AI-assisted engineering, with evidence
Build features with agents. Keep engineering in control.
PromptGrinder turns AI coding prompts into deterministic, reviewable delivery trains: smaller slices, explicit ownership, durable Git checkpoints, safe recovery, and isolated parallel worktrees.
New here? Read From feature idea to a safe AI delivery train.
Start with the repository
Install PromptGrinder, then let it inspect the project before defining work:
promptgrinder discover
promptgrinder roles enhance
promptgrinder roles review latest
The resulting project and role evidence gives workers useful context while keeping their ownership boundaries explicit.
Support at a glance
| Surface | Current status |
|---|---|
| macOS Apple silicon and Intel | Supported release-candidate surface |
| Headless, Terminal.app, and iTerm2 | Supported execution surfaces |
| Codex CLI | Supported for run, run-folder, and named workers |
| Antigravity | Supported for named workers through its agy adapter |
| Models | Selected from the signed-in runtime’s live catalog and bounded by repository policy |
| MCP servers | Configured and owned by the selected engine; PromptGrinder-native management is future work |
| Linux, Windows, other terminals, and other engines | Not qualified or supported yet |
Codex CLI 0.150.x is the known-qualified band. Other versions can run as
provisional when they pass PromptGrinder’s non-mutating adapter capability
probe; a missing required capability blocks the train before worker creation.
Then run a train
promptgrinder run-folder docs/features/FB-XXX \
--repo . \
--checkpoint \
--commit-each \
--require-clean-git
Read the installation guide, explore the source repository, or start with the full guide.