What does this skill do, and when should you use it?
i-have-adhd is a Claude Code skill that, once activated, reshapes every response in the session for a reader with ADHD: the first line is the next executable action, multi-step work is numbered, progress is restated every turn, time estimates use concrete minutes, and errors state only cause and fix. It adapts *The Adult ADHD Tool Kit* (Ramsay & Rostain) into ten hard output rules plus six documented exemptions (full explanations, destructive actions, debug spirals, ambiguity, rule-vs-task conflicts, harness conflicts). Invoke it with /i-have-adhd; it stays on until you say "stop adhd mode". The repo bundles two skills — this profile covers only the one at .cursor/skills/i-have-adhd/SKILL.md.
- Rewrites the first line of each reply as an immediately actionable step, with commands, paths, or snippets up top
- Forces multi-step work into numbered lists of bounded actions, cutting or folding unnecessary steps
- Restates session state each turn ("Step 3 of 5 done") so nothing relies on working memory
- Converts vague estimates into concrete units ("about 15 minutes", not "a bit of work")
- Suppresses tangents, forbids preambles/recaps/closers, and reports errors as cause plus fix
- Runs a pre-send check: first and last line alone must reveal the next action and what just happened
- An ADHD developer in Claude Code who is tired of three paragraphs of context before the actual fix wants "edit src/auth.ts:42" as the opening line
- A multi-turn refactor where the reader loses track between messages and needs the agent to restate progress automatically
- A reader who struggles with effort estimation and needs "about 15 minutes" instead of "some work" to decide whether to start now
- A debugging session where the agent keeps appending "by the way, your dependencies are stale" and tangents should be split off to the end
- Situations that need long teaching-style walkthroughs — the skill permits full explanations on "explain", but its overall orientation is compression, not depth
- Users on non-Claude Code platforms who won't adapt it: the disable-model-invocation flag and /i-have-adhd slash invocation are Claude Code mechanisms and need rewriting elsewhere
How do you install this skill?
- Static review only: no code or skill instructions were executed; scores are low-confidence source-level judgments.
- The skill overrides output style for the entire session; you must remember the 'stop adhd mode' off phrase, or it keeps shaping every response.
- Rules suppress preambles and hedging; safety caveats or uncertainty could be over-compressed — re-verify in safety-critical contexts.
- Publisher is unverified and there is no changelog; diff the .cursor copy against the canonical copy before upgrading.
- No declared Chinese/multilingual output support; Chinese-language users should verify the rules' effect on Chinese responses themselves.
The README's install path is a paste-in prompt that delegates to the repo's AGENTS.md; no per-step manual commands are documented:
# Paste into your CLI prompt:
Install the i-have-adhd skill/plugin from https://github.com/ayghri/i-have-adhd, refer to the repo's AGENTS.md for instructions.Claude Code (plugin marketplace route implied by the README's tuning section):
claude plugin marketplace add ayghri/i-have-adhd
claude plugin install i-have-adhd@i-have-adhdNote: the marketplace commands are reverse-engineered from the fork-tuning steps; INSTALL.md contents are not in the source, so verify against the repo.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- /i-have-adhd
- Fix the failing test at auth.spec.ts:42 and lead with the next action
- Break this refactor into numbered steps with concrete time estimates for each
- stop adhd mode
Activate with the slash command:
/i-have-adhdRules persist for the whole session and survive topic changes. Turn them off by saying "stop adhd mode" or "normal mode"; the agent confirms in one line and returns to its default style. Inside a harness with a task/plan tool, the skill directs the checklist to do the state restating instead of narrating the plan in prose. Built-in exemptions: full expansion on "explain"; confirmation before destructive actions (rm -rf, force push, schema migrations); after three consecutive "still broken" turns it stops iterating, names the possibly-wrong assumption, and asks one diagnostic question.
What are this skill's strengths and limitations?
- Ten rules are concretely verifiable (e.g. no step contains "and then" twice), not vague style advice
- Explicit off-switch and six exemption scenarios prevent the rules from backfiring on explanations or safety checks
- Zero runtime dependencies: pure Markdown instructions, no shell, network, or filesystem needed
- Rule 9 explicitly scopes the 5-item list cap to presentation only — no silent information loss
- Effectiveness rests entirely on model instruction-following; no tests or benchmarks are provided for any platform
- disable-model-invocation and the slash command are Claude Code-specific; other Agent Skills clients need frontmatter and trigger adaptations
- The compressed style may not suit readers who want context-first answers; no graduated tuning is documented
- "No ADHD diagnosis needed!" is claimed but the value for non-ADHD users is unevidenced
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| i-have-adhd: ADHD-Friendly Output Mode this page | 65 · Recommended | ★ 44k | 1mo ago | MIT |
| ELI5 Audience-Adaptive Explanation Skill | 64 · Recommended | ★ 1.7k | 6mo ago | MIT |
| Claude Council | 47 · Use with care | ★ 95 | 2mo ago | — |
| Project Memory Manager | 59 · Recommended | ★ 2.9k | 3d ago | Apache-2.0 |
| PlanMySaaS | 55 · Use with care | ★ 12 | 5mo ago | MIT |
The source cites *The Adult ADHD Tool Kit* by Ramsay & Rostain as its loose inspiration, but that is a human time-management book, not a competing skill; no alternative skills are named.
How did FollowSkills review this skill?
Pure output-formatting instructions: no permissions requested, no external actions, no sensitive-data flows. Destructive actions (rm -rf, force push, migrations) explicitly require confirmation (override rule 2); session persistence has a clear off switch ('stop adhd mode'). Deducted for: unverified publisher; attribution rests only on MIT copyright, with no independent signature or provenance endorsement.
Instructions are self-consistent: no contradictions among the ten rules, and dedicated handling for rule conflicts, harness conflicts, debug spirals, and ambiguity — stronger edge-case coverage than most prose skills. Static calibration caps this at 10: the repo has tests and a sync CI, but they target hooks and the eval harness, not the skill file's own key paths, so real output behavior cannot be verified statically.
Trigger semantics are clear: explicit /i-have-adhd invocation, disable-model-invocation: true prevents false auto-triggering, explicit off phrase, and override rules (explain, destructive actions, debugging, ambiguity) bound non-fit cases. Deducted for: no declared Chinese/multilingual output support, all examples in English, and boundaries that remain partly inferential.
Well-layered docs (SKILL.md body, multilingual README, INSTALL, credits to Ramsay & Rostain), explicit MIT, version 0.4.1 in package., CI enforcing sync between the .cursor copy and the canonical copy (#55 fix documented). Deducted for: no CHANGELOG file, no explicit maintenance-ownership or update-path commitment, and README points to skills/ while the assessed path is the .cursor copy (CI mitigates but ambiguity remains).
Concrete before/after contrast is directly reusable; each rule ships paired good/bad examples; output needs no rework in principle. Deducted for: static review cannot confirm the rules are followed consistently in live sessions; marginal benefit is scenario-dependent (ADHD readers), the cited third-party video review covers an earlier version, and no current effectiveness data is committed.
Auditable primary material exists: test suites (hooks, install docs, judge), an evals framework, and a sync CI workflow; README cites a third-party video and an academic source. Deducted for: static read cannot reproduce execution, eval result data is not committed with the source, and the third-party corroboration targets an older version.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →