Iron Law

Never conclude first. Evidence determines the conclusion.

Red Flags

• "Let's use SWOT, it always works" → STOP. Framework must fit the problem, not habit.
• "Everyone knows they're the market leader" → STOP. Reputation ≠ evidence. Need 3+ data points per section.
• "The challenge round found nothing" → STOP. A challenge that confirms everything is a rubber stamp.
• "I'll fill in the weaker sections later" → STOP. Uneven depth hides gaps.
• "Data doesn't support it but the conclusion feels right" → STOP. Evidence determines the conclusion.

Analyze

Apply a strategic framework to a topic, then stress-test it with a challenge pass.

When to Use

• Strategic analysis, prioritization, positioning, or planning
• The user explicitly names a framework
• Another skill needs a structured analysis artifact

Supported Frameworks

Workflow

1. Detect the framework and load skills/analyze/references/frameworks/{framework}.md.
2. Gather sources before analyzing. Read internal source files for the subject; for external entities, read current docs, source, changelogs, or feature pages. Record source paths for citation.
3. Ask strategist to apply the framework with evidence and clear implications, enforcing Source Requirements.
4. Ask devil-advocate to attack the weakest 3 points. At thorough depth with mmbridge detected, also dispatch mmbridge debate in parallel.
5. At thorough depth, run a second challenge pass focused on sourcing quality.
6. Synthesize both into a balanced output with recommended actions.

Options

Framework Auto-Detection

When --framework is not specified, detect from the prompt:

1. Explicit name wins.
2. Intent keywords map common prompts to the closest framework.
3. If multiple frameworks fit, ask instead of guessing.
4. If nothing matches, default to swot and state the assumption.

Source Requirements

When analyzing competitors, products, or external entities:

1. Read actual sources, not reputation.
2. Use at least 3 concrete data points per section or quadrant.
3. Cite every external claim inline; mark uncited claims as [unverified].
4. Distinguish fact from inference.

Depth Levels

• quick maps to --depth quick
• standard maps to --depth standard (default)
• thorough maps to --depth thorough

Output

# {Framework} Analysis: {topic}

## Analysis
(framework sections with evidence)

## Challenge
- 3 weakest points
- blind spots and assumptions

## Balanced Insight
(synthesis)

## Recommended Actions
1. ...
2. ...
3. ...

Auto-Save

• Path: .captures/analyze-{framework}-{slug}-{YYYY-MM-DD}.md
• {slug} = topic lowercased, spaces to hyphens, max 40 chars
• Write the full analysis using the Write tool and tell the user the saved path.

Challenge Round

Mandatory at standard (1 round) and thorough (2 rounds). See references/challenge-round.md for the full protocol.

MMBridge Debate Enhancement

When mmbridge is detected (see references/mmbridge-integration.md), analyze can dispatch mmbridge debate during the challenge round.

When to Use

• --depth standard: skip mmbridge debate
• --depth thorough: run it in parallel with the internal devil-advocate

Dispatch

At Step 4 (Challenge Round 1), also run via Bash:

mmbridge debate "<proposition from analysis>" --rounds 2 --json > /tmp/mmbridge-debate-${RUN_ID}.json

Merge

• Parse the mmbridge debate JSON output for arguments tagged FOR and AGAINST
• AGAINST arguments are merged into the challenge round as additional attack vectors
• FOR arguments that contradict internal devil-advocate findings are noted as "disputed by external model"
• Synthesize both internal and external challenges into the "## Challenge" section of the output

Cost Note

mmbridge debate costs more than a single internal challenge, so only activate it at thorough depth.

Gotchas

• Do not force equal depth across all framework sections.
• Do not present generic claims without names, numbers, or evidence.
• Keep the challenge output in the final synthesis.
• Do not analyze competitors by reputation alone.
• Apply the same evidentiary standard to every option being compared.

When Called from PDCA

analyze spans both Plan and Do phases:

• Plan: synthesize research findings into a structured framework.
• Do: apply a production-oriented framework to shape the artifact being built.

Add --context plan|do to override auto-detected PDCA phase.

Subagents

analyst: { model: sonnet, constraint: "pre-process raw research data into structured patterns before framework application" }
strategist: { model: sonnet, constraint: "apply framework, cite evidence, fill required sections" }
devil-advocate: { model: sonnet, constraint: "attack weakest 3 points, surface blind spots" }