Extract

From the conversation/context/user:

• The ask, confirmed — intent's work order, or as good: the WHAT the user has already settled.
• Its open choices — each place more than one way to build it remains, where the ways differ in what gets built.

Gate

Proceed only when: the ask is confirmed, and more than one way to build it remains that differs where it matters.

Anything else, say which in one plain line — never decide a path anyway:

• The ask itself is still rough — starting would mean inventing what the user means → use intent skill.
• Only one real way exists → nothing to decide; state it and proceed with the work.

Route

Before routing any choice: retrieve what settles it without the user — their taste from CLAUDE.md and project rules, facts from the repo, docs, the web. Only choices that stay genuinely open after retrieval route to a skill.

Route each open choice by what keeps it open, invoking the named skill with the Skill tool:

• A claim about something that exists — a library fits, the schema holds, the current design bears the load → use judge skill.
• The user can't choose from nothing — nothing exists to react to → use draft skill: the fewest versions that show a real fork, differing where it matters (who owns the data, where the seam sits, what fails how), never one approach renamed.
• The user holds taste they can't yet state → use elicit skill.
• Their stated preference reads two ways that would build different things → use clarify skill.
• They're missing something already settled about the technology → use explain skill.

When the work iterates — an agent loop, a refinement cycle — which loop is itself one of the choices: read loop-selection.md.

Each decision re-enters the routing: a resolved choice can open the next one or conflict with an earlier one; a conflict is one more choice, surfaced to the user, never silently settled. Route again until every open choice has the user's decision or confirmation and none conflict.

Output

The technical path is the result, decided: each choice resolved with its reason and what was turned down. The user owns the path.

Then check what still stands open:

• Nothing, and a human will watch each step of the work → proceed with the work.
• The work will run without a human watching, optimizing toward something → use target skill before it runs.

The shaping is spent once the user owns the path — anything further runs under the branch picked here, never as another round of deciding the path.