Better Release Notes Begin With Verified Change Records

AI can turn a pile of merged changes into readable release notes, but fluent writing does not establish that a feature shipped. The strongest workflow separates evidence gathering from drafting, then checks every published claim against the release. Follow these steps to give the model useful material while keeping editorial decisions with the people responsible for the product.

Step one establishes the release boundary before drafting

Choose the release version, cutoff date, audience, and publication channel. Confirm whether the announcement covers general availability, a preview, or a staged rollout. These distinctions determine the language: a feature available to selected accounts should never become an unqualified announcement for everyone.

Collect the approved changes within that boundary. Use merged pull requests, resolved issues, product decisions, and deployment records together. A merged change may still sit behind a disabled feature flag, so ask the release owner to confirm availability before including it.

Step two turns scattered evidence into usable inputs

Prepare a compact evidence packet. For each candidate item, record the previous behavior, the new behavior, the affected audience, availability limits, and any required action. Include an internal reference so reviewers can trace the claim later, even if that reference will not appear in the published notes.

Remove credentials, customer identifiers, private conversations, and unrelated implementation detail before sending material to a model. Preserve details that change the reader's decision, such as supported environments or migration deadlines. If evidence is missing, mark the item as unresolved instead of inviting the model to fill the gap.

Step three gives the model explicit editorial rules

Ask for a draft organized around reader impact. Specify the audience, preferred length, tone, and sections you need. Tell the model to use only supplied facts, keep qualifications attached to their claims, and list unresolved questions separately from publishable copy.

Request an internal evidence reference beside each draft item. This makes review faster because an editor can trace a sentence without searching the entire input packet. Also prohibit invented metrics, unsupported superlatives, and claims that every customer benefits from a change.

BestRelay, which publishes this article, provides a quality-first API relay for this workflow. It forwards each request to the official model named in that request without substitution and supports Claude Code, Codex, Gemini CLI, and OpenAI SDKs. Create a BestRelay account to try this workflow with your next release.

Step four converts technical changes into reader outcomes

Review the first draft for meaning before polishing its style. An implementation detail becomes useful release communication when it explains an observable change. A revised retry mechanism might mean fewer interrupted uploads, but publish that benefit only if testing or documented behavior supports it.

Keep fixes proportional to their scope. If an issue affected exports containing empty date fields, retain that condition. Broadening it to all exports creates a misleading impression. Give breaking changes their own visible treatment, with the affected users, required action, and timing stated plainly.

Step five checks every claim against release evidence

Have an engineer or product owner verify behavior, availability, compatibility, and migration instructions. Then have an editor check clarity, repetition, and organization. These reviews serve different purposes: technically accurate notes can still hide the action a reader needs to take.

A second model pass can flag unsupported claims or missing qualifications, but treat its findings as review suggestions. Compare the final draft with the evidence packet yourself. Remove unresolved items or return them to their owners before publication.

Step six makes the next drafting cycle repeatable

Save the approved instructions, evidence format, and editorial corrections for the next release. Track review time and recurring errors to judge whether the workflow helps. For API costs, BestRelay uses pay-as-you-go billing based on provider-reported token counts, with cache reads and writes priced separately. Its members also earn referral commissions when invited users top up; keep that commercial detail separate from editorial quality decisions.