People
Contributors and maintainers
- Does
- Bring problems and ideas, discuss, test, review, decide, and take responsibility for every release.
- Limits
- Only people hold approval authority. That is a principle of the approved charter.
How we build
AI helps with the work. People make the decisions, and take responsibility for whatever ships. Here is the path an idea takes, and who does what along the way.
What is agreed, and what is proposed. The approved Founding Charter sets the principle: AI can help throughout, and releases need human oversight. The detailed steps and roles below come from pull request #3 (on GitHub) and pull request #5 (on GitHub), which are still open for discussion and not yet in force.
The charter's development approach
- Step 1: Identify a problem
- Step 2: Discuss ideas
- Step 3: Design
- Step 4: Build
- Step 5: Test
- Step 6: Review
- Step 7: Release
- Step 8: Maintain
From idea to release
Steps with the peach mark are decisions only a person can make.
Step 1: A person describes a problem
People
Anyone can open an issue explaining who is affected and what they have tried.
Step 2: Research and a specification
ChatGPT, checked by people
ChatGPT can help research existing tools and draft a specification. People check its claims.
Step 3: An approved issue
Decided by people
People decide the work is worth doing and record the decision publicly.
Step 4: Building on a branch
Claude or Base44
Claude implements the issue on its own branch with tests. Base44 is used only for apps it has been assigned to.
Step 5: A pull request with evidence
Claude, with automated checks
The pull request records what changed, how AI was used, which tests ran and what remains uncertain.
Step 6: Human review checkpoint
Decided by people
A maintainer reads the change, checks the evidence and approves it or asks for changes.
What reviewers check
- Does the change do what the issue asked, and nothing it should not?
- Were the tests really run, and do the results match what the pull request claims?
- Is it accessible, private by default and safe to maintain?
- Is AI assistance disclosed, so reviewers know what to check closely?
Under the draft policy, security-sensitive changes need review by someone other than their author, and high-risk features stay unreleased until a qualified, independent person has reviewed them.
Step 7: A release decision
Decided by people
A named person approves the release, with a checklist, known limitations and a way to roll back.
Step 8: Maintenance
People, helped by AI
Maintainers fix problems, review updates and say clearly when something is no longer supported.
Who does what
GitHub
The source of truth
- Does
- Holds the code, issues, proposals, reviews and recorded decisions. If this website and GitHub ever disagree, GitHub is right.
- Limits
- A repository being on GitHub does not make its code reviewed or safe.
ChatGPT
The ideas engine
- Does
- Helps research problems, compare existing tools and draft specifications.
- Limits
- Cannot set priorities or approve anything. Its suggestions are checked like any other source.
Claude
The creative engineering agent
- Does
- Designs and builds approved work on branches, writes tests and opens pull requests with honest evidence.
- Limits
- Cannot approve, merge its own work, publish a release or deploy to production.
Base44
Builder and host for designated apps
- Does
- Builds and hosts specific existing apps it has been assigned to. This website is hosted on Base44.
- Limits
- Its GitHub sync does not automatically follow pull request rules, so its changes need the same human checks.
Proposed permission levels for AI agents
The draft policy gives AI agents only the access a task needs. Release and production authority stays with people. This is a proposal from pull request #3, not yet in force.
A0: Research and drafts
Examples: Summarise public material, propose ideas, draft text and designs.
Required control: People review anything before it is published.
A1: Isolated development
Examples: Write code and tests in a sandbox, create branches, propose pull requests.
Required control: Temporary, limited permissions. People approve before anything merges.
A2: Sensitive changes
Examples: Dependencies, CI, authentication, encryption, permissions, file changes, network access.
Required control: Approved in advance, with security-focused human review and a written risk assessment.
A3: Release and production
Examples: Merge protected changes, sign releases, deploy, manage production secrets.
Required control: Reserved for human maintainers. No standing AI authority.
Being open about AI
- Pull requests say whether and how AI was used, and which parts a person checked.
- An AI-written test report is not proof that a test ran. We link to real results instead.
- AI suggestions are evidence to examine, not decisions to follow.
- Issues, web pages and files are treated as untrusted input that cannot change an agent’s task.
About this website
This website was designed and built by Claude, on Base44, at the founder's request, from the public website blueprint (on GitHub).
The founder approved it for publication on 14 November 2026. That decision was made by a person, not by Claude. See what was checked.
Open-source code, proprietary tools
ChatGPT, Claude and Base44 are proprietary services run by separate companies. Comutability's own software is intended to use recognised open-source licences. Using these tools to help build something does not make them open source, and you never need any of them to contribute.