Attention Control is a reusable output style that tells an agent how to structure its replies. It leads with the action, keeps wording controlled, and uses a fixed shape so the next step is easy to find. The repo ships the style in several forms: a Claude plugin, a Codex plugin, cursor rules, Gemini settings, and shared prompt files. It also includes evals and sync scripts so you can test and copy the style across agents.
Builders who want their agent replies to be easier to scan and act on.
You can get agent replies that lead with the next action and stay consistent turn after turn.
Defines 11 rules for order, length, and structure, including leading with the next action and ending with one concrete next step.
Limits word choice, tense, and sentence shape so the agent writes in a more direct, consistent way.
Ships the same style for Claude Code, Codex, Cursor, Gemini CLI, and manual installs.
Supports a one-time skill invocation as well as persistent setup through settings or `AGENTS.md`.
Includes scripts and cases to validate and score the style against an unstyled baseline.
Uses `scripts/sync_style.py` to regenerate agent-specific copies from the shared style file.
claude plugin marketplace add aaddrick/attention-control
claude plugin install attention-control@attention-control
/attention-control:attention-control
codex plugin marketplace add aaddrick/attention-control --ref main
codex plugin add attention-control@attention-control
$attention-control:attention-control
Attention Control
Air traffic control discipline for agent output.
Written for a reader with ADHD.
English · 简体中文 · 日本語 · 한국어 · Tiếng Việt · Português (BR)
Claude Code is the only agent here with a native output style slot. Run these two commands in your terminal:
claude plugin marketplace add aaddrick/attention-control
claude plugin install attention-control@attention-control
Then run /config, select Output style, and pick Attention Control. It
takes effect after /clear or the next session.
To skip the picker, add outputStyle to ~/.claude/settings.json. It is a
top-level key. It does not go inside env, permissions, or any other block:
{
"model": "opus",
"env": { "EXAMPLE_VAR": "1" },
"outputStyle": "Attention Control"
}
model and env stand in for keys you may already have. Keep them. Add the
outputStyle line beside them.
For one session instead of every session, use the skill the plugin also ships:
/attention-control:attention-control
Say "stop attention control" to turn it off.
Codex has no output style slot, so the rules ship as a skill.
codex plugin marketplace add aaddrick/attention-control --ref main
codex plugin add attention-control@attention-control
Inside Codex, /plugins opens the plugin browser instead.
Start a new thread, then type the skill:
$attention-control:attention-control
Codex prefixes a plugin skill with the plugin name. Say "stop attention control"
to turn it off. For every turn instead, put the always-on snippet from
INSTALL.md in ~/.codex/AGENTS.md.
See INSTALL.md. None of them has an output style slot, so the
rules ship as a skill, a rules file, or an AGENTS.md block. On a skill route,
type /attention-control and say "stop attention control" to turn it off.
Air traffic control phraseology exists for one reason: a distracted human under load will mishear an instruction. Aviation solved this with two disciplines. Controlled vocabulary makes each word mean one thing. Fixed message shape puts the instruction first and the background last.
This style applies both disciplines to your coding agent. The agent leads with the action you can run, and writes each sentence so one word means one thing.
The style targets one reader: a reader with ADHD. That reader is where the shape rules come from. See Why the shape rules exist.
Before
|
After
|
Shape decides what you say and in what order. 11 rules:
Language decides how each sentence reads. The short version:
Full text: output-styles/attention-control.md.
