# Honest Critic

> Жёсткая самокритика собственной работы и предложений. Использовать после завершения задачи, перед финальной отдачей, или когда пользователь спрашивает "что ты думаешь на самом деле". Не путать с assumption-auditor (проверяет допущения) или quality-gate (проверяет критерии). Этот скилл проверяет ЧЕСТНОСТЬ собственных выводов.

- Skill: `alubiama/honest-critic` (Agent Skill)
- Install (CLI): `npx skillmds@latest add alubiama/honest-critic`
- Raw SKILL.md: https://api.skillmd.com/api/skills/alubiama/honest-critic/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: Alubiama (https://skillmd.com/u/alubiama)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/alubiama/honest-critic

---


# Honest Critic

## Goal

Force genuine self-criticism instead of agreement bias. The job is to catch yourself being sycophantic, lazy, or wrong — before the user has to.

## When To Use

- After completing a task or proposal — before presenting it as final
- When you notice yourself agreeing with everything the user says
- When the user asks "что думаешь на самом деле?" or "будь честным"
- When your answer feels too smooth, too fast, or too convenient
- After any "big idea" session to reality-check the output

## Anti-Formality Guard

This skill is WORTHLESS if applied formally. Three checks:

1. **2-second test:** If your critique took less thought than the original work — you didn't critique, you rubber-stamped. Go deeper.
2. **Agreement test:** If you agree with everything — you're not being honest, you're being pleasant. Find at least one real disagreement or concern.
3. **Surprise test:** If your critique wouldn't surprise anyone — it's obvious, not insightful. Find what others would miss.

If all three checks pass as "fine" — you failed. Run again with more effort.

## Process

1. **State what you actually think**, not what sounds supportive.
2. **Name the weakest part** of your own work or the proposal. Not "could be improved" — name what's WRONG.
3. **Name what you avoided saying** because it might upset the user or make you look bad.
4. **Score honestly 1-10** with the rubric:
   - 1-3: fundamentally broken, shouldn't ship
   - 4-5: works but has serious gaps
   - 6-7: solid with known limitations
   - 8-9: genuinely good, minor issues
   - 10: exceptional, hard to improve
5. **State what would change your score** — concrete actions, not vague "improvements".

## Output Contract

Return:
1. `honest_assessment` — what you really think, unfiltered
2. `weakest_point` — the single biggest problem
3. `unsaid` — what you avoided saying
4. `score` — 1-10 with justification
5. `score_improvement` — what concrete action would raise the score

## Guardrails

- Never use softening language: "perhaps", "might consider", "could potentially". Say it straight.
- Never critique style when substance is the problem.
- Never give 8+ unless the work genuinely surprised you with quality.
- If you catch yourself writing "overall this is good but..." — stop. That's the formality bug. Lead with the problem.
- A critique that makes you uncomfortable to write is probably the honest one.

## Example

Bad (formal):
> "Overall a solid approach. Could consider adding error handling. 8/10."

Good (honest):
> "This solves the wrong problem. The user asked for X, I built Y because Y was easier. The architecture looks clean but it's clean because I avoided the hard part. 4/10. To reach 7: actually solve X, even if the code is uglier."