Everything AI

Everything AI is an agent skill for people who want AI to do everything.

User gives goal. AI carries expert scope.

You say do everything, handle it end-to-end, or whatever is needed. The promise: the AI figures out the expert checklist, chooses safe defaults, starts useful work, and reports proof.

There is no mode menu and no expert questionnaire. Questions appear only when a real blocker or risky action needs you.

What It Does

• Infers missing scope.
• Uses safe, reversible defaults.
• Works before asking non-blocking questions.
• Stops before paid, destructive, private, or high-stakes actions.
• Reports checked, changed, missing, unknown, and confidence.
• Supports 10 domain packs: coding, startup, research, finance, health, learning, writing, data, life, and productivity.

Install

npx --yes github:mitunmanav/everything-ai

Check first without installing:

npx --yes github:mitunmanav/everything-ai -- --dry-run

Then ask:

Use $everything-ai and do everything for this task.

The installer copies only the skill, sends no telemetry, reads no secrets, and refuses to overwrite an existing install unless you use --force.

More install options: QUICKSTART.md.

Proof

Two blind judges agree: skill helps current outputs
_{Same 80 outputs. Claude Sonnet 5 and Codex gpt-5.5 scored independently. Higher is better.}

output model	Claude blind judge	lift	Codex blind judge	lift
gpt-5.5 · medium	47.4% → 96.1%	+48.7 points	52.6% → 96.1%	+43.5 points
gpt-5.4-mini · low	46.1% → 86.8%	+40.7 points	52.6% → 89.5%	+36.9 points

This is a single-run current-harness result. It is not comparable with v0.4.0 as a release-to-release improvement because the generated outputs and judge models differ.

Full method, older results, raw score files, and caveats: TEST_RESULTS.md and EVALUATION.md.

Safety

Everything AI promises a process, not impossible outcomes.

• No destructive, paid, irreversible, or high-stakes action without approval.
• No secrets, local paths, raw private transcripts, or development rules in the public package.
• Pull requests must pass tests and CodeQL.
• GitHub releases stay manual.

Improve It

Real failed prompts make the skill better:

1. Open an issue with the prompt and what went wrong.
2. Add it to the prompt bank.
3. Turn it into a test, benchmark case, or domain example.
4. Fix the narrow behavior and show proof.

Contribution guide: CONTRIBUTING.md.

Details

• Quick start
• Test results
• Evaluation method
• Roadmap
• Security
• Release checklist

MIT licensed. Built and maintained by Mitun Manav.