Thinking Critically

Rigorous pre-implementation analysis that simulates a paranoid staff engineer review. This skill prevents costly architectural mistakes by forcing systematic examination of assumptions, trade-offs, and failure modes before writing code.

Phase 2 of the 9-Phase Feature Development Cycle:

Interview -> THINK CRITICALLY -> Plan -> Branch -> Implement -> Verify -> PR -> Merge -> Wrap-Up
              ^ YOU ARE HERE

Philosophy

"The best time to find a design flaw is before you write the code."
"Every assumption you don't question is a bug waiting to happen."

When to Use

• Planning a new feature or system (Phase 2 of Feature Cycle)
• Writing or reviewing a PRD
• Making architectural decisions
• Before implementing anything complex
• When something "feels off" but you can't articulate why
• After getting initial requirements, before diving into code

Prerequisites

• Feature spec completed (Phase 1: Interview complete)
• docs/features/FEAT-XXX/spec.md exists with decisions
• docs/project.md exists (for project context)
• docs/architecture/_index.md exists (for technical constraints)

Process

0. Read Context First

# Read feature spec (REQUIRED - output of Interview phase)
cat docs/features/FEAT-XXX/spec.md

# Read project context
cat docs/project.md 2>/dev/null

# Read architecture constraints
cat docs/architecture/_index.md 2>/dev/null

# Check current status
cat docs/features/FEAT-XXX/status.md

0.1 Determine Analysis Depth

AUTOMATIC ABBREVIATION RULES

Condition                        | Steps Executed     | Reason
---------------------------------|--------------------|-----------------------
New feature + new system         | All 11 steps       | Max architectural risk
New feature + existing patterns  | 1-2-3-5-9-11      | Medium risk
Small/clear scope feature        | 1-2-5-11          | Low risk
Bug fix / hotfix                 | SKIP entirely      | No architectural risk

When in doubt -> execute ALL 11 steps

The 11-Step Critical Analysis Protocol

Execute ALL steps in order. Do not skip steps unless explicitly told to or abbreviation rules apply.

Step 1: Problem Clarification & Constraints

Define exactly what we're solving and within what boundaries.

Step 2: Implicit Assumptions Identification (CRITICAL)

Surface hidden assumptions that could invalidate the entire approach. Most failures stem from unstated assumptions.

Step 3: Design Space Exploration

Enumerate possible approaches before committing to one.

Step 4: Trade-off Analysis

Explicitly weigh competing concerns.

Step 5: Failure-First Analysis

Systematically identify what can go wrong.

Step 6: Boundaries & Invariants

Define what must ALWAYS be true and what the system will NEVER do.

Step 7: Observability & Control

Ensure the system can be monitored and managed.

Step 8: Reversibility & Entropy

Assess how hard it is to undo decisions and recover from mistakes.

Step 9: Adversarial Review (Paranoid Staff Engineer Mode)

Attack the design from every angle.

Step 10: AI Delegation Assessment

Determine what AI (Ralph Loop) can safely handle vs what needs human oversight.

Step 11: Decision Summary

Synthesize everything into actionable conclusions.

Critical Rules

1. NEVER skip Step 2 (Assumptions) - most failures stem from here

2. If ANY assumption has Low confidence + High impact -> STOP
   Validate with user before proceeding

3. Step 9 must be genuinely adversarial - attack your own design

4. All 11 steps for complex/critical features
   Steps 1-2-5-11 for smaller features

5. In Ralph Loop: auto-PAUSE if confidence is Low in Step 11

Ralph Loop Integration

When running autonomously via Ralph Loop:

Iteration 1: INTERVIEW -> reads spec.md, pauses if TBD -> emits INTERVIEW_COMPLETE
Iteration 2: THINK CRITICALLY -> executes protocol -> emits ANALYSIS_COMPLETE
             AUTO-PAUSE if:
             - Low confidence + High impact assumption found
             - Step 9 identifies critical red flag
             - Step 11 confidence level is "Low"
Iteration 3: PLAN -> reads spec.md + analysis.md -> generates design.md + tasks.md

How Analysis Feeds Plan Phase

The Plan phase reads analysis.md in addition to spec.md to:

• Select recommended approach (from Step 11)
• Incorporate failure mitigations (from Step 5) into design
• Include invariants/boundaries (from Step 6) as validations
• Use AI delegation matrix (from Step 10) to decide Ralph automation scope
• Add observability requirements (from Step 7) to implementation tasks

Output

After completing the protocol, save output to docs/features/FEAT-XXX/analysis.md

Status Updates

After completing analysis:

1. Update status.md: Phase: Critical Analysis complete

2. Update _index.md: Feature status updated with phase progress

3. Update context/decisions.md with key decisions from Step 11

4. Append to status-log.md with action record