Compose

Extract

From the conversation/context/user:

• The ask and its aspects — the distinct domains it spans.
• The skills in play — the ones the user named, plus the ones whose listed conditions the ask meets.

Gate

Proceed only when: the ask has multiple aspects that different skills carry, or the user named multiple skills.

Anything else, say which in one plain line — never compose anyway:

• One skill's condition covers the whole ask → invoke that skill alone; its procedure governs.
• The ask's words admit two readings that would build different things → use clarify skill, then compose over what the pick settles.

Plan

• Use a subagent for exploratory reading, keeping the main thread's context for the composition itself.
• Choose the skills that matter to the ask, scaled to it — never two skills that deliver the exact same outcome.
• Define the stages the ask needs — within this session or across sessions. A single skill can fire more than once, at different stages.
• For each chosen skill, conclude its useful-when: the condition in this conversation that will call for it.

Invoke

Work the ask, invoking each skill as the conversation reaches its condition, in dependency order:

1. Invoke it via the Skill tool. As its body loads, conclude its wrong-if — the conditions under which it refuses.
2. Emit one line before the work: [compose] <skill> — <why, and the hoped-for outcome>.
3. After the output, silently check the skill's wrong-if against what came out. Speak only if it trips — one line, corrected within the same turn.

Amend

A wrong-if trip, or a turn where no chosen skill's condition fits, amends the one held belief that failed — re-derive the chosen set only if needed.

The loop ends when every stage the plan defined has had its skill run, or the user stops the composing.

Output

The worked ask is the result. Each invoked skill's output stands as that skill handed it back — composing never restates them in its own words, and never assembles a verdict none of them reached.

Name what ran: the skills invoked, in the order their conditions arrived. A chosen skill whose condition never came is named as never invoked, never dropped in silence. When the loop ended because the user stopped the composing, the stages that never ran are named as never run, never counted as unneeded.

Then check whether anything still unresolved would change what gets built:

• Nothing would → proceed with what the worked ask produced.
• The ask itself came open while composing → use intent skill.
• A claim about what a skill produced needs a committed verdict → use judge skill.

The composing is spent once the loop ends — anything further runs under the branch picked here, never as another round of composing.