lite/SKILL.md. Available to skill authors who legitimately need to reference patterns in docs.
• 2 new pytest cases (test_scanner_ignore_block_suppresses_findings, test_scanner_ignore_single_line) → total 37 (was 35).

Changed

• SKILL.md trimmed from 2073 → 1800 words. Removed command-list duplication that already lived in references/commands.md, removed "Why no install counts" from SKILL.md (lives in CLAUDE.md as maintainer-facing rationale), tightened --customize and Platform Management sections, replaced the duplicated "When to activate" block with a one-line non-redundant note. Added "Common Mistakes" table.
• Pattern descriptions in SECURITY_RISK_PATTERNS rewritten to use CWE references rather than echoing the literal trigger ("Dynamic code execution call (CWE-95)" instead of "eval() — arbitrary code execution risk"). Detection behavior unchanged.

Notes

• Local-install copies in ~/.claude/skills/skills-curator/ were synced with the source during this work; future installs from npm/skills.sh will pick up the changes automatically.

The Lite-promotion release. skills-curator-lite is now feature-parity with the v4.3 Python engine and ships as the default tier. The Python full version remains available as the performance/regression-tested tier.

Added (skills-curator-lite v2.0.0)

• Feature parity with the Python v4.3 engine. Lite gains everything that previously needed Python: --auto-equivalent project fingerprint, --customize-equivalent fork generation, GitHub topic-search catalog enrichment via curl, multi-target migration, and the 55-platform catalog.
• Five verbs (was three): RECOMMEND, EVALUATE, AUDIT, PLATFORMS, MIGRATE — plus the bonus CUSTOMIZE flow.
• Embedded catalog expanded to 19 entries with hand-written pros/cons matching the Python KNOWN_SKILLS.
• Tag-rivals table + customization hints — agent surfaces a customize hint when a skill's tags imply a stack mismatch with the project.
• First-activation orientation — same two-line intro as the Python version, adapted for Lite.
• Live catalog enrichment via curl to GitHub Search API. Honors SKILLS_NO_TELEMETRY=1.
• Platforms catalog (all 55 supported agents) embedded as YAML.
• Symptom→skill mapping expanded from 11 → 17 patterns.

Changed

• Lite is now the default install tier. install.sh and install.ps1 install Lite unconditionally and add the Python full version on top only when Python 3.10+ is available (or --with-python/-Tier with-python is passed).
• Plugin description rewritten to lead with two tiers (Lite + Python full).
• Order in plugin.json skills array flipped — Lite is now first.

Notes

• Lite intentionally drops one capability: cross-device Gist sync. Users who need sync should add the full Python tier.
• The two skills don't conflict — different registry paths.

Install: npx skills add captkernel/Skills_Curator

Skills Curator v4.2.0 — Customize + Lite

Two big additions.

1. --customize — fork external skills for your project

Take any skill (registered, local, or owner/repo@skill), score each section by relevance to your project's stack, and emit a customized fork at ~/.claude/skills/<name>-for-<project>/SKILL.md. Per-section actions: keep, keep-trim, rewrite-stack, drop-or-rewrite, rewrite-frontmatter. The engine produces the structured plan; the agent rewrites Vue examples to React, Django to FastAPI, etc.

python registry.py --customize agent-browser
# Or preview only:
python registry.py --customize agent-browser --no-fork

2. Skills Curator Lite — same model, no Python

A new companion skill at skills/skills-curator-lite/. Pure Markdown — the agent reads project files, applies an embedded catalog + symptom map + 14 grep-based security patterns, and writes the registry directly via Bash. Use when Python isn't available.

Both skills now ship in the plugin manifest. Install one or both — registries don't conflict.

Install

npx skills add captkernel/Skills_Curator

See CHANGELOG.md for the full diff vs v4.1.0.

Skills Curator v4.1.0 — Intelligence layer

The proactive turn. Skills Curator now activates without being asked.

New

• --auto — fingerprints the project (key dep/config/doc files + mtimes) and only re-runs the recommendation engine when the project actually changes. Cheap to call on every session start.
• --symptoms "<phrase>" — maps user complaints ("slow tests", "ugly UI", "manual deploy", etc.) to skill categories. 17 built-in patterns covering testing, CI/CD, UI, deploy, docs, performance, auth, refactor, accessibility.
• --find <term> — alias for --discover, matches the verb users learn from npx skills find.
• New SKILL.md "Proactive activation" section instructs the agent to run --auto at session start and --symptoms whenever the user describes a pain point.
• New references/discovery.md: Skill Categories table (10 domains), single-skill response template, search tips.

The intelligence layer is the USP

v4.0 was "evaluate when asked". v4.1 is "learn the project, listen for symptoms, recommend without prompting." See the README for the full pitch.

Install

npx skills add captkernel/Skills_Curator

See CHANGELOG.md for the full diff vs v4.0.0.

Skills Curator v4.0.0 — Initial public release

Decide once. Re-decide never. The first Claude Code skill that evaluates skills before you install them and persists your verdict so you never re-decide.

What is this?

A Claude Code skill that handles judgment about other skills:

• Pre-install evaluation (pros / cons / conflicts / verdict, all persisted)
• Project-aware recommendations ranked by fit, not popularity
• Pre-install security scan (14 risk patterns)
• Stack audit, health scoring, version drift detection
• PR-ready markdown export of every decision
• Cross-agent migration (9 agents, verified paths)

Install

npx skills add captkernel/Skills_Curator

What ships in v4.0.0

• 26 pytest cases (Linux + macOS + Windows ×ばつ Python 3.10 / 3.11 / 3.12 / 3.13)
• 3 slash commands: /skill-evaluate, /skill-recommend, /skill-audit
• Stdlib-only Python — no pip install step
• Cross-device sync via private GitHub Gist
• SKILLS_NO_TELEMETRY=1 for full offline mode

See CHANGELOG.md for the complete list of changes, fixes, and removals.