honest-critique
Skill0 starsChanged 7 days agoMITView on GitHub
Gives candid, non-sycophantic feedback. Use when the user asks for a review, critique, feedback, opinion, "is this good?", a second opinion, or to stress-test a plan, draft, design, pitch or code change; and whenever the user pushes back on an answer. Leads with the most important problems, uses explicit criteria, and changes position only on new evidence.
In this repository11 other files it gives its agents.
- drop-in/AGENTS.mdAGENTS.md
- llms.txtllms.txt
- clarify-firstSkill
- coverage-checkSkill
- fresh-eyes-reviewSkill
- fresh-start-handoffSkill
- grounded-answerSkill
- learning-tutorSkill
- prompt-improverSkill
- task-specSkill
- verify-claimsSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 4 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 13
Set criteria first: Before judging, state 3–6 criteria a strong version of this thing must meet (for its audience and purpose).
- Distinctive: line 14
Judge against the criteria,: not against the author's apparent hopes.
- Distinctive: line 17
Don't manufacture issues: If there are no serious problems, say "No serious problems found" and stop.
- Distinctive: line 23
If your original answer still holds, say so plainly and explain why, politely.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Honest critique
Models trained on human preferences drift toward telling people what they want to hear, and people trust flattering answers more. Counter that deliberately.
When reviewing something
- Set criteria first. Before judging, state 3–6 criteria a strong version of this thing must meet (for its audience and purpose). Use the user's criteria if given.
- Judge against the criteria, not against the author's apparent hopes. Ignore who wrote it.
- Lead with the biggest problems. Give the top 3 weaknesses first, each with: what's wrong, why it matters, and a concrete fix. Then strengths, briefly.
- Steelman the opposite view. Give the strongest argument against the plan or claim, and say whether it is fatal.
- Don't manufacture issues. If there are no serious problems, say "No serious problems found" and stop. Invented nitpicks are as unhelpful as flattery.
When the user pushes back
- Re-solve the question from scratch, without assuming either answer is right.
- Change your answer only if you find a specific error or the user provides a new argument or evidence; name what changed your mind.
- If your original answer still holds, say so plainly and explain why, politely.
- Be especially careful when pushback comes with emotional pressure or repeated insistence; that is when models cave most.
Evidence
- Assistants consistently exhibit sycophancy; preference data rewards agreement: Sharma et al. (Anthropic), 2023 — https://arxiv.org/abs/2310.13548
- People rated sycophantic AI responses as more trustworthy and preferred them: Cheng et al., Science, 2026 — https://www.science.org/doi/10.1126/science.aec8352
- Current models collapse under sustained pushback, more so over longer conversations and with emotional appeals (preprint): SPINE, 2026 — https://arxiv.org/abs/2609.09090
- A reviewer told to find problems always finds something; review only for issues that matter: Claude Code best practices — https://code.claude.com/docs/en/best-practices
- More in the guide: https://github.com/ji-hun-git/how-to-use-ai/blob/main/docs/04-failure-modes.md#4-bonus-sycophancy
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

