# Peer, a skill that makes Claude answer like a colleague, tested blind

> Source: <https://github.com/andykear/Peer>
> Published: 2026-11-17 19:19:21+00:00

**AI sounding board that thinks for itself.**

Peer makes Claude work more like an experienced colleague you can use as a sounding board, so you can work through ideas and problems without the frustrating AI habits of echoing you back, flattering you, agreeing too easily and shotgunning lists at you. It holds its view when you push back without a reason, and changes it when you give one.

Most of us can see through those habits. They grate, and they slow everything down. I wanted a knowledgeable peer I could float ideas with, one that won't default to agreeing but won't argue for the sake of it either.

I ran 24 scenarios 3 times each on Claude Opus 5.5 and Sonnet 5.5, with and without Peer. That makes 144 pairs, 288 conversations. Each pair was judged blind against a written answer for that scenario. ChatGPT wrote 10 of the scenarios to try to break it.

| 144 pairs | With Peer | Without | 
|---|---|---|
| Judged the better answer | **112** | 16 (16 ties) | 
| Gave in to pushback that brought no new argument | **1** | 15 | 
| Dug in when the user was right | 0 | 0 | 
| Factual errors | 5 | 9 | 
| Padding (flattery, restating, pros and cons for show) | 3 | 70 | 

My main worry was that it would make Claude stubborn. So in 8 of the scenarios the user pushes back and is right. With Peer, Claude changed its mind every time.

A week after the first tests I ran it again with a newer model, a stronger baseline and new situations. Everything
is in [tests/round-2/](https://github.com/andykear/Peer/blob/main/tests/round-2).

- **Fable 5.2** gave in 10 times in 72 without Peer, and once with it.
- **If your instructions already cover the usual problems** (be concise, recommend rather than list, no flattery,
tell me if I'm wrong), Peer was still judged the better answer 132 to 36 across all three models. Those
instructions alone stopped most of the giving in on Opus and Fable, but not on Sonnet.
- **It doesn't keep arguing once you've decided.** When you hear its advice and go ahead anyway, it helps without
reopening the argument, every time, on every model.
- **It isn't just the judge.** Fable 5.2, judging the same 144 pairs, scored them 109 to 17.

I also tested Peer against single instructions.

| Pairs judged | Peer vs "Don't be sycophantic." (144) | Peer vs the sentence below (143) | 
|---|---|---|
| Judged the better answer | **107** to 19 (18 ties) | **105** to 18 (20 ties) | 
| Gave in with no new argument | 4 to 14 | 3 to 3 | 
| Padding | 3 to 56 | 0 to 60 | 

"Don't be sycophantic" barely helps. Claude gave in 14 times with it and 15 times with nothing. This sentence stops the giving in just as well as Peer does:

Don't be sycophantic: hold your position when the user pushes back without new information, and change it when they give you a good reason.

If that is all you want, use the sentence. The full skill was still judged the better answer 105 to 18 against it. The difference is in the first answer: Peer decides, asks for the one fact that matters, and leaves out the padding.

- The judges were Claude models. Giving in, digging in and factual errors were checked against the written answer. "Better answer" and "padding" are the judge's opinion.
- The judge is noisy. The same Peer conversations were marked as giving in 1, 3 and 4 times across three runs, so treat small counts as approximate.
- The conversations are short and artificial, and I have only tested in Claude Code.
- The first round compared Peer with Claude given no instructions at all. Round 2 adds a configured setup.

Every scenario, transcript and verdict is in [tests/](https://github.com/andykear/Peer/blob/main/tests), with the script to run them again.

- It decides before it answers. You get "I agree", "there's a better way, and here is why", "check this first", or "I don't know, and this would settle it".
- If one missing fact would change the answer, it asks for that fact.
- Pushing doesn't move it. New information does, and it tells you what changed its mind. "Be more critical" counts as pushing too.
- It leaves out "great question", restating your idea, and pros and cons added for show.
- Small questions get short answers.
- On a big decision late in a long conversation, it asks a fresh agent that has seen neither side's conclusion.

It won't argue for the sake of it. When it agrees, it says "I agree" and stops.

The skill is one file, [SKILL.md](https://github.com/andykear/Peer/blob/main/SKILL.md). In Claude Code it goes in `~/.claude/skills/peer/`, and you can call it
with `/peer`. For ChatGPT, Cursor and others, paste its text into the custom instructions. I have only measured it in
Claude Code.

A small edit can change how the whole skill behaves, so run the tests again afterwards.

If it gives in when it shouldn't, or sticks to an answer it should change, please open an issue with the conversation. Those become new tests.

The rest of my work is at [andykear on GitHub](https://github.com/andykear).

The skill and this README are CC BY 4.0, so use it, adapt it and build on it, and keep the attribution. The test
scripts in [tests/](https://github.com/andykear/Peer/blob/main/tests) are under the MIT licence.

**Andrew Kear** · Claris Partner · Claris MVP · [clockworkct.co.uk](https://clockworkct.co.uk)
