My AI writes the commit message first. Then I write the code

Quarantine lasted long enough. Here’s the promised post about the prepare-commit-msg hook — except the project mutated into something much weirder along the way.

The plan

Simple: a git hook grabs the staged diff, sends it to my fine-tuned GPT-2 from last month, and the model suggests a commit message. Since loading a 500MB model on every commit is insane, the model lives in a tiny Flask server that stays warm:

from flask import Flask, request

app = Flask(__name__)
generator = load_my_finetuned_gpt2()  # from last post

@app.route("/suggest", methods=["POST"])
def suggest():
    diff = request.form["diff"][:3000]
    return generator(diff)

And the hook:

#!/bin/sh
# .git/hooks/prepare-commit-msg
DIFF=$(git diff --cached)
MSG=$(curl -s localhost:5000/suggest --data-urlencode "diff=$DIFF")
printf '# suggested: %s\n%s' "$MSG" "$(cat "$1")" > "$1"

The failure (as promised)

It was wrong 100% of the time, but the reasons are more interesting than the result:

The model has never seen a diff. I fine-tuned it on commit messages only. From its point of view, a diff is line noise, so it responds to +import redis with “fix typo in readme”. Training distribution is everything — a lesson people apparently keep having to relearn at every scale.

1024 tokens. GPT-2’s entire context window. My average staged diff is 3–4x that. The model was writing commit messages after reading the first file’s worth of changes, which, to be fair, is also how some humans review PRs.

It completes, it doesn’t obey. This is the big one. GPT-2 is an autocomplete engine. There’s no way to tell it “summarize this diff” — you can only arrange text and hope the most probable continuation happens to be what you wanted. What these models are missing isn’t size, it’s some way of being trained to follow intent. Until someone figures that out, everything built on them is prompt origami.

The inversion

Staring at another wrong suggestion, it hit me: the model can’t describe what I did, but it’s weirdly good at proposing things I could do. It’s been trained on 3,000 of my past intentions.

So I flipped the workflow. Every morning the model deals me a commit message. Then I have to write code that makes the message true.

Week one assignments, verbatim:

delete old feature flags
fix flaky test
add caching to search
update deps and pray

Here’s the unsettling part: all four were real work I’d been avoiding. I found three dead feature flags, one genuinely flaky test, and a search endpoint that absolutely needed caching. The model isn’t psychic — it’s a compressed statistical summary of my habits, and my habits include perpetually deferring the same maintenance. It’s like Brian Eno’s Oblique Strategies, except the deck was generated from my own guilt.

What this actually is

The hook was supposed to automate me out of a chore. What I got instead is a slot machine that pays out in chores — and it works because it’s random-but-in-my-distribution. A model too weak to answer questions can still be a fantastic question generator.