Adhd

ADHD

Stop picking the textbook answer. The first three answers the model would give are the answers a senior engineer would give in thirty seconds. Correct. Forgettable. The interesting answers live past number three, in the awkward middle nobody walks into. This skill makes the model walk there.

When to use vs siblings

• /adhd — generate the option space when you don't have candidates yet.
• /oracle (compare mode) — evaluate options you already have.
• /plan-ceo-review — challenge whether the thing should be built at all.
• /verify — adversarially check a conclusion you've already reached.

They compose: /adhd to widen, /oracle to weigh the shortlist.

Pre-flight (run before Phase 1)

This skill is expensive. About 10 Agent calls, 30 to 90 seconds wall clock, 5 to 10x a single answer. Do not pay that cost when a direct answer is better. Run this gate before Phase 1.

Step 1. Explicit invocation check.

If the user typed /adhd or explicitly asked for ADHD mode, "use the adhd skill", or "run ADHD on this", SKIP the rest of this section and go straight to Phase 1. The user opted in. Do not second-guess.

Step 2. Self-judge (only if Step 1 did not match).

Ask yourself three questions. If the answer to any is no, ABORT.

1. Open-ended? Would a senior engineer give multiple viable answers here, or is there one canonical answer? If canonical, abort.
2. High-stakes? Is the cost of the obvious answer being wrong actually high? Architecture decisions, public API surfaces, naming a real product, fuzzy bugs with no known root cause, schema design = yes. Side project at 11pm = no.
3. Open phrasing? Did the user avoid words like "quick", "standard", "canonical", "textbook", "just", "one-line"? If they used any of those, they want the direct answer. Abort.

If all three checks pass, proceed to Phase 1.

If any fails, ABORT and answer the question directly. Optionally append one sentence: "If you want a wider exploration under parallel cognitive frames with explicit trap detection, run /adhd <your problem>."

Standalone Codex batching

Read the live concurrency limit before Phase 1 and reserve a slot for the main session. Launch only the independent generator frames that fit. Create each frame with spawn_agent and wait for it with wait_agent; then launch the remaining frames in the next batch. Use send_message to deliver context to a running frame, followup_task to trigger another turn for an idle existing frame, and interrupt_agent only to stop a frame's current turn.

Every frame still receives only the original problem, user context, its own vantage prompt, and the generator instruction. Never pass output between frames, even across batches. Apply the same batching to the three deepen agents. Do not require five simultaneous slots or promise Sonnet routing.

The loop

Two strict phases. Mixing them kills idea quality, because the critic strangles the generator.

Phase 1 — Diverge (no critic)

For the problem P:

1. Pick 5 cognitive frames from the table below. Bias toward engineering tags when the problem is code-shaped. Always include at least one wild frame to keep range.

2. In Claude, spawn 5 parallel Agent tool calls in ONE message. In standalone Codex, use the independent batches above. One agent per frame. Each Agent gets only:

   • the problem P
   • any context the user provided
   • the chosen frame's vantage prompt
   • a system instruction that forbids evaluation

   The exact instruction to give each Agent:

   You are in DIVERGENT mode. You are a generator, not a critic. Generate 6 short distinct ideas under this frame. Each idea is one phrase or one sentence. Do not evaluate. Do not rank. Do not 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 before or after. [{"text": "...", "rationale": "..."}, ...]

3. Critical invariant. Branches must be isolated. Claude runs all five in parallel; Codex may batch them only to respect its live slot limit. Do NOT pass one branch's output as context to another. Branches that see each other anchor each other and the whole method collapses to a wider single thought.

Phase 2 — Focus (critic on)

After all branches return:

1. Score. Rate each idea on three axes 0 to 10: novelty (distance from the obvious default), viability (could it actually ship), fit (does it address the stated problem). For any idea that looks attractive but is a trap (hidden cost, false economy, will not scale, premature abstraction), flag it with a one-line reason.

2. Cluster. Group ideas into 3 to 6 clusters by their underlying angle, not by surface keywords. Label clusters by angle: "remove the server plays", "cache-shaped plays", "batched-window plays", "race-multiple- backends plays".

