planning-with-files
Optional persistent file-based planning for multi-phase or long-running work that benefits from recovery across sessions or…
Explain a subsystem through direct codebase exploration. Use for how-it-works questions, walkthroughs, caller discovery, placement, or layering decisions.
$ npx -y skills add khaneliman/khanelinix --skill how --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/howContext preview
The summary Claude sees to decide when to auto-load this skill.
Explain a subsystem through direct codebase exploration. Use for how-it-works questions, walkthroughs, caller discovery, placement, or layering decisions.
name: how description: "Explain a subsystem through direct codebase exploration. Use for how-it-works questions, walkthroughs, caller discovery, placement, or layering decisions."
Explain actual behavior from code and caller evidence. Give the reader a working mental model, not annotated source. When called inside another lifecycle, return the traced evidence to that owner without starting another workflow.
Identify the subsystem, feature flow, runtime trace, or placement question. Resolve facts from code and tools. State a reasonable scope assumption when the request permits it; ask only when a material choice cannot be resolved safely. Reuse current evidence rather than repeating an exploration already completed.
when one bounded lookup would materially reduce context or latency.
Keep tightly coupled reasoning and synthesis in the parent. Do not create overlapping searches or a separate synthesis worker merely to satisfy phases.
Use [explorer-prompt.md](references/explorer-prompt.md) for delegated discovery. Each packet names its paths, question, constraints, required evidence, and exit criteria. Give only the relevant facts, not the conversation history.
Inspect returned paths and verify claims that determine the explanation. Reconcile contradictions and distinguish observed behavior from inference. The parent owns the conclusion and may substantially correct or rewrite worker output. Never forward an unverified worker explanation as the final product.
Use [explainer-prompt.md](references/explainer-prompt.md) for explanation shape. Scale the output to the question; a small helper does not need a subsystem report.
For architectural issues or improvement requests, establish the current model before critique. Reuse the explanation above or existing verified grounding. Use [critique-rubric.md](references/critique-rubric.md) to frame one fresh reviewer; [critic-prompt.md](references/critic-prompt.md) supplies its evidence contract. Escalate to `interrogate` for contested, high-risk, or explicit multi-angle review, not every architectural question.
The parent classifies findings as act on, consider, noted, or dismissed, with reasons. Present the explanation and any requested critique verdict. Return to the caller when this skill is only a grounding method.
Run packages directly with: Or install from the packages output. For example:
Optional persistent file-based planning for multi-phase or long-running work that benefits from recovery across sessions or…
Audit coding-agent sessions and mine cross-harness history for recurring corrections, skill-routing gaps, and harness faults. Use…
Design and audit agent instructions, skills, plugins, subagents, prompts, hooks, and provider configuration. Use for AI-tool…
Sketch interfaces, state, and module structure before code. Use for explicit design-led implementation or as a caller-owned…
Spawn N parallel candidates at the same task, pick a base, graft the strongest parts of the losers into it. Use for /arena,…
Bevy workflows for ECS, migrations, first-party UI/input/rendering/assets/shaders, plugin selection, scenes, scheduling, hot…