analyze

Structured Analysis Skill

Apply CIA/IC Structured Analytic Techniques to produce defensible, evidence-based analytical assessments. Every claim must be cited. Every judgment must trace to technique outputs.

Invocation

/analyze                          → Adaptive mode (auto-select techniques)
/analyze <technique>              → Direct mode (run one technique)
/analyze --guided                 → Guided mode (walk through all phases)
/analyze --resume <analysis-id>   → Resume or update existing analysis
/analyze --iterate <analysis-id>                → Re-run full analysis with new evidence
/analyze --iterate <analysis-id> <technique>    → Re-run specific technique(s)
/analyze --lean                   → Lean mode (abbreviated technique set)
/analyze --comprehensive          → Comprehensive mode (full rubric, adversarial + deception checks)
/analyze --no-osint               → Disable web research

Techniques: customer-checklist, issue-redefinition, restatement, brainstorm, kac, ach, inconsistencies, cross-impact, what-if, premortem, counterfactual, narratives, bowtie, opportunities, devils-advocacy, red-hat, alt-futures, deception

Flags combine: /analyze --guided --no-osint is valid.

Execution

You MUST read the orchestrator protocol before proceeding. It contains mode routing, technique selection logic, and the technique routing table.

Step 0 — Context Inference

Before parsing explicit arguments, scan the conversation history for implicit inputs. Users often invoke /analyze mid-conversation after discussing a problem, providing data, or sharing links.

Extract from conversation context:

• Problem statement: What is the user trying to analyze? Look for questions, concerns, scenarios, or decisions under discussion.
• Implicit technique hints: Did the user mention assumptions, hypotheses, competing explanations, risks, or scenarios? Map these to techniques (e.g., "I'm not sure which explanation is right" → ACH, "what could go wrong" → Premortem).
• Implicit flags: Did the user indicate they want something quick (→ --lean), don't want web research (→ --no-osint), or want to walk through everything (→ --guided)?
• Prior analysis: Are there existing analyses in analyses/ for the same topic? (→ suggest --resume or --iterate)
• Evidence already provided: Files shared, URLs pasted, data discussed — these become Tier 1/2 evidence.

Step 0.1 — Validate Assumptions

If context inference produced any results, present them to the user for confirmation before proceeding:

Based on our conversation, here's what I'm picking up:

**Problem**: [inferred problem statement]
**Mode**: [inferred mode + rationale]
**Techniques**: [inferred techniques, if any]
**Flags**: [inferred flags, if any]
**Prior context**: [files, data, or evidence already in conversation]

Does this look right? Adjust anything before I proceed.

If the user provided explicit arguments, those always take precedence — but still surface any useful context (e.g., "You asked for ACH. I also noticed you shared [file] earlier — I'll include that as evidence.").

If no conversation context exists and no arguments were provided, proceed directly to Adaptive mode (the orchestrator will prompt for a problem statement).

Steps 1–6 — Main Execution

1. Read protocols/orchestrator.md (relative to this skill's directory)
2. Parse explicit arguments → determine mode and flags (merge with Step 0 inferences, explicit args win)
3. Follow the orchestrator's instructions for the detected mode
4. For technique execution, follow the orchestrator's Technique Execution Contract:
   • 1 technique (Direct mode): Execute in-context — read protocol, read template, execute SETUP → PRIME → EXECUTE → ARTIFACT → FINDINGS → HANDOFF, write artifact to analyses/<id>/working/
   • 2+ techniques: Dispatch to background subagents in dependency-aware tiers — each subagent reads protocol/template, executes the technique, writes the artifact, and returns only a compact findings summary. Main context accumulates summaries and file paths, not full technique work.
5. For evidence gathering: read and execute protocols/evidence-collector.md
6. For report synthesis: dispatch per protocols/report-generator.md Phase A/B architecture

Self-Correction (3 Layers)

• Layer 1 (after each technique, silent): Protocol compliance check — all steps completed? All template sections filled? No unfilled {{PLACEHOLDER}} tokens?
• Layer 2 (before report, silent): Analytical self-critique — 8 checks:
  • 3a. Assumption audit
  • 3b. Evidence balance
  • 3c. Confidence calibration
  • 3d. Alternative check
  • 3e. Missing voices
  • 3f. Internal consistency audit (cross-artifact validation)
  • 3g. Analytical bias scan (sycophancy, anchoring, vividness, completion, authority)
  • 3h. Quality score (quantitative 1-5 with pass/fail threshold)
• Layer 3 (before finalization): Human review gate — present summary (including quality score), incorporate feedback
• Auto-Remediation Gate (between Phase A and Phase B): HIGH-severity Layer 2 flags (evidence imbalance >2:1, unstated critical premises, strong counter-arguments, sycophancy/anchoring bias, quality score < 3.0) trigger automatic remediation — the orchestrator invokes the iteration handler to collect targeted evidence, re-run flagged techniques (max 3), and regenerate the report before the user sees it. