Shared standards for Claude Code, Cursor and Codex

Use one team standard for Claude Code, Cursor and Codex. The tools differ in where the agent runs. The standard does not: a person names the task and the boundary, the instructions live in the repository, and a person reviews the change before merge.

That is the same methodology we teach in every workshop. The point of a shared standard is that anyone can pick up anyone else's work.

Put the rules where the agent will read them

Write the working agreement in the repository. AGENTS.md, a project instruction file, or a short team note is enough if it says which work may be delegated, what evidence must come back, and who owns the merge.

Keep it short enough that people will maintain it. A rule nobody opens is not a standard. Name the allowed files and commands for a task class, and say what must be true before review: a diff, a test result or an explanation of why the test could not run, and a note of remaining risk.

Skill rubrics for AI coding teams is a longer version of that agreement.

Measure the gap before you standardise the tool

Ask five engineers how they use AI for coding and you often get five answers. One delegates a whole task. Another still uses autocomplete. A shared standard fails when the team has no shared picture of that gap.

The team proficiency matrix describes five levels, from Assisted to Self-improving. At the Delegated level, an agent does a whole task, such as a bug fix or a small feature, and a person reviews every change and checks the evidence. The Systematic level is the shared way of working: the same instructions, context files and review checks in the repository.

Engineers answer 16 questions. Product and design answer 12. The baseline is the level three in four people have reached, and training starts there. The film and the level guide are in AI coding maturity.

Roll the standard out on one repository

Pilot the written rules on one repository before you copy them everywhere. Track which review comments the team accepts and which it dismisses. Extend the standard only when that pilot is boring and repeatable.

Training runs this on your own codebase, onsite or remote. Services keeps the habit going with a shared channel, a briefing every two weeks, and a monthly live session. Book a 30-minute call to choose the starting repository.

Frequently asked questions

Do Claude Code, Cursor and Codex need different team rules?

The tools differ. The team rule does not. Scope the task, keep instructions in the repository, and have a person own the merge.

Where should the shared instructions live?

In the repository, where engineers already look, such as AGENTS.md or the project instructions your tool reads.

How do we know the team is ready for one standard?

The team proficiency matrix measures the current level. Training starts from the level the group has reached, then moves one level up.