Zeroshot

zeroshot

Write your QA process down as a graph. zeroshot makes every agent run follow it.

zeroshot.sh · Discord · Docs · Blog

 

Single-agent sessions degrade. Context gets buried under thousands of tokens, and the model optimizes
for "done" over "correct". In our study, all 100
sessions told to check their own work stopped after one pass.

zeroshot fixes this with independent agents that check each other's work, following a graph you
define for your repo. Reviewers start in a fresh session, can run a different model, and aren't
grading their own work. Fail the check? The reason goes back to the writer, and the fix goes through
every review again until it actually works.

The graph is built around your repo and your QA requirements, not a fixed pipeline.

See the difference on zeroshot.sh: a flight game and
a landing page, each built by an agent alone and again through
a zeroshot graph.

Who is this for?

- Senior engineers who care about correctness and reproducibility
- Teams automating PR workflows and code review gates
- Infra/platform teams standardizing agentic workflows
- Open-source maintainers working through issue backlogs
- AI power users who want verification, not vibes

Get started

Install it once:

This also installs the zeroshot skills for Claude Code, Codex, Copilot CLI and Pi. Open one of them
in your repo and ask:

It reads your build, test and CI setup, asks how you want changes delivered, and shows you the graph
and every agent's instructions before saving anything. Then it asks which models to use and saves the
graph, as your default if you want.

Then plan the work before you hand it over:

Your agent reads the code, asks only what changes the outcome, and writes a plan where every
acceptance criterion comes with a check: a test, a…