Skip to content

v0.5.2 -- Lean pipeline, rich output

leanbrand-syncquality-gatestastecontext-management

Lean Prompts

AI-driven prompt audit tool (P1--P7 automated checks + semantic analysis) identified dead weight, contradictions, and over-specification across 27 skills, 14 agents, and 12 system prompts. Result: 848 lines cut. Prompts slimmed to role-only headers -- agents own their methodology instead of repeating it from skills.

Brand Feedback Loops

Two new mechanisms keep brands in sync with shipped projects:

  • Build-time feedback -- at three pipeline checkpoints (foundation review, extraction scan, revision mode), the build detects when changes affect brand-level values and asks whether to propagate back. A background agent syncs tokens.json and brand chunks while the build continues.
  • Standalone /gsp:brand-sync -- run anytime to diff a project's shipped state against its source brand across four dimensions: tokens, voice and tone, visual patterns, and personality. Shows before/after divergences with file-level evidence. User picks what to sync back.

Build Quality Gates

Three new checkpoints in the build pipeline:

  • Component extraction -- post-build scan for duplicated Tailwind clusters, hardcoded values, and repeated JSX patterns. User-approved, not auto-applied.
  • Progressive tokens -- foundations writes global tokens only; screen agents add screen-specific tokens as needed.
  • Preview verification -- opt-in check after foundations compile (catches blank pages, broken tokens, missing fonts).

Visual Taste

The critic now produces dual scores: usability (/50, Nielsen's heuristics) + taste (/75, 15-item checklist covering typography tension, color restraint, spatial rhythm, surface depth, and more). Three new references give agents a concrete visual vocabulary: anti-patterns (AI convergence counters), advanced visual effects (spotlight borders, magnetic buttons, liquid glass), and advanced block compositions (bento 2.0, curtain reveals, sticky sequences).

Context Management

SessionStart hook re-injects active brand/project state after context compaction. Per-chunk and per-phase output budgets keep agent output focused. Context recovery script handles the handoff automatically.

Installer Fix

Fixed a bug where copyAgents() wiped its own work when called twice for main + custom agents. Also redesigned the installer banner -- left-aligned box layout.