Zimster

A practical workflow for coding agents.

Zimster gives an agent clear steps, useful tests, and review at the points that matter. One implementation owner stays in charge. Extra agents are used when there is a reason, not as a default.

Build software with AI without turning every task into a committee.

Zimster keeps the workflow focused: plan when planning is useful, write tests that prove the behavior, review the risky parts, and only call the work complete when the evidence supports it.

See how the workflow is structured →

Choose the package that fits your setup.

The project is intended for more than one coding-agent environment. The commands differ, but the workflow stays the same.

Codex

Add the public repository as a marketplace source, install the plugin, and confirm it appears in the installed list.

codex plugin marketplace add https://github.com/wazimmerman/zimster.git
codex plugin add zimster@zimster --json
codex plugin list
Codex install notes →

Claude Code

Use the Claude-oriented package surface and the install flow documented for the current release.

/plugin marketplace add wazimmerman/zimster
/plugin install zimster@zimster
/reload-plugins
Claude Code install notes →

Portable bundle

Use the portable ZIP when you want the skills and supporting files without starting from a harness-specific package.

Download the portable ZIP
Extract it into your plugin or skills location
Follow the notes for your coding agent
Portable install notes →

Less ceremony. Clearer responsibility.

The plugin keeps useful engineering practices and trims back the parts that tend to make agent runs slow, repetitive, or hard to follow.

One owner stays in charge

The main agent keeps the implementation coherent instead of splitting every task across a growing tree of agents.

Tests have to prove something

RED-GREEN-REFACTOR is used as evidence, not as a label attached after the code already works.

Delegation is a choice

Zimster can use other agents, but only when the work is bounded and the split makes the result better.

Model choice comes later

A cheaper model does not create a reason to delegate. Model routing is considered only after delegation already makes sense.

Review follows the risk

Small work stays small. Public interfaces, migrations, security, and other load-bearing changes get a higher proof bar.

Completion states mean different things

Self-review, checkout integrity, independent approval, and final completion are tracked separately.

The work moves in a straight line.

  1. Classify the task

    Set the risk level, the Git disposition, and the minimum proof required.

  2. Design and plan when needed

    Use a plan for real choices and multi-step work. Skip the extra ceremony when the task is already clear.

  3. Build with one owner

    The implementation owner writes the tests, code, and corrections while keeping the whole change in view.

  4. Delegate bounded work

    Send out a scout, reviewer, or implementation slice only when the boundary is useful and easy to verify.

  5. Verify and review

    Run the affected checks, review the risky seam, and keep the approval tied to the exact candidate.

  6. Finish with evidence

    Map the requirements to proof and report only the completion state the evidence supports.

The core set

Twelve skills cover the workflow without turning it into a catalog of overlapping personas.

Using Zimster

Selects the smallest useful workflow for the task.

Designing Work

Clarifies consequential choices before coding.

Writing Plans

Turns larger work into coherent, testable slices.

Owner-Driven Development

Keeps one implementation owner responsible for the change.

Test-Driven Development

Uses observable RED-GREEN-REFACTOR.

Parallel Agents

Delegates only genuinely independent workstreams.

Risk-Adaptive Review

Matches review depth to the risk in the change.

Receiving Code Review

Checks feedback before changing the implementation.

Systematic Debugging

Reproduces the problem and proves the root-cause fix.

Verification Before Completion

Requires fresh evidence before a success claim.

Git Worktrees

Creates a clean place for risky changes when needed.

Finishing a Development Branch

Verifies the result and hands off a clean branch.

What has been tested

The labels below show how far v0.6.0 has been tested in each environment. Some packages have been run directly; others have only had their structure and packaging checked.

Coding agentTesting statusWhat was checked
CodexInstalled-package verifiedThe package and install surface were checked in Codex. This is narrower than claiming every workflow path was exercised live.
OpenCodeLive verifiedThe current release includes live verification in OpenCode.
Claude CodeStructurally validatedThe package shape and related files were checked; hands-on verification is still narrower.
CursorStructurally validatedThe adapter and package surface were checked; marketplace packaging remains a separate step.
Kimi CodeStructurally validatedThe package surface was checked, with less live evidence than Codex or OpenCode.
PiStructurally validatedThe package declaration and adapter surface were checked; ecosystem distribution work remains.

Release files

All downloads come directly from the v0.6.0 GitHub release.

ZIP

Codex package

Installable package for the Codex plugin surface.

Download →
ZIP

Claude package

Release package for the Claude-oriented install path.

Download →
ZIP

Portable bundle

Skills-first package for more general use.

Download →
TXT

SHA-256 checksums

Use this list to verify the release files.

Download →

Found a problem?

Open an issue with the Zimster version, your coding agent, and the steps needed to reproduce it.

Open GitHub Issues