Local. Open source. Built by Lucio.

Writ Workflow governance for Claude Code.

Approve the plan. Review the tests. Keep the reasons behind the code.

MIT licensed Runs on your machine

Inside a real Claude Code session1:21
Actual screen capture. Edited for length. Demo notes

The workflow

Two approvals before implementation.

See the full sequence
  1. 01

    Plan

    You review the scope and files.

    Your approval
  2. 02

    Tests

    You review the test skeletons.

    Your approval
  3. 03

    Implement

    Supported source changes can proceed.

    Ready to build

Work mode has documented path exemptions. An approved plan change re-arms the gates.

RulesProject docsPast decisions
Current task

Context

The rules that fit the work.

Writ retrieves relevant project context and adds file-specific guidance at supported write points.

Recall
DecisionCommitChanged files

Continuity

Recall the reasons behind a change.

Recall brings selected decisions and recorded file-change reasons into later sessions.

Acknowledgements

Ideas Writ builds on

Superpowers

Jesse Vincent's Superpowers formalizes the development discipline Writ builds enforcement around: planning, testing and reviewing the work.

Jolli

JolliAI's work on preserving development reasoning helped shape Writ's decision memory. Recall's digest eviction policy is adapted from Jolli's ContextCompiler policy.

Writ brings these influences together with its own context retrieval, approval checks and decision records. Read the source acknowledgements.

Before you install

A local runtime.
Your own setup.

Linux or macOSPython 3.11+Docker + Neo4jClaude Code
Installation guide →
What do the gates actually check?

In Work mode, non-exempt source changes wait for your approval of a plan and test skeletons. The checks validate presence and structure. You judge whether the plan and tests are good. Adding a written rule does not automatically add a blocking check.

What does Recall remember?

Selected project decisions and recorded reasons behind file changes. It is not a complete conversation recording. Capture depends on the configured hooks and local service.

What are the security limits?

Writ constrains a cooperative agent. It is not a security sandbox. Some checks refuse actions, some ask and others report. Shell inspection has documented gaps. Read the security model.

Who builds Writ?

Lucio Saldivar at Infinri. Writ is free and MIT licensed. Ask a question or report a bug on GitHub.