thanos

/thanos

A prompt architect for Claude. Give it any raw prompt — a one-liner or a full PRD — and it hands back an optimized, paste-ready prompt with the right agent loop, the skills to invoke, and the model to run it on.

It never runs your task. It designs the prompt that will.

Why

Most prompts fail for two reasons:

1. There is no definition of done.
2. There is no way for the agent to check that it is done.

/thanos fixes both — every prompt it returns carries a testable stop condition and a machine-checkable acceptance gate.

Examples

Goal loop — you know what done looks like:

/thanos get my failing tests green

/goal make `npm test` exit 0, stop after 4 tries

Fix one failure per iteration. The test runner's exit code
is the only source of truth — do not self-assess.

Do not: delete tests, weaken assertions, touch unrelated files.

Time loops — recurring work, watching something external:

/thanos keep an eye on my open PRs

/loop 30m check my open PRs: apply reviewer feedback and reply —
stop when everything is merged

/schedule every morning at 9: summarize overnight issues into one
message — skip if there are none

…plus one line naming the model to run it on.

The four loops

A loop is an agent repeating cycles of work until a stop condition is met.

/thanos always picks the simplest loop that fits. You only see its choice on the first line of the prompt.

What you get back

The prompt, and almost nothing else:

1. The prompt — one paste-ready block. The loop command goes on line one, with skills named inline, machine-checkable acceptance criteria, and a Do not: list. When a job needs different models for different steps, the switches are written into the prompt.
2. Model — one line.
3. Heads-up — one line, only when something needs your yes or won't fit one run.

There are no diagnosis tables and no rejected alternatives to read through.

Built-in guardrails, enforced by a self-check before every answer:

• Fits /goal — Claude Code rejects goals over 4,000 characters, so every /goal block is measured and trimmed to fit, without dropping a deliverable.
• Grounded gates — it never invents a check against a file it hasn't opened.
• Full coverage — every deliverable you asked for gets an acceptance criterion.
• Risk scan — publishing data or handling API keys pulls in a security skill automatically.
• No stalled loops — anything that needs your yes gets its own second prompt, never buried inside /goal.

Installation

Claude Code — as a plugin

claude plugin marketplace add https://github.com/thanospapage/thanos

claude plugin install thanos@thanos

Claude Code — as a plain skill

git clone https://github.com/thanospapage/thanos.git

cp -r thanos/skills/thanos ~/.claude/skills/

Restart Claude Code, then type /thanos.

Claude.ai and Cowork

1. Clone or download this repo.
2. Zip the skill folder:

cd skills && zip -r thanos.zip thanos/

3. In Claude.ai open Settings → Skills → Add → Upload a skill.
4. Drop in thanos.zip. Cowork reads the same skill library.

Other LLMs and agents

The skill is a single markdown file — skills/thanos/SKILL.md — so any tool that accepts custom instructions can run it. One caveat: /goal, /loop, and /schedule are Claude Code commands; other tools still get the full diagnosis and optimized prompt, and should treat the loop line as the stop condition to enforce.

Codex CLI — append the file to your global or project instructions:

cat skills/thanos/SKILL.md >> ~/.codex/AGENTS.md

Then start a message with /thanos <your prompt>.

Cursor — add it as a rule, project-level or global:

mkdir -p .cursor/rules && cp skills/thanos/SKILL.md .cursor/rules/thanos.mdc

Or paste the file into Settings → Rules → User Rules to have it in every project.

Gemini CLI — append it to your context file:

cat skills/thanos/SKILL.md >> ~/.gemini/GEMINI.md

Anything else (ChatGPT, Hermes, local models) — paste the contents of SKILL.md as the system prompt or custom instructions, then send your prompts as /thanos <your prompt>.

Usage

/thanos <anything you want to ask Claude>

Also triggers on "optimize this prompt", "what loop does this need", "turn this PRD into a prompt". It won't trigger when you just want the task done, or when you ask to optimize code — those are engineering jobs, not prompt jobs.

Keeping costs down

• Every loop needs a stop condition — one without it is a bug.
• Pilot on a small slice before a big run.
• Script deterministic steps instead of re-reasoning them each cycle.
• Don't run a routine more often than the thing it watches changes.
• Send repetitive work to cheaper models; save the big one for judgment.

If /thanos saved you a broken loop, a ⭐ helps others find it.

References

• Loop taxonomy, from the Claude Code team: https://x.com/claudedevs/status/2074208949205881033
• Pipeline structure adapted from the prompt-optimizer skill in ECC (MIT)

License

MIT