Five facts about ADHD reading drive all 11 shape rules. Each fact below names the rules it produces.
| The fact | What the agent does |
|---|---|
| Working memory is small. Anything not on screen is gone. | It never writes "keep in mind X". It restates the state every turn: "Step 3 of 5 done: I changed the schema. Next: run scripts/backfill.py." (rules 6, 10) |
| Knowing the answer is not doing the answer. Work dies in the gap between the two. | It does the work it owns instead of handing it back. It gives the command, not the label. "Add the missing header" is a label. Authorization: Bearer ${token} is a fix. (rules 1, 2, 3) |
| Starting is the hardest step. | The first line is small, obvious, and doable now. The last line names one action that takes under two minutes. "Open the file" counts. (rules 1, 4) |
| Time estimates feel uniform. "A bit of work" and "a few hours" register the same. | It writes "about 15 minutes if tests cover this, an afternoon if not". It never writes "some work". (rule 7) |
| Dopamine is scarce. A buried win does not register. | After a change, it names the result in concrete terms: "Login works with magic links. Run npm run dev and open /login." (rule 8) |
Two more rules protect attention itself. Rule 5 suppresses tangents, so one open thread stays one open thread. Rule 11 removes the preamble and the closer, so the answer starts on line 1.
This is why the style is not "be terse". Terseness that drops the command, the number, or the condition costs the reader a round trip, and a round trip costs the thread. Rule 9 follows from the same logic: an error gets a location, a cause, and a fix, with no "Uh oh" in front of it. Alarm is not information, and it competes with the information for the same attention.
You need no ADHD diagnosis for this to help. A tired reader, a reader on a phone, and a reader with 40 open tabs all read the same way.
Four targets, one rule each. Code, commands, file paths, identifiers, and error messages stay verbatim. Quoted text stays verbatim. Code comments and commit messages match the style of the repository around them. Only the prose the agent writes itself follows this style.
Those are instructions, not a lock. An output style is system prompt text, and nothing outside the model enforces it. Treat the boundary as a strong default, and check the output when it matters.
Accuracy beats brevity. A rule never removes a fact, a number, a condition, or a scope qualifier. A hedge that carries real uncertainty stays.
The harness compares response quality against an unstyled baseline. It does not measure length.
python3 scripts/run_evals.py validate
python3 scripts/run_evals.py plan --trials 3
24 cases, 6 scored dimensions, and a release gate that blocks a candidate that regresses correctness or safety.
The judge is the weak point, so the harness targets it. blind hides the
condition and balances the positions. The judge scores every group twice with
the order reversed, then reports how often the two passes disagree. The runner
runs in an empty directory and reads none of your config. Design notes and the
measurements behind them: evals/README.md.
Fork, edit output-styles/attention-control.md, then regenerate every
agent-specific copy:
python3 scripts/sync_style.py
Swap your copy in, one command at a time:
claude plugin uninstall attention-control
claude plugin marketplace remove attention-control
claude plugin marketplace add <your-username>/attention-control
claude plugin install attention-control@attention-control
This style combines two existing works. Neither author takes part in this project.
Shape layer: i-have-adhd by
Ayoub G. (MIT). The eval harness derives from the same project.
Language layer: the
asd-ste100 output style
by L1nefeed, itself a condensation of
ASD-STE100 Simplified Technical English, Issue 9.
This project reproduces no text from the ASD specification. The ASD does not certify, endorse, or sponsor it. Details in NOTICE.md.
MIT. See LICENSE.
Sign in to join the discussion.
No comments yet. Be the first to say what this is good for.
All parts of Claude Code's system prompt, 27 builtin tool descriptions, sub agent prompts (Plan/Explore/Task), utility prompts (CLAUDE.md, compact, statusline, magic docs, WebFetch, Bash cmd, security review, agent creation). Updated for each Claude Code version.
HERO = Hashing · Edge cases · Rubrics · Overbuild — the four shapes coding agents over-defend in. A paste-in contract that stops them. Works with Claude Code, Codex, Antigravity, Cursor, Copilot, Windsurf, Gemini CLI.
Make Claude talk like a human: 17 installable output styles for Claude Code, built on credited methodologies
Stop agents from over-engineering. Inspired by Ponytail, but without the plugin madness. Just one AGENTS.md file.
Independently authored prompt templates for AI coding agents — system prompts, tool prompts, agent delegation, memory management, and multi-agent coordination. Informed by studying Claude Code.
A collection of production-ready slash commands for Claude Code