Skip to content

About

Response-style skill for AI coding agents, for twice-exceptional (2e: gifted + ADHD) users: expert depth, ADHD-friendly form. Answer first, one thread at a time. Derived from i-have-adhd.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

82 Commits

Folders and files

Repository files navigation

twice-exceptional-skill: a rainbow infinity with a star and lightbulb (gifted) and a lightning bolt with open tabs (ADHD)

Full expert depth, ADHD-friendly form. Answer first, no filler.

License: MIT

Response-style skill for coding agents (Claude Code, opencode): it changes the form of the output, not the technical level.

Twice-exceptional (2e) describes people with high ability in one or more domains and one or more disabilities, such as ADHD, a learning disability or autism (NAGC position statement, after Reis, Baum and Burke 2014). The two can mask each other, so neither is obvious. This skill targets giftedness with ADHD: expert-level understanding, with an attention and executive cost in turning "understood" into "done". Its rules help any reader who wants the answer first.

Derived from ayghri/i-have-adhd by Ayoub Ghriss (MIT); see Credits and tooling.

What it does

Expert-peer level, with a form that lowers the attention cost of going from "understood" to "done".

Before After
Context, preamble, "there are several aspects to consider…" First line = command, path or decision
Warning at the bottom ⚠️ above the steps, and only for real risks; guesses are labeled "Hypothesis:"
Long prose Tables, diagrams, hierarchies
Several topics mixed One thread at a time, switches flagged, 🅿️ parking
"Hope this helps!" No preambles or courtesy closings
"You could also try…" after every answer Extras only for real problems; a finished task ends on "Done: …"
Re-answering the same question in new words After the third time: name what would change the answer, offer to build it
Agreeing when the user pushes back Position changes only for a new fact or argument

Install

git clone https://github.com/JohnnyTriggerDev/twice-exceptional-skill.git
cd twice-exceptional-skill
./scripts/install.sh

Idempotent; installs for Claude Code and opencode and loads the skill in every session. On Windows: powershell -ExecutionPolicy Bypass -File scripts\install.ps1. For claude.ai, manual steps and verification: docs/INSTALL.md.

Turn it off

Scope How
Current session Write "normal mode" or "stop twice-exceptional-skill"
Permanently Remove the loading lines, see docs/INSTALL.md#disable

Examples (TL;DR)

Five invented prompts, same model, with and without the skill. Full replies side by side: docs/EXAMPLES.md.

Prompt Without the skill With the skill Length
Remove a Docker volume that may be in use A tutorial with three steps and the caveats at the end docker ps -a --filter volume=<name>, then ⚠️ back up first, above the steps 1,759 → 862 (-51%)
SQLite or Postgres for ~100 users? "Probably SQLite", then four long lists and a question back "Use SQLite.", a ⚠️ with the exception, a table of alternatives 1,966 → 1,392 (-29%)
i want to add caching to the flask api, redis or in memory, A two-option guide with a comparison table "Use Redis.", ⚠️ on the per-worker cache, asks "Plan first or execute?" 3,502 → 1,591 (-55%)
A rough call with the manager, thinking of quitting Empathy, then a list of questions Empathy, "no need to decide tonight", one question 832 → 408 (-51%)
Move cities to be closer to family? Three blocks of questions and a regret test No verdict: the two questions that decide it, offers to draft a message to the family 1,877 → 812 (-57%)

Five of eight pairs, picked for readability, one run each: a demo, not a measurement. Without the skill a reply is sometimes the more complete one.

How it works

Always-on through global instructions, not only when the model picks the skill (docs/ARCHITECTURE.md).

flowchart TD
    A[User message] --> P{"Bare preamble ending on a comma?"}
    P -- yes --> P1["Reply only: Yes, go on."]
    P -- no --> B{"Deviation applies?"}
    B -- "emotional topic / explain / destructive action / ..." --> C[Deviation handling]
    B -- no --> D["First line: answer or action"]
    D --> E{"Risk in the steps?"}
    E -- yes --> E1["⚠️ above the steps"]
    E -- no --> F
    E1 --> F["Steps / table / diagram: one representation per concept"]
    F --> G{"Does a Why add new information?"}
    G -- yes --> G1["Why / Details"]
    G -- no --> H
    G1 --> H{"Thread dropped in this message?"}
    H -- yes --> J["🅿️ Parked"]
    H -- no --> I["📌 To remember, only if it changed"]
    J --> K["Last line: Next on the current topic, Done, or Done, not run"]
    I --> K
Loading

Shape of a response:

┌──────────────────────────────────────┐
│ 1. Answer / next action              │  ← the reader can stop here
├──────────────────────────────────────┤
│ ⚠️ Risks, only if there are any      │
├──────────────────────────────────────┤
│ Steps / table / diagram              │
├──────────────────────────────────────┤
│ Why / Details, only if new           │
├──────────────────────────────────────┤
│ 🅿️ or 📌 (optional) · Next or Done   │
└──────────────────────────────────────┘

Across a session: "Now:" when the goal changes, "Back to:" when a parked thread resumes, steps named rather than numbered, and the same term for the same concept (docs/RULES.md).

