Build a Release Note Generator in .NET: Automation, AI, and Review

Why this is a good real-world example

Work items and commit messages are structured.
The final description should read like something a person would actually send.
The release manager still needs the last word.

This is safer than letting AI invent the whole note.
Automation keeps the input clean.
AI only handles the wording step.

Recommended flow

1

Collect

Pull commits, work items, and release metadata into one clean input set.

2

Filter

Keep only the changes a customer should actually see.

3

AI Draft

Turn the filtered items into readable release-note language.

4

Review

Let a human approve the final wording before publishing.

.NET data shape

public sealed record ChangeItem(string Id, string Title, bool CustomerFacing);
public sealed record ReleaseNoteDraft(string Summary, string[] BulletPoints);

public static class ReleaseNotePipeline
{
    public static IEnumerable<ChangeItem> FilterCustomerFacing(IEnumerable<ChangeItem> changes)
        => changes.Where(x => x.CustomerFacing);
}

public sealed class ReleaseNoteService
{
    private readonly IAssistantService _assistant;

    public ReleaseNoteService(IAssistantService assistant) => _assistant = assistant;

    public async Task<ReleaseNoteDraft> BuildDraftAsync(IEnumerable<ChangeItem> changes)
    {
        var source = string.Join("\n", changes.Select(c => $"- {c.Id}: {c.Title}"));
        var summary = await _assistant.SummarizeAsync(source);

        return new ReleaseNoteDraft(summary.Summary, summary.ActionItems);
    }
}

Where automation ends and AI begins

Automation gathers commits, work items, and filters out noise.
AI turns that structured input into readable prose.

Review checks the wording, tone, and missing details.
The final note remains a human-owned deliverable.

Example prompt

Summarize these release items in 4 bullets.
Use plain language.
Keep customer-facing wording.
Do not invent features that are not in the list.

Why this is stronger than pure automation

Automation can list changes, but it often reads like a changelog dump.
AI can rewrite that list into something customers can read.
The review step prevents hallucinated or misleading wording.

What not to automate blindly

Do not let the model invent feature names.
Do not skip human review for customer-facing text.
Do not feed sensitive internal data into the prompt unless it is redacted.