Shared workflows: the cross-tool hub
The Workflows section holds the practices that stay the same whether the agent is Claude Code, Codex or Cursor: managing context, building the harness, specifying, building, testing, shipping, operating and automating. Each subgroup covers one lifecycle stage. Tool-specific mechanics, such as commands and settings files, live in each tool’s own section.
This hub is for developers and tech leads who already have an agent running and now need it to produce work they can trust. Maybe the agent writes code quickly but reviews pile up, or the same mistake returns every week. Find the stage where your work stalls, and open that subgroup.
Which workflow group do you need?
Section titled “Which workflow group do you need?”New here? Start with the AI-native lifecycle, which shows the whole loop from intent to production, and the method, which covers the habits that hold across every stage, such as TDD, human checkpoints and prompting. Then read the second column below as the question you are asking today. Each subgroup’s overview page lists its guides.
| Subgroup | Open it when you are asking… |
|---|---|
| Context | “Why does the agent ignore my conventions or lose track in a large repository?” |
| Harness | “Which permissions, hooks, skills and sandboxes should bound the agent?” |
| Plan and specify | “How do I write a spec and acceptance criteria the agent cannot misread?” |
| Build | “How do I run agents on monorepos, legacy code, APIs and databases?” |
| Test and verify | “How do I trust tests I did not read, and catch plausible but wrong code?” |
| Review and ship | “What must an agent’s pull request prove before it merges, and how do I review it without reading every line?” |
| Operate and learn | “How do I observe agents in production, respond to incidents and keep the codebase healthy?” |
| Automate | “How do I go from an issue to a pull request, or from intent to production, with nobody at the keyboard?” |
Automation depends on the stages above it. An unattended agent is only as safe as the tests and gates that check its output, so fix Test and verify and Review and ship before you automate.
Where do Claude Code, Codex and Cursor differ?
Section titled “Where do Claude Code, Codex and Cursor differ?”The practice is the same in all three tools, and all three support hooks, skills, unattended runs and worktrees. What differs is the file names and commands. Each tab below fills the same four slots. The tool map compares every stage side by side.
- Instructions:
CLAUDE.md. - Deterministic checks: hooks, with repeatable procedures in skills.
- Unattended runs:
claude -p. - Parallel work:
claude --worktree.
Flags checked against Claude Code 2.1.283. See the Claude Code section.
- Instructions:
AGENTS.md. - Deterministic checks: hooks, with repeatable procedures in skills.
- Unattended runs:
codex exec. - Parallel work:
codex --worktree, or cloud environments.
Flags checked against Codex CLI 0.157.1. See the Codex section.
- Instructions: Rules.
- Deterministic checks: Hooks, with repeatable procedures in Agent Skills.
- Unattended runs: the Cursor CLI in print mode (
-p). - Parallel work: Worktrees, or Cloud Agents.
Feature names checked against Cursor’s docs on 2026-08-28. See the Cursor section.
How do you know a workflow worked?
Section titled “How do you know a workflow worked?”A workflow is done when it produces evidence that a machine checks. Your feeling that the diff looks right is not evidence. Across every subgroup, the evidence falls into three layers:
- The oracle: tests and acceptance criteria that the agent cannot edit. See how strong your oracle is and how to protect it.
- The gates: type checks, lint, fitness functions and a review agent that run in CI on every change.
- The bundle: each pull request states what changed and why, and shows its proof in an evidence bundle. A human signs off on the evidence, not on every line. See how to review an agent’s pull request.
Paste either prompt below into any of the three tools, at the root of your repository.
What goes wrong when teams pick workflows?
Section titled “What goes wrong when teams pick workflows?”- Automating before verifying. A team jumps to issue-to-PR automation, then drowns in pull requests nobody trusts. To recover, pause the automation, and run the audit prompt above. Then close the test and ship gaps it reports before you turn automation back on.
- Fixing every mistake with another rule. The instruction file grows past the point where anyone knows which lines still matter. To recover, move repeatable checks into hooks or CI, as the harness section describes. Then prune the file.
- Reading every diff because nothing else is trusted. Review becomes the bottleneck, and speed gains disappear. To recover, add an evidence bundle to your pull request template, and require it before anyone reviews.
- Copying a tool-specific recipe into another tool. A command or settings key from one tool does nothing in another. To recover, keep the practice from this section, and take the mechanics from that tool’s own section.