Test Your Move From Claude to Pi in One Sitting

In one sitting you will know whether Pi can stand in for Claude Code on your own repo. You’ll have a diff to prove it, not a vibe.

Why now: a post titled Goodbye Claude, Hello Pi is circulating on Hacker News. It’s one developer’s account of leaving Claude for Pi. Reading it won’t tell you how Pi behaves on your code. A controlled run will.

The method is a bakeoff. Same commit, same instructions, same task, two agents.

1. Tag your current commit so both agents start from identical code.
2. Copy your Claude instructions into a neutral filename.
3. Write one real backlog task into task.md. Pick one whose correct answer you already know.
4. Run it in Claude Code. Save the diff and the test output.
5. Reset to the tag. Open Pi in the same repo and paste in task.md.
6. Compare the two diffs and the two test results.

Paste this to run steps through:

git tag bakeoff-base
cp CLAUDE.md AGENTS.md

# Claude Code run (print mode)
claude -p "$(cat task.md)"
git diff > claude.diff
npm test > claude.test.txt 2>&1

# reset, then start Pi in this repo and paste task.md
git reset --hard bakeoff-base
git clean -fd

# after the Pi run
git diff > pi.diff
npm test > pi.test.txt 2>&1
diff claude.diff pi.diff | less

Swap npm test for whatever your project actually runs.

The gotcha: your instructions may lean on Claude Code features. Skills, subagents and MCP servers are Claude Code mechanisms, and a different harness can’t be assumed to honor them. You’ll recognise the problem when the second agent skips a step your file takes for granted, like running tests before it declares success. Don’t score that as the model failing. The test is rigged until the instructions stand on their own, so rewrite any rule that depends on a Claude-only feature as plain text and rerun.

My read: the lock-in in agent tooling lives in your instructions and habits, not in the model. If the two diffs pass the same tests on your tasks, switching is a cost and preference decision, not a capability one. If they don’t, you’ve learned what your setup actually depends on, which is worth knowing either way.

If your CLAUDE.md is a pile of accreted rules, start with the breakdown of Claude.md, skills, subagents and MCPs to see which parts are portable and which are not.

Want the next agent-tooling test in your inbox? Subscribe. One AI signal a day. 90 seconds. No fluff.