Clear plans. Reviewed results. You set the direction.
Workflow helps your coding agent plan a change, implement it, review the result, and fix the issues it finds. It works with Cursor, Codex, OpenCode, and compatible coding agents.
- Agree on the work first. Clarify the goal, important decisions, and how to check the result.
- Understand what you get. See what changed, what was checked, and what still needs attention.
- Control the next step. Request planning, implementation, review, and correction separately.
Use the latest stable GitHub Release for installation and updates. Choose manual installation or let your own agent handle it. No repository checkout, Node.js, or npm is needed.
Manually: download the release archive for Cursor, Codex, or OpenCode together with SHA256SUMS and provenance.json, verify the files, and follow the manual installation steps. For updates, follow the manual update steps to replace the complete package. Rollback is installing an older GitHub release.
With your agent: for a first installation, paste this into Cursor, Codex, or OpenCode in a mode that can make changes:
Install the latest stable Workflow release from https://github.com/geldmacher/workflow/releases/latest for my current host. Read the install-release skill and its linked installation instructions from the matching release tag, then perform the installation.
Your agent needs GitHub access and permission to install the plugin. The selected release must contain install-release.
Update with your agent: use the included install-release skill:
- Cursor:
/install-release Update Workflow to the latest stable release for Cursor. - Codex:
$install-release Update Workflow to the latest stable release for Codex. - OpenCode:
@install-release Update Workflow to the latest stable release for OpenCode.
Follow the reported activation steps: reload Cursor; in Codex, restart the app and install or refresh Workflow in the Plugins Directory; in OpenCode, start a new session. See the installation guide for details and troubleshooting.
The recommended starting point is plan → implement → review → correct if needed → review again. Open a project and request each phase separately.
1. Plan the change. Ask:
Use plan-work to add CSV export to the orders page. Export only the currently filtered orders, using the visible columns. Empty exports should contain headers. Follow the existing download conventions.
Ordinary language works. To select the skill directly, use $plan-work in Codex, /plan-work in Cursor, @plan-work or /plan-work in OpenCode, or plan-work in portable clients. Use the same prefix for the skills below.
2. Approve and implement. Check the proposed behavior, scope, and checks, then select Implement Plan in Cursor or Codex. That button stays the default when Plan mode produced the plan. When Plan mode is not used, ask /implement-work in Cursor or $implement-work in Codex and supply the plan artefact. In OpenCode, switch to the build agent and ask /implement-work or @implement-work with that plan. You receive the changes, checks, and any open issues.
3. Request review. After implementation, ask:
Use review-work to review the CSV export against the approved plan and implementation report.
4. Correct if needed. If review finds an issue, ask:
Use correct-work to fix the CSV export findings from this review and run the affected checks.
Then request review-work again. A positive review with the required proof completes the repository work. Publishing or deploying needs an explicit request.
For a small, obvious edit, a direct request to your agent may be enough. For the full sequence, a bug-fix example, and tips on reusing checks, follow the working guide.
- Set up or maintain project feedback with clear project purpose, useful checks and agent guidance.
- Save useful lessons in project guidance after review, when you request it. Learning
- Keep checks repeatable with documented steps for testing real behavior. Verification
- Choose a fitting method for a bug fix, feature, or performance problem. Playbooks
Working guide · Development and releases · Latest release
Experimental: Auto-Work. An optional approach connects the phases through native agents. Its Light/Dark modes, requirements, and limits are explained in the separate guide.