Ask ZSL

A thin interactive router. You're mid-session, you know roughly what you want to do, but you're not sure which ZSL skill fits — ask-zsl asks you a couple of situational questions and points you at the right one. It's routing logic only: it doesn't re-describe each skill (the Skills overview and each SKILL.md already do that) — it just gets you to the right door.

User-invoked only (disable-model-invocation: true): you ask it; it never fires on its own.

How to route

Ask the user a short, situational question, branch on the answer, and name the skill (and why it fits) — then stop. Don't run the skill for them; hand them the slash command. If they're already deep in one phase, jump straight to that branch.

"What are you trying to do right now?"

• Mature a fuzzy, multi-session idea before it's a PRD → /decision-mapping (decision map under .scratch/decision-maps/, then hands to /to-prd).
• Stress-test a concrete plan / align on what to build → /grill-me (non-code) or /grill-with-docs (sharpens CONTEXT.md + ADRs inline).
• Package the current conversation into a PRD → /to-prd.
• Break a PRD into independently-grabbable slices → /to-issues, then /triage each child to ready-for-agent.
• Build it → see the build sub-questions below.
• Hunt a bug or perf regression → /diagnose.
• Understand an unfamiliar repo before touching it → /teach-me-the-codebase.
• Fight architectural entropy → /improve-codebase-architecture.
• Ship / review / commit → /code-review, then /commit or /commit-push-pr (and /git-branch first if you're still on the default branch).

Build: one slice or many?

• One slice, here, now → /tdd (single-issue red-green-refactor on your branch).
• Several unblocked [AFK] slices of a PRD, integrated into one PR → /tdd-parallel (fans out, integrates, auto-chains coverage, opens one PR).
• Manual [HITL] steps an agent can't do → /human-itl first — /tdd-parallel refuses while any [HITL] is open.
• /tdd vs /tdd-parallel? One issue at a time on the current checkout → /tdd. A whole PRD's worth of parallel slices into a single integration PR → /tdd-parallel.

Verify and prove coverage

• Is every PRD user story actually covered by a passing, non-vacuous test? → /verify-coverage (Tier A maps to an existing test; Tier B generates one and mutation-proves it). Usually auto-chained by /tdd-parallel; invoke it directly to audit a PRD whose slices shipped elsewhere.

Run it overnight (remote agents)

• Schedule unattended overnight runs of several PRDs → /afk-fanout in the evening (one remote routine per PRD, 2h apart).
• What each scheduled routine fires → /afk-worker (per-PRD remote executor; you don't invoke it by hand).
• Walk the overnight results in the morning → /morning-review (reconciles the afk-runs ledger, surfaces PRs to verify/merge and halts to re-triage).

Off the loop

• Throwaway exploration of a design → /prototype.
• Compact this session for the next agent → /handoff.
• Standup notes from recent sessions → /timesheet.
• Author or sharpen a skill → /writing-great-skills.

Keep routing in sync

ask-zsl and the "Which skill do I want?" decision tree in docs/skills/index.md are two views of the same routing. When a loop skill is added or removed, both must change together — this pairing is recorded in CLAUDE.md's skill-sync contract. If you find yourself routing to a skill that isn't in the tree (or vice versa), one of them is stale.

What this skill is not

• Not a doc. It defers every skill's description to that skill's page — it only owns the routing.
• Not an executor. It points you at a slash command; it never runs the skill on your behalf.