config-chunks

First-run guided onboarding that sets up an agent for a less-technical user — leads with a one-shot recommended set of guidance "chunks" (or 2–3 plain-language questions to tailor), confirms, applies them by driving the config-chunks engine into CLAUDE.md / AGENTS.md, then offers safe permission setup. Use when the user says "/ai-setup", "set up my agent", "set up my agent guidance", "onboard me to chunks", "configure my CLAUDE.md guidance", "first-run setup", "recommend chunks for me", "help me get started with config-chunks", "which chunks should I use", or otherwise wants a guided walkthrough to subscribe to guidance chunks. Re-runnable any time; for piecemeal management afterward use /chunks, for permissions use /permission-setup, and to author new chunks use /chunk-review.

Score a candidate guidance chunk for inclusion in the always-on agent-instruction bundle (the concatenated context injected into every agent session). Use when reviewing or authoring a chunk before opening a chunk PR, or when asked to "/chunk-review", "score this chunk", "review a guidance chunk", "review this chunk", "is this chunk worth shipping", "should this go in the bundle", or "is this chunk-worthy". Applies a 7-axis rubric (Universality, Altitude, Not-a-skill, Non-redundancy, Concision, Impact, Stability), enforces hard size gates, checks for redundancy against existing chunks, and emits a score plus a keep / revise / cut verdict.

Manage guidance chunks — toggle subscription to chunk groups and standalone chunks in `~/.claude/config/chunks.yaml`, set which host files receive the bundle, or run a doctor check on the config. Use when the user says "/chunks", "list my chunks", "what chunks am I on", "add a chunk group", "opt into a chunk", "subscribe me to <group/chunk>", "unsubscribe from <group/chunk>", "set chunk targets", "diagnose my chunks config", "fix my chunks.yaml", or otherwise wants to change which guidance chunks land in their CLAUDE.md / AGENTS.md bundle. The skill drives the engine script — it never hand-edits YAML.

Walk a less-technical user from a vague want to a well-framed, systems-level problem — separating the stated ask from the real need, eliciting success criteria and constraints, reflecting the framing back, then decomposing it with them. Use when a session opens with a vague or non-technical ask ("I want X", "can you build me something that…", "I need a thing for…"), when the user has named a solution but not the problem, when "frame before you solve" is in play, or on "/ideation", "help me think this through", "I'm not sure what I actually need", "where do I start". Pairs with the engineering-mindset and problem-framing chunks; this skill is the procedure those posture-chunks point to.

Recommend safe, conservative agent permission settings for a less-technical user, then delegate the write — never mutate security config directly. Detects the provider (Claude Code / Codex / Cursor / others) and surface (CLI / Desktop-IDE / Cloud), explains permissions in plain language, recommends the "Cautious" least-privilege posture as behavior then concrete config, names that provider's bypass footgun by name, and shows the exact config + location for the user (or a platform config skill) to apply. Use when the user says "/permission-setup", "set up permissions", "set up my agent permissions", "is my agent safe", "make my agent safe", "what permissions should I use", "lock down my agent", "are my settings secure", "review my agent permissions", or finishes /ai-setup. Re-runnable any time; reads references/permission-profiles.md live.