ADHD

Spawn many heads thinking differently, in parallel, then pick. The first three answers a senior engineer would give in thirty seconds are correct and forgettable. The interesting answers live past number three. This skill walks there.

Pre-flight

Expensive: ≈8 peer spawns plus an orchestrator critic turn (5–10x a single answer). Gate every run.

If the user typed /adhd or explicitly asked for "ADHD mode", skip the gate and run.

Otherwise, all three must hold:

1. Open-ended. Multiple viable answers, not one canonical correct one.
2. High-stakes. Architecture, public API, naming a product, fuzzy bug with no known root cause, schema, migration. Not throwaway work.
3. Open phrasing. User did not say "quick", "standard", "canonical", "textbook", "just", "one-line", "show me how to".

If any fails: abort, answer directly, optionally offer /adhd <problem> as an explicit follow-up.

Phase 1 — Diverge

Pick 5 frames from the table below. For code-shaped problems: 4 code or design tags plus 1 wild. Otherwise mix.

For each frame, spawn_peer in the user's workspace with a description like adhd-diverge: <frame> and a seeded message containing the frame's vantage prompt, the problem, any user-provided context, and this instruction:

You are in DIVERGENT mode. Generator only, no critic. Generate 6 short distinct ideas under this frame. Each idea is one phrase or one sentence. Do not evaluate, rank, or hedge. The first three obvious answers everyone would give are banned — push past them into the awkward middle. Output a JSON array only, no prose: [{"text": "...", "rationale": "..."}, ...].

Spawn the 5 peers in parallel. Branches must not see each other. Do not pass one branch's output as context to another — anchoring eliminates the diverge.

Phase 2 — Focus

Two passes:

1. Score + cluster (inline, orchestrator does this). For each idea, rate novelty / viability / fit 0–10. Flag traps with a one-line mechanistic reason (not vague risk). Cluster the pool into 3–6 groups by underlying angle, not surface keywords ("remove-the-server plays", "cache-shaped plays").

2. Deepen top 3 (spawn one peer per survivor). Rank by 0.35*novelty + 0.40*viability + 0.25*fit, exclude traps, take top 3. For each, spawn_peer with a descriptive name like adhd-deepen: <cluster-or-idea-shorthand> and this seeded message:

You are in FOCUS mode. Take one promising idea and connect dots. Sketch how it would work in 4–8 sentences. Name the load-bearing risk. Name the first concrete step a coder would take. Then generate 3–5 sub-ideas (variations, hybrids, unlocks). JSON only.

Cleanup and resume

After all acks land, kill_peer all 8 (5 diverge + 3 deepen). The process is what's expensive; the session persists in storage and can be resumed later.

In the final reply, include each deepen survivor's session id or peer name so the user can resume the one they want to push on. Diverge sessions are also resumable but rarely revisited.

Pick descriptive names on spawn — the resume list shows descriptions, not internal ids.

Output shape

Render in this order. The structure is the point — do not collapse to prose.

1. Brief. One or two lines confirming the problem and any reframe.
2. Wide set. Full pool grouped by cluster, each cluster labeled by angle. Each idea is one phrase with score chips [N7 V8 F9].
3. Converge. 2–4 idea shortlist with one line each on why. Mark the non-obvious-but-viable pick with ★. List traps separately, each with the one-line reason it's a trap.
4. Focus. The 3 deepened branches: sketch, load-bearing risk, first concrete step, child ideas. Include the resumable session id for each.
5. Provocation. One wildcard question or idea that opens a new direction.

Frames

Pick 5 per run. Vary across sessions so the same problem produces different candidates when re-run.

Anti-patterns

• Convergence disguised as divergence. Ten variations of one idea is not breadth. If every candidate shares the same underlying assumption, you decorated, not diverged.
• Weird-for-weird's-sake with no convergence. A pile of 30 unsorted absurdities is as useless as one safe answer. Always converge with a real opinion.
• Skipping isolation. If you simulate parallel branches by writing them sequentially in one orchestrator turn, you have not done ADHD. Use real spawn_peer fan-out — each spawned peer is a fresh context.
• Refusing to commit. "Here are 20 ideas, you decide" is a cop-out. Generate wide, converge with a position.
• Forgetting to kill. Leaving 8 idle peers after every run pollutes the mesh. Kill on cleanup; sessions remain resumable.

Calibration

Scale to stakes. Quick "name this function": 3 frames × 4 ideas. "How should we position this product": 5 frames × 8 ideas. Default 5 × 6.

Stop diverging when new candidates start repeating the shape of existing ones. The space is mapped; do not pad to hit a number.