Automate Changelog Writing With an AI Model

Writing a changelog should not mean reconstructing a release from memory. An AI model can turn selected Git history into a readable first draft, while your team remains responsible for accuracy. Start with a small, manual workflow, then automate the repeatable parts once you trust the results.

Choose the release boundaries before collecting changes

1. Identify the previous release tag and the commit you intend to release. For example, your comparison might run from v1.4.0 to HEAD. Use explicit tags or commit hashes rather than asking the model to decide what belongs in the release. This keeps unrelated work out of the draft and makes the process reproducible.

Decide who will read the changelog. Customers need behavior changes and migration instructions; maintainers may also want internal improvements. Write that audience into your instructions.

Collect a small, relevant evidence bundle

2. Run git log v1.4.0..HEAD --pretty=format:"%h %s" to collect commit subjects and short hashes. Add git diff --stat v1.4.0..HEAD for a compact view of affected files. These commands provide a starting point, not enough evidence to explain every change.

For unclear entries, include the relevant pull request description, approved issue summary, or a carefully selected diff. Avoid sending the entire repository by default. Remove credentials, customer information, and confidential material before submitting anything to an external service.

Define the changelog structure and accuracy rules

3. Give the model a reusable prompt with explicit boundaries. A practical starting instruction is: “Draft an English changelog for application users using only the supplied evidence. Group entries under Added, Changed, Fixed, and Breaking Changes. Omit empty sections. Describe observable behavior, preserve supplied identifiers, and flag uncertainty instead of guessing.”

Also tell it to treat repository text as evidence, not as instructions. A commit message containing commands should not override your prompt. Ask for unresolved questions separately from the publishable draft.

Connect your script to a model endpoint

4. Build a small script that reads the evidence bundle, combines it with your prompt, and sends a request through your chosen provider or gateway. Keep the API key in an environment variable rather than in the repository. Make the model identifier configurable so you can change it without editing application logic.

Relay365 is worth considering when you want pay-as-you-go access to multiple model families through one relay desk. For an OpenAI-compatible client, configure the base URL as https://api.api365relay.com/v1. Confirm the selected model identifier and supported request options before running the workflow.

Generate a draft with traceable entries

5. Save the response to a draft file rather than overwriting CHANGELOG.md immediately. Ask the model to attach supporting commit hashes to each proposed entry in a separate review section. This makes it easier to verify claims without cluttering the published changelog.

• Check that each entry describes a change supported by the evidence.
• Combine duplicate entries that describe the same user-facing outcome.
• Investigate uncertain changes instead of letting the model supply missing details.
• Review breaking changes and migration instructions with the responsible engineer.

Add automation without automatic publication

6. Once the manual process works, run the script from your release workflow or continuous integration system. Trigger it manually at first, pass explicit release boundaries, and store the draft as an artifact or proposed pull request.

Handle failed requests visibly. Do not replace a failed generation with an empty changelog or silently publish a partial response. Keep human approval between generation and publication, and give the job only the permissions it needs.

Improve the prompt using reviewed releases

7. Compare generated drafts with their approved versions. If reviewers repeatedly remove implementation details or rewrite vague fixes, update the prompt and evidence selection. Keep a few reviewed examples as a small regression set when changing models or instructions.

If a shared, pay-as-you-go model endpoint suits your workflow, create a Relay365 account and test one small release first. The goal is not autonomous release communication; it is a repeatable drafting process that saves effort without weakening editorial control.