3. Deepen the top 3. Rank by weighted score (novelty 0.35 + viability 0.40 + fit 0.25), exclude traps, take top 3. For each, spawn one Agent call that produces:

   • a 4 to 8 sentence sketch of how the idea works
   • the load-bearing risk
   • the first concrete step a builder would take
   • 3 to 5 child ideas (variations, hybrids, unlocks)

   Deepen Agent instruction:

   You are in FOCUS mode. Take one promising idea and connect dots. Sketch how it would actually work in 4 to 8 sentences. Name the load-bearing risk. Name the first concrete step a coder would take. Then generate 3 to 5 sub-ideas that branch off (variations, combinations with other domains, things this unlocks). Output JSON only.

Scoring, clustering, and the final synthesis stay in the main session — that is judgment work. In Claude, generator and deepen agents inherit CLAUDE_CODE_SUBAGENT_MODEL (Sonnet). Standalone Codex makes no model-routing promise beyond the live host's configuration.

Frames

Pick 5 per run.

Picking frames

For code-shaped problems: pick 4 frames tagged code or design, plus 1 tagged wild. For open product or strategy problems: a mix from all tags. Vary the picks across sessions so the same problem produces different candidate sets when re-run.

Output shape

After Phase 2, render in this order. Do not collapse it into a wall of prose. The structure is the point.

1. Brief. One or two lines confirming the problem and any reframe used.
2. Wide set. Full pool grouped by cluster. Each cluster labeled by underlying angle. Each idea is one short phrase. Show score chips like [N7 V8 F9] next to each.
3. Converge. A 2 to 4 idea shortlist. State why each is on the list. Mark the non-obvious-but-viable pick explicitly with ★. List traps separately, each with the one-line reason it is a trap.
4. Focus. The 3 deepened branches. For each: the sketch, the load- bearing risk, the first concrete step, and the child ideas.
5. Provocation. One wildcard question or idea that opens a new direction the user can push into if nothing landed.

Anti-patterns

These are how this skill goes wrong. Watch for them.

• Convergence disguised as divergence. Ten minor variations of one idea is not breadth. If every candidate shares the same underlying assumption, you have not diverged. You have decorated.
• Weird-for-weird's-sake with no convergence. A pile of 30 unsorted absurdities is as useless as one safe answer. Always converge.
• Walls of equally-weighted prose. Cluster, label, pull out the best. Structure is half the value.
• Refusing to commit. After diverging, take a position on what is actually promising. "Here are 20 ideas, you decide" is a cop-out. Generate wide, but converge with a real opinion.
• Skipping the isolation invariant. If you simulate parallel branches by writing them sequentially in one context, you have not done ADHD. You have done a wider single thought. The Agent tool gives each branch a fresh context. Use it.

Calibration

• How many ideas? Scale to stakes. Quick "name this function" = 3 frames × 4 ideas. "How should I position this product" = 5 frames × 8 ideas. Default is 5 × 6 = 30.
• How weird? Read the room. Serious strategy work: flag the wild cards clearly so they do not read as unserious. Open brainstorming or play: let it run loose. Absurd ideas earn their place by seeding viable ones.
• When to stop diverging? Stop when new candidates start repeating the shape of existing ones. The space is mapped. Do not pad to hit a number.

Cost

5 diverge + 1 score + 1 cluster + 3 deepen ≈ 10 Agent calls per run. About 5 to 10x a single-shot answer. Not for every keystroke. For decision points where the cost of the obvious answer is high. In Claude, diverge/deepen agents run on the Sonnet subagent pool, so the Opus/Fable cost of a run is one synthesis pass. Standalone Codex reports its own live routing and slot limits instead.

Attribution

Ported from UditAkhourii/adhd (MIT). Upstream ships the same loop as an npm CLI (adhd-agent) plus evals and a source spec on divergent ideation; this port keeps the skill-only form (no install) and adds the cc-settings sibling-skill routing and subagent model notes.