glm-hammer

Color prospector (crucible prospecting, one of 4 dimension prospectors). Dispatched by the crucible skill in parallel with the type/layout/motion prospectors, after the vein-reader returns. Receives the Direction Brief plus any user-provided references and hunts real-world palettes; returns ≥3 references with concrete hex values. Read-only plus web research.

Plan critic (forge panel, seat 3 of 3). Dispatched by the forge skill to verify a plan fully covers the user's request, contains no placeholders, and states only binary-decidable acceptance criteria with adequate verification. Read-only; returns APPROVE or REJECT with evidence.

Plan critic (forge panel, seat 1 of 3). Dispatched by the forge skill to judge whether a plan's tasks can actually be implemented as written against the real codebase. Read-only; returns APPROVE or REJECT with evidence.

Design assay judge (crucible). Dispatched by the crucible skill between smelting and the designer panel to verify that every token in tokens.json traces back to a real prospected reference or the Direction Brief — no invented values, no invented references. Read-only; returns APPROVE or REJECT with evidence.

Design critic (crucible design panel, seat 1 of 2). A picky art-director dispatched by the crucible skill to judge whether the tokens actually sing the story — palette family, typographic voice, spatial rhythm, motion character. Read-only; returns APPROVE or REJECT with evidence.

Light planning mode. Use when the user requests a small, clearly-scoped change (roughly ≤3 files, no new interfaces, no ambiguity) that still deserves a written plan before code. For ambiguous or multi-component work, use the forge skill instead. Produces an executable plan with binary acceptance criteria and a self-review, then hands off to hammer.

Storyline-driven design mode. Use when the user makes a storyline- or reference-driven design request — design tokens, mood/visual identity, a style guide — especially when they say "디자인", "무드", "스토리", "레퍼런스", "design tokens", "style guide". Prospects real references with a multi-agent panel, smelts them into W3C design tokens, assays fidelity to the story, then requires approval from a 2-designer critic panel before handing off to forge. Hooks enforce the gates.

Strong planning mode. Use when the user requests a non-trivial feature, refactor, or system change — especially when the request is ambiguous, spans multiple files, or says "제대로", "강하게", "빡세게", "thoroughly", "strong plan". Runs deep codebase recon BEFORE asking any question, drafts an executable plan, then requires unanimous approval from a panel of 3+ critic subagents before the plan may be presented. Hooks enforce the critic gate.

Implementation loop. Use when a plan document exists (from forge or blueprint) and implementation should run — including when the user says "implement", "execute the plan", "구현해", "실행해", or approves a plan. Executes tasks through a worker → validator → implementation-critic loop, then a security/QA review gate; any failure loops back to rework. Hooks block turn completion until every task and review passes.

Presentation deck mode. Use when the user wants a presentation, slide deck, or PPTX built from a topic or script — especially when they say "발표", "슬라이드", "프레젠테이션", "발표 자료 만들어줘", "슬라이드 뽑아줘", "presentation", "slide deck", "pptx", "deck". Secures a script, chains into crucible for design tokens (resume protocol), then builds a token-styled self-contained HTML deck. Hooks enforce the gates.