Constitutional Self-Critique

Constitutional self-critique provides explicit principles for the model to critique against. Rather than generic quality dimensions, the critique tests compliance with a stated set of principles — essentially, the model checks its own output against a specification. This is directly analogous to constitutional AI: the critique is governed by a constitution of principles.

# Constitutional self-critique

CONSTITUTION (principles your response must satisfy):
1. PRECISION: Every recommendation must be specific enough to implement without clarification
2. SCOPE ALIGNMENT: Recommendations must address only the stated problem, not adjacent problems
3. CONSTRAINT COMPLIANCE: Every recommendation must be feasible within the stated constraints (budget, timeline, headcount)
4. MEASURABILITY: Every recommendation must include a measurable success criterion
5. RISK DISCLOSURE: Significant implementation risks must be identified

After drafting your response, apply the constitution:

For each principle (1-5), determine:
PASS: The response satisfies this principle
FAIL: The response violates this principle — [specific location and correction needed]

If any principle FAILS, revise the response to address the failure before outputting.

Output format:
---CRITIQUE---
1. PRECISION: [PASS/FAIL] [explanation if FAIL]
2. SCOPE ALIGNMENT: [PASS/FAIL]
3. CONSTRAINT COMPLIANCE: [PASS/FAIL]
4. MEASURABILITY: [PASS/FAIL]
5. RISK DISCLOSURE: [PASS/FAIL]
---FINAL RESPONSE---
[response here, revised if needed]

1. Write the Constitution First The principles must be written before the task prompt, not derived from it. The constitution is the quality standard. If you write it after seeing the output, you are rationalizing, not specifying.
2. Keep It to 5-7 Principles More than 7 principles creates a checklist so long the model skips items or provides superficial assessments. Five to seven principles, each with a clear pass/fail criterion, produces usable critique output.
3. Make Principles Binary "The response is clear" is not binary. "Every recommendation includes a specific metric" is binary. Binary principles produce actionable critique: PASS or FAIL, with a clear correction when it fails.