Same brief
Each tool receives the same repository, constraints, and acceptance criteria.
TESTS / REPRODUCIBLE
Public benchmark tests
The brief comes first. Readers should be able to reproduce the task, inspect the acceptance criteria, and challenge the result.
Bench 001
Build and verify one production-ready web feature inside the same repository, using the same requirements and time boundary.
Visible evidence
We keep prompts, artifacts, failures, costs, and the final verification output.
No invented winner
A recommendation remains planned until the test has actually been run and checked.
Freshness built in
Every claim carries a checked date because these products change quickly.
Tools on the first bench
- Codex vs Claude Code
- Cursor vs Claude Code
- Cursor vs Windsurf
- Gemini CLI vs Claude Code
TESTED / MINIMUM EVIDENCE
A finished-looking result is not enough
A review can use the Tested label only after its evidence package is complete and the stated acceptance checks have run.
Required evidence package
- Product version and checked date
- Locked task, inputs, constraints, and acceptance criteria
- Steps another builder can execute
- Elapsed work, attributable cost, and verification results
- Failed attempts and human intervention
- Inspectible artifacts, logs, screenshots, or source links
- Limitations, unknowns, and the boundary of the conclusion
ARTICLE / TEMPLATE
Reusable review structure
Every comparison, workflow, and best-of page follows the same audit trail so another builder can apply or challenge it.
Problem and scope
State the decision being tested, who it is for, and what the review will not claim.
Test environment
Record tool version, plan, repository, hardware, date, dependencies, and starting conditions.
Procedure
Publish the fixed brief, inputs, steps, time boundary, and acceptance checks in execution order.
Measured results
Separate direct measurements from observations, estimates, and unknown values.
Failures and intervention
Keep failed attempts, corrections, retries, and the human work required to reach the result.
Cost and time
Report attributable spend and elapsed work without implying precision the evidence cannot support.
Limitations
Name confounders, missing evidence, version drift, and scenarios the result does not cover.
Fit, not hype
Conclude who the tool fits, who should avoid it, and which facts support that boundary.
Reproduction checklist
Give readers the inputs, commands, checks, and evidence locations needed to repeat the test.