The Pain: Answers Buried Under Filler
Anyone who has used an AI coding assistant knows the pattern. Ask Claude why a token verification fails, and the response starts with "Great question! Let me think about this," restates the moving pieces, mentions a related topic "by the way," and finishes with "Hope this helps! Let me know if you want to dig deeper." The actual instruction—what to do next—lands around sentence seven.
i-have-adhd was created to fix this. Created by an ML PhD, it consists of 143 lines of Markdown and zero lines of code, because it is a skill file that reshapes how AI programming assistants produce output. After installation, the same question gets an answer structured like:
> Run npm install jsonwebtoken@latest, then edit src/auth.ts:42.
> 1. Open src/auth.ts
> 2. Replace verifyToken (lines 42–58) with the snippet below
> 3. Run npm test -- auth.spec.ts
>
> Next: paste the first failing line if any test fails.
The first line is the next action. Steps are numbered. The closing line is one concrete next step. No "Great question," no "Hope this helps," no "by the way."
Why It Went Viral: A Universal Pain Point
9,236 stars in two months from a 143-line Markdown file is an unusual growth curve. For comparison, stop-slop—a similar de-AI-flavoring skill—took three months to reach 2,000 stars, and llama.cpp needed two years to reach 50,000. The pain point is universal: every AI coder is tired of "Great question!" and has lost context by the third turn. The novelty was not the goal—de-AI-flavoring skills already existed—but the lever used to reach it.
Five Neuroscience Facts as Design Principles
The project opens not with rules but with five facts about the ADHD brain. Every rule derives from these facts.
1. Working memory is small. Anything off-screen is forgotten. Rule: lead with the next action. 2. Knowing the answer is not doing the answer. Friction between knowledge and execution is where work dies. Rule: number multi-step tasks so the first step is small and immediate. 3. Starting is the hardest step. Rule: end each turn with one concrete next step so the reader does not have to decide what comes next. 4. Time estimates feel uniform. "A little work" and "a few hours" register the same. Rule: use minutes, never "in a while." 5. Dopamine is scarce. Visible progress matters. Rule: make wins visible so the brain registers a reward.
These facts come from consensus in ADHD cognitive neuroscience: working memory deficits, executive dysfunction, and reward-pathway dopamine insufficiency. The README cites *The Adult ADHD Tool Kit* by J. Russell Ramsay and Anthony L. Rostain, a CBT-based guide for adults with ADHD. Designing for the most fragile users often benefits everyone—an instance of Universal Design.
Ten Derived Rules
From the five facts come 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 (in minutes) 7. Make wins visible 8. Matter-of-fact errors (no "oops" theater) 9. Cap lists at 5 items 10. No preamble, no recap, no closers
Rule eight is especially notable: AI loves to dramatize failures ("Oops, the test failed, let's see what's going on"). The replacement is just: "auth.spec.ts line 23 failed, expected true, got false."
The Pre-Send Check: An Output Linter
The most actionable part of SKILL.md is the Pre-send check—a compact checklist that compresses the ten rules into one workflow.
Delete:
- The first sentence if it announces what you are about to say
- The last sentence if it asks "anything else?" or restates what just happened
- Any "by the way" aside
- Any vague adverb that adds no information ("maybe," "perhaps," "probably")
- Context to evaluate whether the suggestion is correct
- A chance to learn the domain
- Ability to catch errors when AI is wrong
- 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 project: stop-slop
Keep: vague hedges that signal genuine uncertainty. Removing them manufactures false confidence.
Replace: idioms and metaphors ("circle back," "get the ball rolling," "align on") with literal actions.
Validate: if the reader reads only the first and last lines, do they know (a) what to do next and (b) what just happened? If yes, send.
This is essentially a linter for AI output—automatic checks at save time, rather than a list of style rules to remember.
CBT for LLMs: Environment Design Applied to Output
i-have-adhd traces back to a CBT-based ADHD toolkit. CBT's core logic is not to change ability but to change the environment to compensate for the ability gap. A person with ADHD does not need to try harder; they need their keys on the entryway shelf and their pill bottle next to the toothbrush.
i-have-adhd applies the same logic in the reverse direction: instead of changing the user's environment, it changes the AI's output environment to compensate for the user's cognitive profile. Default LLM output is shaped for neurotypical readers. For an ADHD reader, buried answers create friction large enough to kill the task. i-have-adhd reshapes output to match how ADHD brains process information.
The Controversy: Action Speed vs. Independent Judgment
The sharpest critique is that i-have-adhd removes the context users need to judge advice themselves. When AI outputs "what to do" without "why," you start faster but lose:
This is a real trade-off: action speed vs. autonomous judgment. For an ADHD user short on working memory, speed may be the right bet. For a neurotypical user, it may reduce them to "AI executor" instead of "AI collaborator." A five-dimension scorecard (correctness 35%, autonomy 25%, actionability 20%, and two more) frames the trade-off more carefully.
Output-Layer Alignment
This project is part of a broader pattern of interventions at different layers:
| Layer | Example | Cost | Effect | |-------|---------|------|--------| | Training | RLHF reward tuning | Very high (retraining) | Global | | Inference | Anti-sycophancy prompts | Medium (prompt engineering) | Context-specific | | Output | i-have-adhd skill | Very low (143 lines Markdown) | Output shape |
The takeaway: alignment does not require top-down retraining. A 143-line Markdown file can reshape the output of a 744B-parameter model at almost zero cost—an example of bottom-up, community-driven alignment.
The Knowing-Doing Gap
The deepest observation comes from fact two: knowing the answer is not doing the answer. The friction between knowledge and action is where work dies. This is not only an ADHD problem; it is a fundamental issue in human cognition.
If your running shoes are by the door, you run. If they are buried in a closet, you don't. If the answer is in line one, you act. If the answer is in sentence seven, you may not. AI output shape determines whether users cross the knowing-doing gap.