Self-Critique - Quality Through Reflection

Self-Critique implements structured self-review workflows inspired by Constitutional AI and self-reflection research.

The core insight: Outputs improve significantly when you pause, review against criteria, and revise—especially your own work.

The Analogy: Code Review (With Yourself)

No Review	Self-Critique
Ship immediately	Pause and review
Miss obvious issues

Catch your own mistakes | | Learn from failures later | Improve before shipping | | Variable quality | Consistently higher quality |

When to Use Self-Critique

Perfect for:

• - 📝 Important writing (emails, docs, proposals)
• 💻 Code before commit (especially complex changes)
• 📊 Analysis and reports (where accuracy matters)
• 🎨 Design decisions (review trade-offs)
• 🤔 Any output with consequences (can't easily undo)

Skip for:

• - ⚡ Drafts or quick notes (speed > quality)
• 🎯 Obvious answers ("What is 2+2?")
• 🔄 Pure exploration (brainstorming)

The Workflow

Step 1: Generate Initial Output

Create your first draft without self-censoring:

CODEBLOCK0

⚠️ Don't self-edit while generating. Get ideas out first.

Step 2: Switch to Critique Mode

Mentally switch roles. You are now the reviewer, not the creator:

CODEBLOCK1

Step 3: Structured Critique

Review systematically against criteria. For each criterion:

CODEBLOCK2

Step 4: Aggregate Findings

Summarize all issues found:

CODEBLOCK3

Step 5: Revise

Address findings systematically:

CODEBLOCK4

Step 6: Second-Pass Critique (Optional)

For critical work, critique again:

CODEBLOCK5

Step 7: Final Output

Ship the revised version:

CODEBLOCK6

Critique Frameworks by Domain

For Code

CODEBLOCK7

For Writing

CODEBLOCK8

For Analysis

CODEBLOCK9

For Decisions

CODEBLOCK10

Severity Levels

When critiquing, classify each finding:

Severity	Action	Example
CRITICAL	Must fix before shipping	Security vulnerability, factual error
MODERATE

Should fix if time permits | Missing documentation, unclear explanation |
| MINOR | Nice to have | Typos, formatting, optional improvements |
| POSITIVE | Keep doing | Good patterns worth reinforcing |

Common Critique Patterns

Pattern 1: The Fresh Eyes

Wait 10 minutes between generating and critiquing:

CODEBLOCK11

Pattern 2: The Role Play

Adopt a specific persona for critique:

CODEBLOCK12

Other personas:

• - Junior developer (confused by jargon?)
• Future maintainer (will this code make sense in 6 months?)
• Adversarial user (how could this be abused?)

Pattern 3: The Red Team

Actively try to break your own work:

CODEBLOCK13

Pattern 4: The Checklist Method

Use domain-specific checklists (see frameworks above):

CODEBLOCK14

Examples

Example 1: Code Self-Critique

Initial Code:
CODEBLOCK15

Self-Critique:
CODEBLOCK16

Example 2: Email Self-Critique

Initial Email:
CODEBLOCK17

Self-Critique:

CRITIQUE:

CRITICAL:
1. No context—recipient doesn't know "the thing"
   FIX: Specify what meeting is about

2. Vague timing—"Tuesday" what time?
   FIX: Propose specific time or ask for availability

3. No purpose—why meeting vs email?
   FIX: State what decision/action needed

4. Too casual for likely audience
   FIX: Professional but friendly tone

MODERATE:
5. No duration estimate
   FIX: "30 minute meeting"

6. Missing call details
   FIX: Include Zoom link or location

REVISION:

Subject: Quick sync on Q3 roadmap - Tuesday 2pm?

Hi Sarah,

Hope you're doing well! I'd like to get your input on the Q3 roadmap priorities before finalizing next week.

Would you be available for a 30-minute call on Tuesday at 2pm PT? I'll send a calendar invite with Zoom details.

If that time doesn't work, I'm flexible Wednesday or Thursday afternoon.

Thanks!
J

### Example 3: Analysis Self-Critique

**Initial Analysis:**

Our API latency increased 50% after the v2 deployment.
The new auth middleware is likely the cause.
We should optimize it.

**Self-Critique:**

CRITIQUE:

CRITICAL:

1. 1. 50% claim not verified—need actual numbers

FIX: Include baseline (100ms) vs current (150ms) p99

1. 2. "Likely" not rigorous—correlation ≠ causation

FIX: Check other variables (traffic, dependencies) FIX: A/B test or profile before/after

1. 3. "Optimize it" too vague

FIX: Specific recommendation with trade-offs

1. 4. No confidence interval or sample size

FIX: Statistical significance check

1. 5. Missing impact assessment

FIX: User impact, revenue impact

1. 6. No mitigation if optimization fails

FIX: Rollback plan, alternatives

REVISION:
CODEBLOCK21

Integration with Other Skills

Self-Critique + Plan First:

1. 1. Plan the task
2. Execute step
3. Self-critique output
4. Revise
5. Next step

Self-Critique + Team Code:

1. 1. Agent generates output
2. Same agent self-critiques
3. Revises
4. Submits
5. Manager integrates

Quick Start Template

CODEBLOCK22

References

• - Research: "Constitutional AI: Harmlessness from AI Feedback" (Anthropic, 2022)
• Related: Self-reflection, Chain-of-Verification, Self-Consistency
• See references/examples.md for more detailed examples