Choose a primary AI engineering harness

A primary AI engineering harness is the tool and repository configuration you can use to move from intent to a verified diff. Claude Code, Cursor, and Codex can all earn the maximum Q1 score. The deciding evidence is versioned context, an explicit plan, bounded tools and data, isolated execution where needed, reproducible tests, and a diff a human can approve.

Scorecard question: How does your primary AI coding setup handle repository work?

Maximum-score answer: The harness uses versioned instructions, plan and artifact gates, bounded tools, and reproducible verification evidence.

Compare the complete workflow

Do not score a logo, model name, or terminal against an editor. Score whether the configured surface supports the job:

Capability	Evidence to collect
Repository understanding	It reads real files, commands, tests, and conventions before proposing work
Durable policy	The repository contains reviewed CLAUDE.md, .cursor/rules/, or AGENTS.md instructions
Planning	A non-editing phase produces plan.md with risks and verification
Controlled execution	Tools, network, credentials, and write access match the task’s risk
Isolation	Parallel or high-risk work uses an appropriate worktree or sandbox
Verification	The session runs exact repository gates and retains output
Review	A human can trace the diff to intent.md, spec.md, and plan.md

Use the cross-tool capability map for verified Claude Code, Cursor, and Codex controls. Choose the surface that fits the team’s environment and constraints; the lifecycle contract stays the same.

Evaluate one representative task

1. Select a bounded change. Use a task that touches several files, has an observable outcome, and does not require production authority.
2. Capture intent and constraints. Write intent.md, name the owner, and record what is out of scope.
3. Configure repository policy. Add or review the primary tool’s instruction file and exact verification commands.
4. Plan without edits. Require file-level steps, risks, rollback, and acceptance evidence in plan.md.
5. Execute inside the boundary. Use only the approved files, tools, credentials, worktree, and network access.
6. Run verification. Retain test output, the final diff, and any unresolved limitation.
7. Score the result. Measure accepted outcome, corrections, review time, elapsed time, and completed-task cost.

Prompts to copy

Read the repository instructions and intent.md. Do not edit yet.
Return a plan with exact files, assumptions, risks, rollback, required tests,
and the evidence a reviewer should expect. Stop for approval after the plan.

Execute the accepted plan within the approved file and tool boundaries.
Run the exact repository checks, inspect the final diff against plan.md,
and report commands, outcomes, limitations, and any required human decision.

Failure modes

The score rewards a product category. Cursor Agent can satisfy the same lifecycle contract as a terminal surface; assess evidence, not interface.

The agent edits before understanding the repo. Start with read-only exploration and an accepted plan.

A passing test hides a wrong outcome. Trace verification to acceptance criteria and inspect the real user or system behavior.

The setup needs broad credentials. Reduce scope or choose a surface whose isolation and approval model fits the task.

Verify the primary setup

• Repository instructions and exact checks are versioned.
• Non-trivial work begins with intent, specification, and an accepted plan.
• The task runs inside explicit data, tool, network, and credential boundaries.
• Test output and the final diff are retained for review.
• A named human owns acceptance and production authority.
• Representative-task metrics justify the choice over alternatives.

Build the operating loop

Start the shared AI-native SDLC and add the selected tool’s compact adapter.

Claude Code adapter Map the lifecycle to Claude Code controls.

Cursor adapter Map the lifecycle to Cursor controls.

Codex adapter Map the lifecycle to Codex controls.