Key points
- The problem: AI coding assistants bury the actual instruction. Asked about a broken auth flow, Claude opens with "Great question!" and the actionable advice appears in the seventh sentence, surrounded by pleasantries, restatement, and filler like "Hope this helps!"
- The project: i-have-adhd (MIT, 9,236 stars in two months) is a 143-line Markdown "skill" file — zero code — that teaches AI assistants to output differently. For comparison, stop-slop (a similar de-slop skill) took three months to reach 2,000 stars.
- Delete: the first sentence if it announces what you're about to do; the last sentence if it asks "anything else?"; any "by the way" asides; vague adverbs (maybe/possibly/probably)
- Keep: genuine uncertainty — deleting it manufactures false confidence
- Replace: idioms and metaphors with literal actions
- Verify: if a reader reading only the first and last lines knows (a) what to do next and (b) what just happened, send it.
- Project: ayghri/i-have-adhd (MIT, 9,236 stars)
- SKILL.md: skills/i-have-adhd/SKILL.md
- Theoretical basis: *The Adult ADHD Tool Kit* by Ramsay & Rostain (CBT for adult ADHD)
- Related: stop-slop
- Author blog: ayghri.me
The neuroscience basis
The SKILL.md opens with five facts about the ADHD brain, and every rule is derived from them:
1. Small working memory → lead with the next action (rule 1) 2. Knowing ≠ doing → number multi-step tasks (rule 2) 3. Initiation is the hardest step → end with one concrete next step (rule 3) 4. Time estimates feel uniform → specific time estimates in minutes (rule 6) 5. Dopamine scarcity → make wins visible (rule 7)
The README notes it is "loosely based on *The Adult ADHD Tool Kit* by J. Russell Ramsay and Anthony L. Rostain" — a CBT-based adult ADHD guide. Using neuroscience as design principles, rather than personal preferences, is what distinguishes i-have-adhd from earlier "de-AI-slop" skills.
The ten rules
1. Lead with the next action
2. Number multi-step tasks
3. End with one concrete next step
4. Suppress tangents
5. Restate state every turn
6. Specific time estimates
7. Make wins visible
8. Matter-of-fact errors (no drama: "auth.spec.ts line 23 failed, expected true, got false")
9. Cap lists at 5 items
10. No preamble, no recap, no closers
The pre-send check
The most operational part of SKILL.md is a four-step deletion filter plus one verification:
This works like a linter for AI output — not memorizing ten rules, but running one filter before sending.
Deeper insight: CBT for LLMs
CBT's logic is: don't change the person's ability — change the environment to compensate. i-have-adhd applies the same logic in reverse: instead of changing the ADHD user's environment, it changes the AI's *output environment* to match how an ADHD brain processes information. The output isn't shorter for its own sake — it's shaped for actionability.
The controversy: cost of deleting context
The sharpest criticism: stripped context removes your ability to judge whether the advice is correct, learn the underlying domain knowledge, or catch AI errors. This is a real trade-off between action speed and autonomous judgment. For ADHD users it may be right; for neurotypical users it risks turning them into executors of AI rather than collaborators. A five-dimension scoring framework mentioned alongside (correctness 35%, autonomy 25%, actionability 20%, plus two more) makes the point that "good output" is not one-dimensional.
Connection to sycophancy research: output-layer alignment
Placed alongside earlier work, i-have-adhd is a third kind of intervention:
| Intervention layer | Example | Cost | Scope | |---|---|---|---| | Training layer | RLHF reward adjustment | Very high (retraining) | Global | | Inference layer | Prompt-level frameworks | Medium | Scenario-specific | | Output layer | i-have-adhd skill file | Very low (143 lines) | Output shape |
It demonstrates the extreme cost-effectiveness of output-layer intervention: no retraining, no RLHF changes — just a skill file reshaping a large model's output. Alignment doesn't have to be top-down; it can be community-driven from the bottom up.
Philosophical coda: the knowing-doing gap
The second "fact" is the deepest: knowing the answer is not doing the answer. The friction between knowing and doing is where work dies — and the width of that gap is set by environment design. If sneakers are by the door, you run; if the answer is in the first line, you act. The AI's output shape determines whether users can cross the knowing-doing gap — true not just for ADHD users, but for everyone.
Links
FAQ
Who is this for? Practitioners, researchers, and students interested in AI, machine learning, and developer tooling.
Core takeaways? Output-layer intervention is cheap and effective; derive rules from neuroscience rather than preferences; shape AI output for actionability.
Open-source? Yes — MIT-licensed, links above.