claude-tweaks
A structured workflow system for Claude Code — from idea capture through build, review, and wrap-up.
What this does
Claude Code is powerful but unstructured. claude-tweaks adds a complete development lifecycle: capture ideas, challenge assumptions, decompose into specs, build with quality gates, and learn from what was built. Every finding is explicitly resolved — nothing silently drops.
What's new in v5.0.0 — Recon v2: LLM-as-judge + scheduled Routine
Recon v2 replaces the v1 mechanical-lens spine with an LLM-as-judge model: the LLM evaluates the repo against a criteria catalog, calling deterministic tool checks as evidence. Area-type routing, content-hash skip, hotspot priority, fingerprinting, and dedup are handled by deterministic helpers. The v1 subagent dance and plan-judgment / ingest-judgment phases are removed. Recon now runs as a scheduled Routine for continuous, hands-off coverage — no manual invocation needed.
What's new in v4.7 — Deep web research + Diagram Design companion
/claude-tweaks:research adds citation-audited deep web research to the plugin. Four runtime modes trade depth for time:
- quick (~2-5 min, 5+ sources) — fast scan
- standard (~5-10 min, 10+ sources) — balanced default
- deep (~10-20 min, 15+ sources) — comprehensive synthesis with broader source pool
- ultradeep (~20-45 min) — multi-persona red-team with adversarial review
As of v4.15.0 this delegates to Claude Code's built-in /deep-research Dynamic Workflow when available, with a lean inline fallback otherwise. Reports land under .claude-tweaks/research/.
Diagram Design companion plugin — a soft-hook integration with cathrynlavery/diagram-design (MIT, separately installed). Unlike Impeccable (wrapped via /claude-tweaks:design) or research (which delegates to the built-in /deep-research), diagram-design has no callable surface — it's a pure-skill plugin that auto-triggers from its YAML description. claude-tweaks adds contextual nudges at three lifecycle moments:
/specify Step 2.5d (new, all surfaces) — when the design doc describes state machines, schemas, multi-actor flows, decision trees, or layered architecture, the spec summary surfaces "consider a {type} diagram" with a suggested output path (docs/diagrams/{slug}.html). Un-gated from frontend — backend specs get architecture / ER / state diagrams too. Caps at 2 suggestions per spec./journeys Step 3.6 (new) — when a journey crosses 2+ personas, has 3+ named decision branches, or sequences 2+ external services, suggests the matching diagram type (swimlane / flowchart / sequence) before commit./review Lens 3i-diagram (extension) — when the diff added structural complexity but docs/diagrams/ has no matching file, emits one informational Lens 3i finding ("Visual documentation gap"). Mirrors the existing "doc-update missed" pattern.
All three hooks are gated by diagram-integration: enabled in CLAUDE.md, written by /init Phase 0.95 (always offered — not frontend-gated). Disabled / missing flag = silent no-op everywhere. claude-tweaks never invokes the plugin directly; the user accepts conversationally and diagram-design's skill auto-triggers. Shared procedure lives at skills/_shared/diagram-integration-check.md — flag-read, signal→type mapping (10 types), canonical phrasing, output convention.
See CHANGELOG.md for earlier release notes (v4.6, v4.5, v4.2, v4.1) and v3→v4 upgrade guidance.
Installation
/plugin marketplace add thomasholknielsen/claude-tweaks-marketplace
/plugin install claude-tweaks@claude-tweaks-marketplace
/plugin install superpowers@claude-plugins-official
/claude-tweaks:init
How it works
SKILL ARTIFACT SUPERPOWERS USED
───── ──────── ────────────────
