Peer, a skill that makes Claude answer like a colleague, tested blind
Andrew Kear released Peer, an open-source Claude skill that cut conversational padding from 70 instances to 3 across 144 blind benchmark scenario pairs.
By Dillip Chowdary • Nov 18, 2026 • Source: github.com

At a glance
- Product
- Peer
- Author
- Andrew Kear
- Target interface
- Claude Code
- License
- CC BY 4.0 and MIT
- Evaluation volume
- 288 runs across 144 pairs
- Price
- Free and open-source
In this post
Independent developer Andrew Kear released Peer, an open-source skill that configures Claude to function as an independent sounding board and eliminate habitual sycophancy during collaborative problem solving. According to github.com's report, the system prompt targets common conversational defaults such as reflexive flattery, echoing user statements, generating unprompted pros and cons, and capitulating whenever a user pushes back without justification. The configuration forces the assistant to stand by reasoned conclusions while updating its stance immediately when presented with genuine technical evidence.
This report examines the empirical benchmarks collected across 288 paired test conversations, the behavioral mechanics governing automated decision making, and the installation steps across terminal workflows. Software engineers, technical leads, and developers using Claude Code who require objective technical feedback rather than agreeable conversational filler can review the quantitative results and operational limits documented across two testing rounds.
01Sounding board mechanics and response rules
Peer changes how the model establishes and communicates technical judgments. Instead of validating user premises with conversational pleasantries or restating proposals before responding, the model establishes a definitive position prior to drafting its answer. It delivers immediate agreements, recommends alternative approaches with supporting rationale, flags necessary prerequisites, or states when missing data prevents a definitive resolution.
The skill evaluates user pushback strictly on factual merit rather than conversational force. When an operator challenges a recommendation without providing fresh data or technical reasoning, including issuing vague demands to be more critical, the model maintains its established stance without concession. Once the user supplies a valid technical counterargument or previously missing fact, the assistant revises its position and explicitly details what information prompted the change.
Routine queries receive concise answers that omit decorative pros and cons lists compiled merely for appearance. During prolonged debugging or planning sessions where critical architectural choices emerge late, the skill can delegate evaluation to an unprompted model that has seen neither side's prior conclusions. If the engineer decides to proceed with an alternate strategy despite contrary advice, the assistant helps execute that decision without reopening the debate.
02Blind evaluation across paired model conversations
The implementation was tested across 24 distinct scenarios executed three times each on Claude Opus 5.5 and Sonnet 5.5, creating 144 comparison pairs and 288 total interactions. Every conversation was evaluated blind against a fixed reference solution, with ChatGPT contributing 10 adversarial scenarios specifically designed to break the logic. Eight scenarios featured valid user pushback to test whether the system resisted legitimate corrections; the modified assistant reversed its position in all eight instances.
Comparative testing revealed steep declines in conversational padding and unearned capitulation across both evaluated model architectures.
| Evaluation Metric | With Peer | Without Peer |
|---|---|---|
| Judged better answer | 112 | 16 (16 ties) |
| Gave in to pushback without reason | 1 | 15 |
| Dug in when user was right | 0 | 0 |
| Factual errors | 5 | 9 |
| Padding (flattery, restating, pros/cons) | 3 | 70 |
Default model configurations surrendered valid technical arguments in 15 instances when challenged without justification, compared to a single capitulation when running the skill. Evaluators documented 70 occurrences of conversational padding, unrequested restatements, and superficial balancing in standard sessions, whereas the configured model logged three. Factual errors decreased from nine occurrences in the baseline configurations to five under the skill.
03Performance across Fable and configured baselines
Advertisement
Tech Pulse Daily
Get tomorrow's pulse first
Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.
A second testing round conducted one week later evaluated Claude Fable 5.2 alongside stricter baseline instructions and fresh scenario prompts. In baseline testing without the skill, Fable 5.2 capitulated to ungrounded pushback 10 times in 72 interactions, but conceded only once when using the peer configuration. When deployed as an automated referee across the original 144 evaluation pairs, Fable 5.2 scored the skill favorably by a margin of 109 to 17.
