I just had to backfill a year of docs release notes for an OSS repo. I ended up writing a small release tag driven generator: it walks git tags, collects merged PRs between releases, buckets them into categories, and renders Markdown/MDX grouped by year -> month -> category -> version. I also added an optional LLM step that outputs structured JSON via Pydantic schema for PR bullets that includes monthly summaries. It is idempotent and preserves manual edits or omissions, so you can auto-generate, then curate over time.
I’m curious what works for you in practice:
- Towncrier?
- reno?
- GitHub Releases / auto-generated notes?
- something else?
What do you like/hate about your current setup? Any tools you’d recommend?
If anyone wants to see the script I wrote, I am happy to share it and would love some constructive feedback.
I am also planning to add some PR templates so contributors include the context up front, which should make any release note generation more accurate.
Are you using any tooling to help with changelog curration? I know towncrier is all about fragments, so contributors must had write a brief summary of their contribution, which would be more in-line with your preference.
One thing I am trying to do is make the generator surface breaking/migration items explicitly, but I still think anything that requires human judgment (migration steps, caveats) should be hand-curated in a dedicated document like you suggested.
I've never seen an auto-generated set of release notes I liked, a list of PRs doesn't cut it.
If you are willing and interested enough to take a quick look, here is what my script generated for our 2025 changelog (no hand-curation yet, this is the raw output):
https://raw.githubusercontent.com/confident-ai/deepeval/refs...
I am curious: does this still seem too noisy in your opinion, or is it getting closer? And what would you want to see for breaking changes/migrations to make it actually useful?
I now have 2024 & 2025 generated; to fully hand-curate two years of history just wasn’t practical, so I’m trying to get the "80% draft" automatically and then curate over time.
I tried to avoid relying solely on contributors to accurately label or tag things correctly. The script is tag-driven only for release boundaries (version tags), while categorization is derived from PR title & body with optional GitHub metadata. The script is idempotent and preserves edits/omissions so you can correct the few bad ones post-generation.
If you are curious, I am happy to share my script and would be genuinely interested whether it reduces the manual cleanup for your workflow. Also, if you run it with `--ai --github` and a PR body is sparse, it fetches a truncated PR diff and uses that as extra context for the LLM summary.
1. Release notes should never be created mechanically but focus on the consumer of the release.
2. The best changelog is your git. It is OK to generate a calcified changelog for an audience that prefers that.
Last and least:
3. The commit messages are a private space where developers communicate. The messages should never end up at the customer without thorough filtering and distillation.
Apart from that: git-cliff is excellent. If you must generate a document from commits, use that.
Your points are well received and largely why I went PR-based (title/body with optional GitHub metadata) instead of commit-based. A PR title and body tend to be focused on the deliverable, whereas commit messages are narrowly focused on the code change at that moment with developers as the intended audience.
Re: git-cliff, I honestly hadn’t evaluated this one, but it looks solid for commit-driven changelogs. I like the rationale behind conventional commits being parsable and templates enforcing consistency. What constraints pushed you toward git-cliff vs writing release notes by hand, and do you have a config/template you have found works well for surfacing breaking changes?