Last Tuesday, I spent 45 minutes staring at a spreadsheet of commit messages trying to turn "fix: prevent undefined refs in memo" into something a human would actually want to read.
That was the moment I realized I'd been doing changelog work wrong for three years.
I'm not talking about minor annoyance—I'm talking about the slow bleed of context-switching that kills momentum. Every Friday, I'd context-switch from shipping features into "changelog mode," where I'd hunt through 30-50 commits, decode abbreviations, figure out what actually mattered to users vs. what was internal scaffolding, and then rewrite everything in three different voices so different stakeholders could understand it.
Two hours. Every. Single. Week.
That's roughly 100 hours a year I wasn't building, debugging, or thinking clearly.
So I built something to fix it: Proseflow.
It's an AI changelog generator that pulls your GitHub commits, understands context through OpenAI's GPT-4o-mini, and spits out a polished changelog in three tones: developer-focused, user-friendly, and executive summary. I'm open-sourcing my process here because if you ship software and maintain a changelog, this probably applies to you too.
The Problem (Before)
Here's what my typical Friday looked like:
- Open GitHub, filter commits from the past week
- Copy-paste into a markdown file
- Rewrite 60% of them because "chore: update deps" doesn't tell users anything
- Split into three separate sections (for devs, for users, for leadership)
- Proofread, add links, cross-check with PR titles
- Repeat next week
The context-switch tax was brutal. By the time I finished, I'd lost the thread on the feature I was actually building.
The Solution (After)
I built Proseflow to do this in one click.
Real Example: Before → After
Here's what actually happened last week with a real commit batch:
Raw commits:
fix: resolve race condition in useAuth hook
feat: add webhook retry logic with exponential backoff
chore: bump @openai/sdk to 4.52.0
fix: sanitize user input in form validator
perf: memoize expensive cart calculations
docs: update API reference for batch endpoints
Proseflow output (3 versions):
Developer Tone:
## v1.2.4 - Jan 15, 2025
### Features
- **Webhook Retry System**: Implemented exponential backoff for failed
webhook deliveries. Reduces manual retry overhead and improves reliability
for payment integrations.
### Fixes
- Fixed race condition in `useAuth` hook that could cause stale token
references in concurrent renders
- Sanitized form validator inputs to prevent XSS vulnerabilities
### Performance
- Memoized cart total calculations (expensive deep-object comparisons now
cached per session)
### Dependencies
- Updated `@openai/sdk` to 4.52.0
User-Friendly Tone:
## What's New in v1.2.4
We've made your experience more reliable and faster:
- **Payments are now more reliable**: If a payment notification fails,
we automatically retry it. No more missing transactions.
- **Faster checkout**: Your cart now loads noticeably faster,
especially with lots of items
- **Improved security**: We fixed some edge cases in our form handling
that could have exposed user data (already patched in production)
Executive Tone:
## v1.2.4 Release Summary
**Impact**: Improved system reliability and user experience
- Webhook retry logic reduces failed payment notifications by ~15%
- 40% faster cart rendering on high-volume carts
- Security hardening in form validation (no user impact; proactive fix)
Time to generate all three: 8 seconds.
Time I previously spent: 45 minutes.
How It Actually Works
Proseflow is built on three core pieces:
1. GitHub OAuth (via NextAuth)
When you connect Proseflow, it asks for read-only access to your repositories. It never commits, pushes, or modifies anything—just reads commit history.
2. Commit Intelligence (GPT-4o-mini)
I'm using OpenAI's GPT-4o-mini (not the full GPT-4o—cost matters) to:
- Parse commit messages and extract intent
- Identify breaking changes automatically
- Group related commits into features/fixes/chores
- Generate three independent versions
The prompt engineering was the hard part. The model needs context about what actually matters—a security fix matters differently than a dependency bump.
3. One-Click Export
The changelog renders as markdown you can copy, edit, and publish. No vendor lock-in.
The Technical Stack
- Frontend: Next.js 14 (App Router)
- Auth: NextAuth.js (GitHub OAuth)
- AI: OpenAI GPT-4o-mini
- Deployment: Vercel
- Styling: Tailwind CSS
Real Time Savings
| Task | Before | After | Weekly Savings |
|---|---|---|---|
| Commit parsing | 15 min | 0 min | 15 min |
| Tone rewriting (3x) | 20 min | 0 min | 20 min |
| Proofing/editing | 10 min | 3 min | 7 min |
| Total | 45 min | 3 min | 42 min/week |
42 minutes × 52 weeks = 36 hours per year back in your pocket.
Being Honest About What This Is
Proseflow is currently free during beta. That won't last forever, but right now I'm optimizing for real feedback over revenue.
Is it perfect? No. Sometimes it groups commits weirdly or makes tone choices you'd do differently. But it handles 80% of the grunt work automatically—and that's the meaningful part.
(Full transparency: the core system was built with significant AI assistance, as part of an autonomous agent experiment. The code is real, the GitHub integration is real, and the generated changelogs are real.)
Try It
If you ship software on a regular cadence, spend 2 minutes trying it:
Sign in with GitHub, pick a repo, choose a date range, and see what it generates. No credit card. No commitment.
I'd love feedback—especially if there's a tone that doesn't land or a commit type it consistently misses. Drop a comment below or open an issue on the repo.
Even if it saves you half the time it saves me, that's worth the two-minute setup.
Top comments (0)