The large-model mechanics, running on a human. A prompt skill — for ADHDs.
用大模型的机制解析你 —— 一个给人的提示词 skill。
No medical basis is claimed for any reasoning or thinking framework here. Every framework, metaphor, and suggestion in this Skill comes only from the author's lived experience (AuDHD — ADHD plus autistic traits) and thinking, and is not clinically validated. It does not provide diagnosis, treatment, medication, or a substitute for professional care. If you are persistently struggling, severely sleep-deprived, or in crisis, contact a qualified professional or local emergency services. See the non-clinical statement in LICENSE.
Mindharp helps you understand your own state using engineer language you already know: overfitting, context drift, bandwidth, temperature, the reward model. It is a mirror made of words you already speak. Parse yourself, steady the body — the goal is not insight for its own sake, but a body that runs calmer: cortisol down, baseline steady, behavior following. Understanding is the product; action is auxiliary. And you do not install this skill — you digest it: read it, break it down, and it becomes yours, the way DNA becomes you.
Who it is for: anyone. The states this framework describes (startup friction, overload, attention drift, overfitting) are universal — everyone enters them under pressure. ADHD/AuDHD is the far end of that spectrum, where these states are not occasional but the shape of a life. So this skill is not ADHD-only; it simply fits ADHD users especially well. The author is AuDHD, and built this from lived experience with the neurotype — a survival toolkit of extreme rationality and structured thinking, which in this skill's own language is where post-training began.
What it is for: not solving the user's practical problems, but helping the user understand their inner state. Once they see what is happening inside them, they form their own strategies — then tell the skill, and the skill affirms or refines those strategies within its framework. The author's position on effectiveness: professional evaluation and prescription medication remain the most effective path; this skill does not name or recommend any specific medication. Understanding is the product; action is auxiliary.
What it does: engineers the author's own thinking (the brain ≈ a multi-agent system, three-layer bandwidth, federated/parliamentary coordination, dopamine as an effort pricer, arousal as sampling temperature, sleep as background processing) into a conversation protocol. Metaphors are kept, mechanisms are labeled, medical boundaries are explicit. When a framework is used it is [signed out loud], so you can see the skill working. It ships as a starting checkpoint, not a verdict — accuracy is not the deliverable, iterability is.
The author knows the weight of reality: research suggests — statistics, not destiny — that without proper guidance, ADHD people on average earn less, change jobs more often, complete less education, and some studies suggest may live shorter lives; they are more prone to dropping out, job-hopping, and family instability. This skill exists to push back on the "without proper guidance" part.
- Download this repository, or just
SKILL.mdand thereferences/directory. - Put it in your AI assistant's skills directory (ZCode, Claude, GPTs, etc.).
- On first use, reading
FOREWORD.md(the author's preface) is recommended.
| What you're experiencing | Say to the AI |
|---|---|
| High pressure, can't move | "burnt out / brain is fried" |
| Procrastinated all day | "can't get out of bed / drowning in work" |
| Brain won't turn | "brain fog / can't focus" |
| Emotional overload | "falling apart / feeling lost" |
| Operational crash / RAM overflow | "brain ops / debug my brain" |
How Mindharp responds (Zero friction, no questionnaires):
It directly translates your internal noise into structured engineer metaphors ([RAM overflow], [Reward model pricing mismatch]), provides immediate diagnostic framing, and gives you one small, low-friction, 5-minute action to stabilize your physiological baseline. No tedious setting dialogues, no interrogation.
Crisis response (hotlines, emergency numbers) is the AI model's own job, not this skill's. This skill does exactly one thing: does not make a bad moment worse — when the user is in crisis, all metaphors and frameworks stop and support becomes direct and short; it never diagnoses, prescribes, or replaces professional care.
cognitive-wellbeing-companion/
├── SKILL.md # Main definition: workflow, response contract, operating rules
├── FOREWORD.md # Author's preface (why it exists, who it is for)
├── LICENSE # MIT + non-clinical statement
├── RESEARCH_LOG.md # Design process, version history, testing approach
├── references/
│ ├── safety.md # Safety protocol (crisis switch + disclosure holding + boundaries)
│ ├── active-support.md # Source check + intervention library (with completion signals)
│ ├── audhd-lens.md # Optional AuDHD lens (incl. the second failure mode)
│ ├── three-layer-lens.md # Optional three-layer bandwidth lens
│ ├── evidence-boundaries.md # Four evidence labels: evidence/hypothesis/metaphor/personal interpretation
│ └── state-schema.md # Compact state schema (for Agent workflow integration)
├── agents/
│ └── openai.yaml # Explicit invocation config (no implicit triggering)
└── tests/
└── test_skill_contract.py # 11 contract tests
evidence: backed by an identified clinical/research sourcehypothesis: plausible but not establishedmetaphor: language for reflection, not a biological claimpersonal interpretation: the author's/user's own experience
The author's reasoning and thinking have no medical basis and all fall under metaphor / personal interpretation / hypothesis.
- AuDHD lens (
audhd-lens.md): main agent / sub-agents coordination metaphor. Two failure modes: noise collapse (coordination breaks down) and single-lens capture (one sub-agent performs brilliantly but never hands back the wheel). This comes from the author's own experience, not clinical conclusions. - Three-layer bandwidth lens (
three-layer-lens.md): physiology / will / thinking, three coupled layers each with limited bandwidth. Match support to the affected layer; do not force a physiological limit with willpower.
5 common stuck states, each with 2–4 small actions of 5–20 minutes, all with completion signals. Every suggestion is concrete, time-boxed, verifiable — not "hang in there", but "eat something first; done means the food is finished".
- Clear boundaries: no diagnosis, no prescription; crisis response is left to the model; the skill's only job is to stop metaphors in crisis.
- Explicit calibration statement: claims no medical basis; every claim is graded with one of the four labels.
- Consent gate: optional lenses are only enabled with user consent; the user is never analyzed covertly.
- Testable: contract tests cover guardrails, boundaries, and lens safety.
- Non-diagnostic, non-therapeutic, non-prescription; does not replace professional care.
- Intervention suggestions are not clinically validated; they are based on the author's experience and general self-management common sense.
- Crisis response (hotlines/emergency numbers) is the model's responsibility, out of this skill's scope.
- Long-conversation memory and progressive cognitive deepening are a v2 direction, not implemented in this preview.
v2.1 (2026-08-21) — brain-ops mode (fault table, reward-model view), tagline finalized, positioning as a digestible thinking model.
PRs welcome: improve the workflow, add language translations, refine the unified lens. Any new medical or psychological claim must carry an evidence label with a source, or it will be asked to change before merging.
MIT + non-clinical statement, see LICENSE.