adhd-friendly

Inherits ~/.claude/rules/nile_patterns.md § Skills Inherit These — ceremony ∝ stakes (this skill's full pipeline is its maximum, not its floor) · one recommendation, not a manufactured slate · the procedure is not the deliverable · signal codes outrank this skill's verbosity · dedup before proposing anything new.

Why this exists, and what it is NOT

A single context is a bad idea-generator for one specific reason: it commits early and then argues for the commitment. Three "alternatives" produced in one thread are three samples from the same neighbourhood.

This skill is already half-owned by three installed skills. Read this before assuming it is new.

⇒ The only two things adhd-friendly adds: genuine context independence and generation rather than critique. If a run does not need both, one of the three above is the cheaper right answer — say so.

Relationship to brainstorming: brainstorming claims "before any creative work" and forbids invoking other skills. It is a one-question-at-a-time requirements interview that narrows toward a spec. adhd-friendly runs upstream of it — widen first, then hand a shortlist into its interview. Never run adhd-friendly inside a brainstorming session; finish or exit that first.

Phase 0 — Pre-flight (always; it is three questions and it is allowed to abort)

1. Is the question actually open? Abort and answer directly if it is a lookup, a known root cause, or phrased closed ("quick", "standard", "canonical", "textbook", "just tell me"). If Nile typed /adhd-friendly explicitly, skip this check — he opted in, do not second-guess him.

2. Say what it will cost, then ask.

This is ~N agent calls (one per frame) plus a converge pass, and maybe 30–90s. Worth it here?

Do not silently spend it. The source project's own docs disagree with themselves by ~4× on cost (README "1.9× output tokens" vs skill body "5–10×") because the real expense is context reload per branch, not output. Assume the larger number.

3. Pick 4–6 frames from references/frames.md, matching tags to the problem, including at least one wild. Name them out loud. Fewer good frames beats grinding the whole bank.

Phase 1 — Diverge (isolated, mechanically restricted)

Spawn one agent per frame, in a single message so they run concurrently, each with subagent_type: read-only-scout.

🔴 read-only-scout is not a suggestion — it is the enforcement. Its tools: Read, Grep, Glob frontmatter is an exclusive allowlist, so a branch is mechanically incapable of writing. Do not use general-purpose with a "don't write anything" instruction: that is prompt-as-enforcement, the exact anti-pattern behind the 2026-06-17 incident, and it is how the source project ships.

Each branch gets only: the problem, any context Nile supplied, its own frame prompt, and this instruction. It does not get the other frames, or any other branch's output.

You are generating, not judging. Produce 6 short, distinct ideas under this frame — one sentence each, plus one clause of rationale. Do not rank, evaluate, hedge, or caveat. The first three obvious answers are banned; those are the ones everyone already has. You may read the repo or corpus for grounding, but you cannot write anything. Return a JSON array only: [{"idea": "...", "why": "..."}]

⛔ The isolation invariant. Branches must be parallel and blind to each other. Do not serialise them, do not feed one branch's output to another, do not "summarise the others first." Branches that see each other anchor each other and the whole method collapses into one wider thought — at N× the price. Independence buys non-overlap; the frames buy coverage (A-AGT-05). You need both.

Phase 2 — Converge

Collect every branch's ideas into one list. Inject the SENTINEL-BAD fixture from references/frames.md into that list, unlabelled, before judging.

Then, in the main context:

1. Cluster near-duplicates across frames. An idea that three frames reached independently is a different kind of signal from one only the wild frame found — note which, but do not treat convergence as correctness.
2. Shortlist 3–5. Include at least one non-obvious survivor, or say plainly that there wasn't one.
3. Name the traps — the ideas that look good and are not, with the reason each dies.
4. Take a position. "Here are 20 ideas, you decide" is the cop-out this skill exists to avoid.

🔴 No scores. Veto, not vote.

Do not rank ideas 1–10, do not emit a rubric, do not print a decimal. pattern-extraction v3 was rewritten specifically to remove that layer: "No decimal unless a tool actually computed it" and "Veto, not vote — one well-grounded counterexample rejects the item regardless of how many other tests passed." Apply that here.

Two independent reasons, both load-bearing:

• Self-made rubrics — scoring against a rubric you wrote optimises the rubric (Goodhart). The source project's evals show exactly this: its huge deltas sit on dimensions its own output format guarantees, while its one format-neutral dimension moved +0.83 at n=6.
• Voting ensembles collude. Use one converge pass with a stated reason, not a panel.

⚠️ The sentinel gate

If SENTINEL-BAD survives into the shortlist, the converge stage is not discriminating. Report that and void the run — do not present the shortlist as if it were trustworthy. A check never shown capable of failing cannot distinguish "clean" from "broken."

Phase 3 — Report

Lead with the answer, not the machinery (the procedure is not the deliverable).

Mandatory accounting line — absence and failure must be distinguishable:

Frames run: 5 of 5 returned · 0 empty · 0 malformed

A branch that returned nothing, timed out, or emitted unparseable output is named, never silently absorbed. Malformed output is expected in practice, not exceptional. If fewer than half the branches returned, say the run was thin rather than presenting a confident shortlist.

Then: the shortlist with a taken position · the traps · one line on what the wild frame contributed (or that it contributed nothing).

Baseline mode — --baseline

Runs the same problem once, in one context, with all frames listed in a single prompt. No fan-out.

This exists because the claim "independence beats one-context-multi-frame" is untested by anyone, including the project the frames came from — its evals compared against a single-shot baseline, which is a strawman for that claim. Run --baseline occasionally on a real question and compare. If the isolated version stops earning its cost, that is a finding worth having, and this skill should shrink to a prompt template.

When this skill is the wrong tool

• The decision is already made → re-litigating it is a rabbit hole with good manners
• One good frame would do → just ask that one question directly, in-context
• The work is execution → adhd-friendly produces options, not progress
• It is tender or personal material → ⛔ a fan-out is the wrong instrument. Personal ground needs one presence, invited, not a panel of agents generating angles on someone's life.
• They said "I have work to do" → stop generating options; give the smallest next action

Self-test

1. Run on any real open question. Confirm the report includes the N of M line.
2. Confirm SENTINEL-BAD was rejected.
3. Control (both directions): remove the sentinel-rejection step and re-run — the run must now report a broken converge stage. If it still reports a clean shortlist, the gate is decorative.
4. Confirm branches were dispatched as read-only-scout, not general-purpose.

Provenance

Frames adapted from UditAkhourii/adhd (MIT) — the mechanism was worth taking, the packaging was not. Full evaluation, including why that project's evidence does not support its headline numbers and why caveman was rejected outright: KB Divergent Ideation — Three Repos, What Survived Review (2026-08-02).

The principles it rests on, stated plainly: independence is not diversity (the frames are what create coverage, not the mere fact of separate agents) · a subagent's value is its clean slate · containment belongs in tooling, never in a prompt instruction · voting ensembles collude · define the convergence criterion before you start iterating.