Changelog

v3.0.0 (unreleased — Phase 1 hardening)

• Deterministic repo validator (scripts/validate-repo.py) with check-sync.sh as thin wrapper
• Static security scan (scripts/audit-security.py) + validate.yml / security.yml CI
• Example-leakage fixes: $setup, $git, $layout, $tokens worked examples are evidence-bound
• ## Gotchas sections across all 18 skills (bodies stay lean: max 806 words outside $craft)
• evals/triggers.json routing data for all 18 skills (10 cases each; runner deferred)
• docs/SKILL-AUTHORING.md canonical model
• allow_implicit_invocation explicit in all 18 agents/openai.yaml
• Trigger runner + confusion matrix (needs model runs — deferred)
• Execution fixtures + graders for $craft / $slice / $context (deferred)
• Baseline benchmark vs v2.1.0 (deferred)

v2.1.0 — Skill depth and Codex plugin

• Restored worked examples, output templates, and use-instead routing guardrails across all 18 skills, with trigger phrases in every description. Bodies stay lean (400–700 words); variant detail remains in references.
• Added .codex-plugin/plugin.json so the pack installs natively in Codex (codex plugin marketplace add mattjmdesign/dengskills, codex plugin add dengskills). Version bumps and sync checks now cover all four manifests.

v2.0.0 — Design workflows in code

• Shortened skill names; migration.json maps every previous name to its replacement. This is a breaking naming change.
• Revised all skills around client intent, existing work, and task-sized outputs.
• Expanded prototype to build and iterate coded prototypes. Added slice for complete product paths and craft for interface quality.
• Integrated craft for interface quality as the canonical copy; the former standalone craft-interfaces repo is folded in and its sync script and upstream manifest are removed.
• Updated prompts, behavioral evaluation cases, manifests, and metadata checks. Use scripts/check-sync.sh --metadata-only for validation without launching client CLIs.

v1.3.0 — Claude Code packaging + release tooling

Packaging only. No SKILL.md changed, and no client's view of the skills changed except Claude Code's, which previously had none.

Claude Code support

• Added .claude-plugin/plugin.json. Claude Code reads its plugin manifest from that path exclusively and ignores a root plugin.json, so the Agent Plugins manifest alone left the repo invisible to Claude Code even though every SKILL.md was already compatible. Verified: claude plugin details dengskills reports all 16 skills.
• Added .claude-plugin/marketplace.json listing this repo as a one-plugin marketplace with source: "./", so claude plugin marketplace add mattjmdesign/dengskills + claude plugin install dengskills@dengskills works without a separate marketplace repository.
• The Claude Code manifest omits the Agent Plugins $schema and carries only fields Claude Code reads. Both manifests hold the same metadata; check-sync.sh fails if they disagree.

Packaging model

• README now states the rule the repo follows: skills/ is the portable core, every root file is packaging for one client, and supporting another client means adding a manifest beside the others — never forking a skill. The packaging table lists Agent Skills, Agent Plugins, skills.sh, and Claude Code as peer layers.

Release tooling

• Added scripts/bump-version.sh: sets the version across every manifest at once, accepts major / minor / patch or an explicit version, and prints the remaining release steps. Version duplication across client manifests is now a one-command edit rather than three hand-edits that can drift.
• scripts/check-sync.sh gained manifest checks: the manifests must agree on version, homepage, repository, and license; the marketplace entry must point at the repo root; and both Claude Code manifests are run through claude plugin validate --strict when the CLI is present.
• README gained a "Releasing a change" section with a per-client refresh table. Documents the non-obvious part: most clients cache a version-keyed copy rather than reading the repo live, so claude plugin update picks up nothing — new skills included — unless the version changed. Verified both ways: adding a skill without a bump was ignored, with a bump it appeared.
• Ignored eval output, local agent-tool state, and node_modules/.

v1.2.0 — Production readiness + revisit

• Added $production-readiness-review (QA & Delivery). Scores a running slice as demo / pilot / production and lists UI-owned risks.
• Skills now explicitly revisit existing work: $product-intent-clarifier and $gap-state-inventory include after-preview guidance.
• Pack scope is no longer “pre-code only.” Testing, deploy, and observability remain guide-only.

v1.1.0 — Guide pairing, trigger discipline, worked examples

The pack is now explicitly paired with the Design Engineering guide: every skill names its corresponding guide page, and the guide links back to its skills.

Trigger discipline

• Added "Do not use when / use instead" guardrails to skills with overlapping triggers (product-intent-clarifier, prototype-fidelity-selector, requirements-from-brief, content-hierarchy-planner, ui-layout-architect, ui-system-initializer).
• Added a routing table to README.md: "If you need X, use Y before/after Z".
• sitemap-planner now covers filters-in-search-params and the route-vs-state distinction (grounds the existing filters-url eval).

Worked examples

• Every SKILL.md now includes one filled example output following its own output format, so agents anchor on a concrete exemplar instead of an empty template.

Evals

• Added negative evals to key skills (e.g. "does not recommend code prototype", "does not invent commands", "does not replace an existing token system").
• Existing evals unchanged and still passing in intent; filters-url is now grounded in the skill body.

Scope

• README and docs now state the pack covers Product Definition through Build Readiness. QA/Delivery phases remain guide-only; post-build skills (testing strategy, pre-launch readiness) are on the roadmap.

Agent Plugins packaging

• Added a root plugin.json manifest (Agent Plugins v1.0.0). The repo is now dual-packaged: installable via npx skills add mattjmdesign/dengskills and loadable directly as an Agent Plugin by any compatible client. skills/ is the standard fixed location and every SKILL.md conforms to the Agent Skills specification; skills.sh-specific files (skills.sh.json, agents/openai.yaml, evals/evals.json) live alongside the portable core.
• Reviewed against the published Agent Plugins 1.0.0 spec: closed manifest fields only, canonical $schema, name constraints, and skills discovered as immediate children of skills/. homepage and all "See also" links now use https://frontendguide.dev.

Sync tooling

• Added scripts/check-sync.sh: validates that frontmatter names/descriptions, README tables, skills.sh.json groupings, and each skill's evals all reference the same set of 15 skill names.

v1.0.0 — Initial pack

15 skills across Product Definition, Project Setup, UI Composition, IA & Flows, and Build Readiness. Published on skills.sh and GitHub.