ADHD & 47 Tabs

An adaptive AI skill for clearer answers, smaller working sets, interruption recovery, and reliable finishing.

Your brain has enough tabs open. Your AI does not need to add twelve more.

ADHD & 47 Tabs makes the useful part of an AI response easier to find, start, resume, and finish. It is not a medical tool and does not diagnose or treat ADHD. No diagnosis is required.

Version 3 moves beyond “be shorter.” It preserves necessary depth while adapting the response to the kind of work and the friction in the moment.

What v3 does

Four base contracts

Need	Response begins with
Answer	The conclusion, result, or recommendation
Action	The smallest meaningful action
Artifact	The finished email, code, plan, prompt, table, or checklist
Project update	The last verified state and one active next item

Six adaptive modifiers

• Friction: shrinks the active scope when starting feels impossible.
• Reorientation: restores the thread with You are here: goal → state → next.
• Memory offload: reuses dates, names, constraints, paths, and decisions already provided.
• Decision: recommends one path and states the deciding criterion.
• Recovery: stops speculative patch loops after repeated failure and runs one discriminating diagnostic.
• Finish: protects the definition of done from optional scope growth.

The goal remains lower cognitive load, not minimum word count. Detailed requests still receive detailed answers. High-stakes questions retain the nuance and escalation guidance they need. Creative work stays creative. Emotional support is not turned into a productivity checklist.

Small working set, complete output

For complex work, the skill keeps:

• Active: one item;
• Ready: up to two;
• Blocked: only what affects the active path;
• Parked: captured but normally hidden.

That limit applies to the working set, not the deliverable. A requested 30-item checklist still contains all 30 items, grouped so it is navigable.

Conversation controls

Natural language works, but these short phrases are convenient:

Say	Result
one thing	Only the current action and stopping condition
map it	Compact route, dependencies, definition of done, and first action
resume	Reorientation breadcrumb and continuation
park that	Capture the tangent without replacing the current goal
more detail / less detail	Change depth without losing the conclusion or safety
why this	Explain the deciding reason
normal mode / stop 47-tabs mode / stop ADHD mode	Disable the defaults until asked to resume

Download and verify

Download the canonical adhd-and-47-tabs.zip.

Verify it against dist/SHA256SUMS:

(cd dist && sha256sum -c SHA256SUMS)

The ZIP is rebuilt deterministically from the source folder during make check. It contains Markdown instructions and references, one YAML host-metadata file, and the MIT license—no executable code, packages, credentials, telemetry, or network instructions.

Install

Claude

Claude custom skills are available on Free, Pro, Max, Team, and Enterprise plans when code execution is enabled.

1. Download the ZIP.
2. Enable Code execution and file creation in Settings → Capabilities. Team and Enterprise workspaces may use organization settings.
3. Open Customize → Skills.
4. Select + → Create skill → Upload a skill.
5. Upload and enable the ZIP.

Official guide: Use skills in Claude.

ChatGPT

1. Download the ZIP.
2. Open ChatGPT's Skills page. In the current web interface, this is under Plugins → Skills in the sidebar.
3. Select Create → Upload from your computer.
4. Upload the ZIP.

Workspace administrators may control skill creation, uploading, installation, and sharing. Personal skills may need to be installed separately on different ChatGPT surfaces.

Official guide: Skills in ChatGPT.

For accounts without uploaded Skills, copy chatgpt-custom-gpt/INSTRUCTIONS.md into a Custom GPT named ADHD & 47 Tabs.

Codex

Inside Codex, invoke the built-in installer:

$skill-installer

Then ask it to install the adhd-and-47-tabs skill from this repository's adhd-and-47-tabs folder. For a manual user-wide install, place that folder at ~/.agents/skills/adhd-and-47-tabs. For a repository-scoped install, place it at .agents/skills/adhd-and-47-tabs inside the repository. Codex detects skill changes automatically; restart only when an install or update does not appear.

The package includes agents/openai.yaml so ChatGPT desktop and Codex can show a clean display name, description, default prompt, and implicit-invocation policy.

Official guide: Build skills for ChatGPT and Codex.

GitHub Copilot

GitHub CLI 2.90.0 or later can preview and install Agent Skills:

gh skill preview zgbrenner/adhd-and-47-tabs adhd-and-47-tabs
gh skill install zgbrenner/adhd-and-47-tabs adhd-and-47-tabs --scope user

You can also copy the skill folder into ~/.copilot/skills, ~/.agents/skills, .github/skills, or .agents/skills, depending on scope.

Official guide: Adding Agent Skills for GitHub Copilot.

Universal Skills CLI

npx skills add zgbrenner/adhd-and-47-tabs --skill adhd-and-47-tabs

This is a third-party CLI. Its documentation says anonymous skill-usage telemetry is enabled by default; set DISABLE_TELEMETRY=1 before the command to opt out, or use one of the manual installation paths above.

Upgrade from v2

The slug is unchanged. Replace the installed ZIP or skill folder with the latest v3 release, then restart or reload the host if it caches skills.

Version 3 changes behavior and the repository evaluation schema, not the package identity. Version 1 used the retired slug i-have-adhd-and-47-tabs; remove that copy so two versions do not load together.

Read the skill

• adhd-and-47-tabs/SKILL.md — always-loaded behavioral core.
• references/interaction-patterns.md — routing, state, controls, and recovery patterns.
• references/examples.md — 32 examples and edge cases.
• references/quick-reference.md — compact host checklist.

The structure follows the Agent Skills progressive-disclosure model: concise metadata and core instructions first, focused references on demand.

Evaluate and validate

The repository includes 47 provider-neutral cases, including nine multi-turn scenarios, and a dependency-free scorer:

python3 scripts/score_responses.py --responses responses.jsonl
python3 scripts/score_responses.py --responses responses.jsonl --format json --output report.json
python3 scripts/score_responses.py --responses responses.jsonl --format markdown --output report.md
make check
make release
make install-hooks

The scorer checks structural regressions such as generic preambles and closers, forced next steps, excessive active lists, missing orientation breadcrumbs, ordered recovery sections, and requested breadth, list, paragraph, and depth preservation. Each case also includes a human-review focus because semantic quality cannot be reduced to regexes.

See docs/EVALUATION.md.

This repository intentionally uses no GitHub Actions or hosted CI. All release checks run locally with Python's standard library and Make.

Research and design

Version 3 synthesizes accessibility guidance and open-source patterns from W3C COGA, the Agent Skills specification, the upstream skill, Leantime, Taskwarrior, Super Productivity, ActivityWatch, Beads, Claude ADHD Skills, Promptfoo, OpenAI Evals, and Inspect AI.

The project adopts principles—not copied prose or code—and records what was deliberately rejected. See docs/RESEARCH.md.

Contributing

Behavior changes should include a regression case. Run make check, regenerate the tracked ZIP and checksum, and explain how the change lowers friction without deleting necessary substance.

See CONTRIBUTING.md, SUPPORT.md, and SECURITY.md.

Attribution and license

Original concept and skill: Ayoub Ghriss, ayghri/i-have-adhd.

Cross-platform adaptation: Zachary Brenner.

MIT licensed. The upstream notice is preserved in LICENSE, and adaptation history is documented in NOTICE.md.