The Pain Every AI-Assisted Coder Knows
You ask Claude: "My auth flow errors out—token verification fails."
The AI replies with a wall of text: "Great question! Let me think about this..." It restates your problem, narrates its reasoning, adds a "by the way" about dependencies, and closes with "Hope this helps!" The actual instruction—"update the package and rewrite that function"—is buried in sentence seven.
The answer is buried.
This is the problem i-have-adhd, a GitHub project that blew up in July 2026, set out to solve. An ML PhD built it with 143 lines of Markdown and zero lines of code, collecting 9,236 stars in two months. It contains no code because it *is* a skill file that teaches AI coding assistants how to respond.
After installation, the same question gets this answer:
> 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 next action is on line one. Steps are numbered. It ends with a concrete next step. No "Great question," no "Hope this helps."
Why It Went Viral: The Pain Point Is Universal
9,236 stars in two months for 143 lines of Markdown is an abnormal ratio. For comparison: stop-slop, a similar de-AI-flavor skill, took three months to reach 2,000 stars; llama.cpp took two years to reach 50,000.
The virality isn't technical novelty—de-AI-flavor skills existed before. It's that the pain point is universal: everyone using AI to code is sick of "Great question!", of digging through paragraphs for the one actionable sentence, of losing context by round three.
But the author did one thing differently: he didn't invent rules from intuition—he derived them from ADHD neuroscience.
Five Facts About the Brain: The Scientific Foundation
The SKILL.md opens not with rules but with five facts about the ADHD brain. Each rule derives from them—the project's smartest design decision:
1. Working memory is small. Anything not on screen is forgotten. Hence: lead with the next action. 2. Knowing the answer is not doing the answer. The core ADHD difficulty isn't ignorance but initiation. Hence: number multi-step tasks—numbering lowers the activation barrier; the first step must be doable *now*. 3. Starting is the hardest step. Hence: end with one concrete next step, so the reader never has to decide what comes next. 4. Time estimates feel uniform. "A bit of work" and "hours" register identically. Hence: specific time estimates—minutes, not "a while." 5. Dopamine is scarce. Buried wins don't register. Hence: make wins visible.
These come from consensus in ADHD cognitive neuroscience (working memory deficits, executive dysfunction, dopaminergic reward shortfalls). The README states it is "loosely based on *The Adult ADHD Tool Kit* by J. Russell Ramsay and Anthony L. Rostain," a CBT-based adult ADHD coping guide.
Designing from neuroscience instead of preference is what separates i-have-adhd from all prior de-AI-flavor skills.
The Ten Rules: From Facts to Operations
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 "oops, tests failed, let's see!" Just: "auth.spec.ts:23 failed, expected true, got false."
9. Cap lists at 5 items
10. No preamble, no recap, no closers
Every rule maps to an ADHD cognitive trait. Neurotypical users don't need these rules, but they don't hurt them either—designing for the most vulnerable users often benefits everyone, the core principle of Universal Design.
The Pre-Send Check: An Actionable Formula
The SKILL.md's best part is the closing pre-send check—a linter for AI output:
Delete: 1. The first sentence, if it announces what you're about to do 2. The last sentence, if it asks "Anything else?" or recaps what just happened 3. Any "by the way" asides 4. Any vague hedges that add no information ("maybe," "possibly")
Keep: hedges that express real uncertainty—deleting them manufactures false confidence.
Replace: idioms and metaphors with literal actions.
Verify: 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.
Like a code linter, it compresses ten rules into one pre-flight pass.
A Deeper Insight: CBT for LLMs
CBT's core logic: don't change your abilities—change your environment to compensate for them. ADHD patients don't need to "try harder"; they need keys by the door and pills next to the toothbrush. Environment design compensates for cognitive differences.
i-have-adhd does the same thing with the roles reversed: instead of changing the ADHD user's environment, it changes the AI's output environment to fit the user's cognition. Default AI output—pleasantries, throat-clearing, closers—is designed for neurotypical reading. For ADHD users, it's where work goes to die.
This isn't "make AI more concise." It's shape-matching: make the output's shape match how the ADHD brain processes information. Not shorter—more actionable.
The Controversy: The Cost of Deleting Context
The sharpest criticism: i-have-adhd deletes the context you need to make your own judgments.
When the AI only tells you *what* to do, not *why*, you lose:
- The context to judge whether the advice is correct
- The chance to learn the domain knowledge
- The ability to detect AI errors
A five-dimension scoring framework adds nuance (correctness 35%, autonomy 25%, actionability 20%, plus two more): "good output" isn't one-dimensional. i-have-adhd maxes actionability but may sacrifice autonomy; different users and scenarios need different trade-offs.
Connection to Sycophancy Research: Output-Level Alignment
i-have-adhd represents a third kind of intervention—no RLHF changes, no training data changes, just shape-shaping at the output layer via a skill file:
| Intervention layer | Example | Cost | Effect | |---|---|---|---| | Training layer | RLHF reward adjustment | Very high (retraining) | Global | | Inference layer | Prompt-engineering frameworks | Medium | Scenario-specific | | Output layer | i-have-adhd skill | Very low (143 lines of Markdown) | Output shape |
i-have-adhd proves output-layer intervention has an extraordinary cost-benefit ratio. This is community-driven alignment—bottom-up rather than top-down.
A Philosophical Observation: The Knowing-Doing Gap
The second fact is the most profound: knowing the answer is not doing the answer. This isn't just an ADHD problem—it's a fundamental human one. You know you should exercise, write that paper, reply to that email. The gap between knowing and doing is bridged or widened by environment design: if the running shoes are by the door, you run.
The shape of AI output determines whether users can cross the knowing-doing gap. Not whether the answer is correct—whether its shape makes action possible. That applies to everyone who has ever struggled to act on what they know. Which is everyone.
Concept Genealogy: Output Shape as an Alignment Dimension
i-have-adhd extends a genealogy of alignment-intervention levels:
1. RLHF — training layer, change the reward function 2. Epanorthosis calibration — inference layer, change the prompt 3. Beyond Sycophancy frameworks — inference layer, change interaction structure 4. i-have-adhd — output layer, change the output shape
Four levels: cost decreasing, customizability increasing. Sometimes alignment doesn't require retraining a model. Sometimes it just takes a skill file.
---
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 Author's blog: ayghri.me