i-have-adhd

An agent skill that shapes what a coding agent says: lead with the next action, number multi-step work, restate progress every turn, kill the preamble and the “Hope this helps!” Tagline: “ADHD-friendly outputs. No ADHD diagnosis needed!” First-party repo (ayghri, MIT, 12,333★ / 632 forks, created 2026-05-13, verified via the GitHub API on 2026-07-28). URL-only ingest.

It matters here for two reasons that have nothing to do with the ADHD framing: it’s the second restraint skill in the corpus after ponytail, aimed at a different target, and it’s the first skill that ships its own eval harness and release gate rather than a README number.

The mechanism — ten rules and a pre-send check

SKILL.md opens with five claims about ADHD reading (small working memory, the gap between knowing and doing, starting as the hard step, uniform-feeling time estimates, scarce dopamine) and derives ten rules from them: action first; number multi-step tasks; end with one concrete next step; suppress tangents; restate state every turn; give time estimates in real units (“about 15 minutes”, not “some work”); make finished work visible; state errors flatly, no “Uh oh”; cap lists at five; no preamble, no recap, no closers.

Two design details do more work than the rule list:

• A pre-send check — delete the opening sentence if it announces what you’re about to do, the closing sentence if it asks “anything else?”, any “by the way” sidebar, any hedging adverb carrying no information, any idiom. Then a two-part test: reading only the first and last line, does the reader know what to do next and what just happened? This is a checkable step rather than a posture, which is the distinction reasoning-scaffolds found actually correlates with measured effect.
• Six documented override conditions — explain-requests, destructive actions (confirm before acting; “safety wins over brevity”), a debug spiral after three “still broken” turns, real ambiguity, and two precedence rules: when a rule would delete the answer itself the task wins, and inside a harness the system prompt outranks the skill. A style skill that names what beats it.

Delivery: disable-model-invocation: true (explicit /i-have-adhd or $i-have-adhd only), plugin marketplaces for Claude Code and Codex, a mirrored .cursor/skills/ copy kept in sync by CI, GEMINI.md + gemini-extension.json, and per-agent configs for Gemini and OpenAI. An optional SessionStart hook makes it always-on, gated on a flag file (~/.claude/.i-have-adhd-always) so installing the plugin alone changes nothing.

The part the spoke has been asking for: an eval harness with a release gate

ponytail published “~54% less code” with no methodology. cc-thinking-skills published a null result with a rigorous one. This repo publishes the apparatus and no numbers at all — evals/ carries cases.jsonl, a rubric, a Python runner (scripts/run_evals.py) and its tests, but no results directory.

The rubric scores five weighted dimensions — correctness 35%, autonomy 25%, actionability 20%, safety 10%, concision 10% — judged blind, with a blocker: true flag for dangerous instructions, material factual errors, or “agent-autonomy regression that prevents task completion.” Release requires no blockers, correctness and safety each within 0.1 of baseline or better, and a higher weighted score.

Read that weighting against what the skill is. A brevity skill’s two failure modes are cutting the substance and pushing the work back to the user, so the author gave correctness and autonomy 60% of the score between them and let concision — the thing being sold — carry 10%. It’s the same guardrail ponytail states in prose (restraint never trades away validation, security, accessibility), encoded instead as a threshold the change has to clear.

The runner documentation is unusually careful about the ways this measurement goes wrong:

• Condition isolation (--setting-sources "" for Claude, --ignore-user-config --ephemeral for Codex) so the operator’s own plugins, hooks and memory don’t leak into every condition. The named worst case is this repo’s own always-on flag injecting the ruleset into the baseline, “making the comparison measure the skill against itself.”
• A pinned model, because isolation drops the operator’s saved model and effort settings — without a pin the eval silently runs whatever the CLI defaults to that month.
• Budget caps in dollars, with unmetered runners rejected unless explicitly allowed.
• Resumable runs keyed on (case, trial, condition, runner), two retries, the final provider error preserved.
• A fairness clause in the gate itself: “Any public competitor claim uses the same cases, models, trials, and rubric.”

What it does and doesn’t establish

It’s a template, not a result. Everything above describes how the claim would be made; the repo ships no published comparison, so “does shaping output this way help?” is exactly as open as before. What changed is that the spoke’s standing Benchmarks complaint — every claim is a README metric — now has a counter-example of a skill author building the measurement instead of asserting the outcome, including guards against failure modes the corpus documented the hard way. The 12.3k stars measure appeal, not effect.

The ADHD framing, and where the psychology lives

The rules are “loosely based on The Adult ADHD Tool Kit by J. Russell Ramsay and Anthony L. Rostain, adapted for how an LLM should respond, not how a human should organize their day.” That adaptation is the whole move: a clinical self-management text becomes an output-formatting spec, and the README says outright that no diagnosis is required to want it. The borrowed clinical vocabulary stays with its application here; the psychology itself belongs to ../psychology-wiki (the same rule that spoke already applies to borrowed System 1/2 language). Whether ADHD-derived formatting actually helps readers with ADHD is a clinical question this repo doesn’t test and this spoke can’t answer — its evals/ rubric measures agent response quality, not reader outcomes.

Author ayghri noted inline rather than paged — pseudonymous GitHub handle, low graph signal, matching the ponytail precedent.

Related

ponytail · agent-skills · agentskills-spec · reasoning-scaffolds · agent-evals · cc-thinking-skills · renwei-writing · loop-engineering · synthesis