crumbs

Interview user, then set up agent-friendly repository guidance and tooling: sane .gitignore, repo-specific AGENTS.md, and optional mise.toml with runtime and check tasks. Use when user wants to start a new repo, prepare repo automation, create AGENTS.md guidance, or set up repo tooling from requirements.

Analyze Auto Guardian bash audit logs and propose safe deterministic allow/block rule improvements. Use when reducing guardian LLM calls or reviewing command approval patterns.

Severe git-status-driven review of current changes for maintainability, architecture, and AI-slop risk. Use for ruthless review, nuclear review, pre-merge quality gate, or when local changes may deserve broader refactoring instead of small patch acceptance.

Design UIs in Pencil (.pen files) and generate production code from them. Use when working with .pen files, designing screens or components in Pencil, or generating code from Pencil designs. Triggers on tasks involving Pencil, .pen files, design-to-code workflows, or UI design with the Pencil MCP tools.

Compress any skill to minimal tokens — rewrites SKILL.md, all references/, and scripts/ comments. Strips verbose prose, filler, redundancy while preserving technical accuracy and functionality. Use when user says "caveman-ify this skill", "compress skill", "shrink skill", or invokes /caveman-ify.'

Create standalone HTML artifacts for plans, prototypes, explainers, code reviews, reports, and custom editing interfaces. Use when the user asks for an HTML file/artifact, visual plan/spec, interactive prototype, explainer page, PR/code review artifact, report, or throwaway editor.

Execute verified refactor work units from refactor-opportunities. Builds dependency graph, dispatches parallel subagents per work unit, validates results.

Refactor and code review across hygiene, over-engineering, runtime lenses. Supports target review and uncommitted-changes review. Reconciles conflicts, checks churn, emits ordered work units, uses language overlays.

Interview user about a plan or design until decisions satisfy stated goal. Use when user wants to stress-test a plan, explore real options, get grilled on design, or mentions "grill me".

Apply user's prose style preferences for drafting, editing, and reviewing writing: casual direct voice, skeptical practical advice, no AI-slop patterns, and stricter self-critique for user-facing content. Use when writing or revising prose, docs, articles, emails, presentations, posts, or long-form explanations.