I automated my changelog with an agent and it lasted 11 days

The pitch: every merged PR gets summarized, grouped, published to Slipway-style page. The reality: the agent wrote "various improvements and bug fixes" for a migration that renamed half the schema, and attributed a reverted PR as shipped. Colm nearly published it.

What actually works now, much dumber: a template filled from PR labels (feature/fix/breaking + one-line description field that is required to merge), then I spend ten minutes editing every Friday. The agent drafts the ten-minute edit. Humans own the merge checklist.

Agents are great interns. Interns should not publish.

Can confirm. I nearly shipped "various improvements" to 400 customers. The required one-line description field has done more for our changelog than any model. Boring process beats clever pipeline.

Same conclusion from forms land: the agent is excellent at turning my messy notes into a clean draft, terrible at deciding what shipped. Decision stays human, prose goes to the machine.

We use the reverse: the agent reads the published changelog and drafts the in-app announcement plus the help-center update. Source of truth stays human-written, derivatives are automated. Eleven days of pain taught you the right lesson fast.

Oh that is good. Derivatives automated, source human. Our help center lags the changelog by weeks. Stealing this.