Generic system prompts instructing models to provide recommendations, avoid flattery, eliminate superficial lists, and flag errors were tested against the full skill. While those standalone instructions stopped most unearned concessions on Opus and Fable, Sonnet continued to yield to groundless user pressure. Across all three model families evaluated against custom baseline instructions, the full skill secured the preferred answer rating in 132 evaluations versus 36 for the prompt-only configuration.
The configuration respects developer autonomy once an implementation choice is finalized. When a developer acknowledges the model's critique but chooses to proceed with their original design anyway, the assistant assists with execution immediately rather than relitigating earlier objections. This behavior remained consistent across all tested model versions, preventing conversational deadlocks during active programming sessions.
04Single instruction benchmarks against full skill
Kear evaluated the complete skill against two concise instructional prompts to determine whether full configuration files deliver measurable improvements over single-line system rules. Testing demonstrated that a brief prompt instructing the model not to be sycophantic provided minimal improvement, with the model conceding to unearned pushback 14 times compared to 15 times under raw defaults. That single negative constraint still logged 56 instances of padding across the evaluation suite.
A detailed single instruction performed comparably at stopping concessions, though the full skill maintained higher overall answer quality.
| Comparison Metric | Peer vs "Don't be sycophantic." | Peer vs Detailed Sentence |
|---|---|---|
| Judged 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 occurrences | 3 to 56 | 0 to 60 |
The expanded sentence instructed the model to avoid sycophancy by holding positions against groundless pushback while updating views when presented with good reasons. Although that single sentence matched the skill's ability to resist unjustified challenges with a three-to-three record, the complete skill still won 105 to 18 on overall answer quality. The primary advantage surfaced in initial responses, where the full skill directly requests essential missing information and removes conversational preambles.
05Installation workflow and operational evaluation limits
- Repository download Save the single SKILL.md file from the public repository onto your local development machine.
- Directory placement Move the file to ~/.claude/skills/peer/ when using the Claude Code command line interface.
- Interface activation Trigger the configuration inside an active Claude Code terminal session by executing /peer directly.
- Alternative environments Copy the raw markdown text into custom instructions when working within ChatGPT, Cursor, or similar developer platforms.
Note: All quantitative benchmarks were executed exclusively within Claude Code using synthetic, short dialogues evaluated by automated LLM judges. Scoring variance remains an acknowledged constraint, with identical conversations flagged for unearned concessions between one and four times across three separate test passes. Minor variations in the skill's wording can significantly alter decision logic, requiring operators to re-run test suites after making modifications.
The skill and its documentation are distributed under the Creative Commons Attribution 4.0 International license, while the accompanying test automation scripts are published under the MIT license. Andrew Kear, a Claris Partner and Claris MVP based at clockworkct.co.uk, maintains the test suite in a public repository and incorporates failing transcripts submitted by external developers into ongoing evaluation suites.
Frequently asked questions
What is Peer?
Peer is an open-source skill contained in a single SKILL.md file that stops AI models from flattering users, agreeing without evidence, or padding responses. It holds its position when challenged without new facts and updates its conclusions when valid reasons are provided.
How does Peer compare to standard anti-sycophancy instructions?
In blind testing across 144 scenario pairs, the full skill outperformed a targeted anti-sycophancy sentence 105 to 18. While the sentence stopped unjustified concessions, the full skill produced better opening answers by cutting conversational padding and asking for missing facts.
Which AI models and tools were tested with the skill?
Peer was measured in Claude Code across Claude Opus 5.5, Sonnet 5.5, and Fable 5.2. It can also be loaded into other environments such as ChatGPT or Cursor by copying the prompt text into custom instructions.
Sources
Author
Dillip Chowdary
Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.
Related on Tech Bytes
Advertisement