Rules at a glance

Scope, precedence, 13 rule sections, 7 deviations and a two-line pre-send check, about 2,400 words. Map and tables: docs/RULES.md. Full text: SKILL.md.

Section What it does
Scope Always on; form proportional to the content (a short answer is two sentences, structure only when the content has that shape); says which skill is active if asked
Precedence Environment instructions, then safety (including crisis language), then deviations, then the rest; nothing suppresses a warning or a confirmation
Answer first, Warnings, Structure Answer or action on the first line with a one-clause reason; ⚠️ above the steps, only for risks; tables and diagrams, one representation per concept
One thread, Switches and parking No extras except real problems; divergent mode; "Now:", "Back to:", 🅿️ Parked
Doing, Verification Do, don't ask, with a side-effect boundary; check with the tools available or mark "Hypothesis:"
Input, Progress Fragments, typos, dates; status, resume points, stop points in long sessions, tedious and avoided tasks
Ready-to-use text, Tone, Memory Texts ready to send; direct, not cold; same terms, 📌 To remember
Closing Last line Next, Done or "Done, not run"; at most one 🅿️ or 📌 block above it
Deviations Emotional topics, explain, destructive action, debug spiral, question loop, ambiguity, options with pros and cons

Limits

  • Prompt-only: it shapes how the agent writes, and the model can still drift from it. CI checks the docs, the diagrams and both install scripts, not the model's compliance; success metrics count corrections in real conversations; one private replay test is summarized in docs/RESEARCH.md.
  • System and environment instructions win over the skill (Precedence 1); the form stays.
  • Not a mental-health tool: on personal and emotional topics the skill steps aside, and it never presents itself as a substitute for a person.

Evidence

The cited papers document the risks the rules target: sycophancy, dependence on always-available assistants, generic referrals, and lost context after interruptions. No paper tests these rules as prompts, and none tests the question-loop deviation or the "Done" closing directly. Five of the fourteen papers are arXiv preprints. A separate table lists the ADHD and giftedness papers behind the traits that the strengths rules address. A replay test on one user's real conversations (27 prompts, 3 sessions) preferred the skill in 29 of 30 blind comparisons, with one tie; the judge used the skill as its rubric, so read its limits first. On DeepSeek V4.1 Flash, judged blind by Claude Haiku and GLM, version 7 beats no skill on 40 synthetic ADHD scenarios (32 to 7) and on the i-have-adhd eval suite (weighted 4.27 vs 3.82 with Haiku, 4.63 vs 4.48 with GLM): benchmarks on other models. On the same suite with Claude Opus 4.8, version 7 scores 4.72 against 4.09 without a skill (Opus judge; 4.74 against 4.17 with GLM). On GPT-6 Luna the same suite gives 4.80 against 4.60 (MiMo judge), but the 40 synthetic scenarios go the other way: version 7 loses to no skill. Full tables, with each finding, the rule it supports and its limits: docs/RESEARCH.md.

Documentation

File Content
skills/twice-exceptional-skill/SKILL.md The skill (source of truth)
docs/RULES.md Map of the rule sections, the deviations and the old rule numbers
docs/INSTALL.md Global install for Claude Code and opencode
docs/ARCHITECTURE.md Why a skill plus automatic loading
docs/EXAMPLES.md Five before/after replies, side by side
docs/RESEARCH.md Papers, findings, use and limits

Credits and tooling

Source Use
ayghri/i-have-adhd, Ayoub Ghriss, MIT Original work this skill derives from: core idea and rule-based structure
Kacper Rutkiewicz, AI Made Simple Video breakdown of an earlier version of i-have-adhd, cited by the original README
Claude Code memory and @import Automatic loading via ~/.claude/CLAUDE.md
opencode rules and instructions Automatic loading via opencode.json

Both documentation links checked on 17 October 2026: @path imports in ~/.claude/CLAUDE.md and the instructions array in opencode.json work as described here.

License

MIT, with the copyright of the original author and of this derivative: see LICENSE.

About

Response-style skill for AI coding agents, for twice-exceptional (2e: gifted + ADHD) users: expert depth, ADHD-friendly form. Answer first, one thread at a time. Derived from i-have-adhd.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages