Write GitHub PR Descriptions with AI
To write a GitHub PR description with AI, paste your diff — or a bullet list of what changed and why — into @vustbot and ask for a PR description with summary, change list and test notes. The bot only writes the text; it never touches GitHub, so you paste the result into the PR form yourself.
The PR description is written at the worst possible moment: the code is done, you're out of patience, and the reviewer's context is the price. Turning a raw diff into that context is a summarization job — a good one for a model, if you feed it the "why" the diff doesn't contain.
What the AI does in this scenario
- Diff or change-list in → summary, grouped changes, test notes out
- Fits your team's template: paste the PR template and it fills the sections
- Flags what the diff can't say — asks for the why when you only gave the what
- Works from
git diffoutput, a commit list, or plain bullets - Per-action Balance; Telegram login, no separate account
Worked example: write github pr descriptions with ai
01Input — what you paste
git diff --stat: 4 files. Changes: switched session cache from in-memory Map to Redis; added REDIS_URL env var; TTL now 15 min (was until-restart); migration: none, cache is disposable. Why: multi-instance deploy broke sticky sessions.
02Output — what comes back
Move session cache to Redis — Running multiple instances broke in-memory sessions (each pod had its own Map). This swaps the session cache to Redis so instances share state.
Changes: cache adapter → Redis client; new REDIS_URL env var (required in deploy); TTL now explicit 15 min.
Notes for review: cache is disposable, no migration; check the TTL suits mobile clients.
Testing: ran two local instances against one Redis, session survived instance switch.
How to write github pr descriptions with ai — step by step
- 1Collect the raw material
Run
git diff main...HEAD --statplus the interesting hunks, or just list the changes in bullets. Under ~200 changed lines, paste the diff itself; above that, your bullet summary plus the stat block works better than a truncated wall. - 2Add the why and the risk
The diff shows what moved — only you know the incident behind it, the alternative you rejected, the part you want eyes on. One sentence of each turns a change log into a review brief.
- 3Generate, trim, paste into GitHub
Ask @vustbot for your template's sections, cut anything a reviewer wouldn't thank you for, and paste into the PR form. The bot never opens the PR — the last step is always your copy-paste, which is also your final review of the claims.
AI vs doing it manually
For a one-file fix, typing two honest sentences by hand beats any pipeline — don't ceremonialize small PRs. AI pays off on the 15-file change where writing the description means re-walking your own week: the model groups the diff into themes in seconds and you spend your effort on the two sentences only you can write — why, and what worries you. The failure mode is the inverse: a lyrical AI description wrapping a diff the author never re-read. Reviewers detect that instantly, because the text answers none of the questions the code raises.
The prompt to copy
Write a GitHub PR description from this. Diff/changes: [PASTE DIFF OR BULLET LIST]. Why: [MOTIVATION — BUG/INCIDENT/FEATURE]. Risk areas: [WHAT REVIEWERS SHOULD CHECK]. Testing done: [WHAT YOU RAN]. Format: one-paragraph summary, grouped change list, "Notes for reviewers", "Testing" — match this template if given: [PASTE TEAM TEMPLATE OR "NONE"]. Tone: factual, no marketing.
Frequently asked questions
Can @vustbot open the pull request or read my branch on GitHub?
No. It has no GitHub access at all — it can't read your branch, your repo or your PR, and it can't post anything. You bring the diff or a change summary into the chat, it returns the description text, and you paste that into GitHub yourself. That last manual step doubles as your fact-check.
What do I paste when the diff is too big for one message?
Paste git diff --stat for the shape, the 2-3 hunks that carry the real logic, and bullets for the rest ("plus mechanical renames in 9 files"). A model reasons better from a curated 80 lines than from a truncated 4000 — and curating them takes you one minute because you just wrote the code.
Will an AI-written PR description oversell the change?
Left alone, models drift toward release-notes cheerfulness — so pin the tone in the prompt ("factual, no marketing") and keep the risk section: a description that names what might break reads as engineering, not promotion. The template on this page bakes both in.
Related in Developers
Try it on your real task
The welcome bonus covers a first run — send the prompt above with your own facts and judge the output yourself.
Open @vustbot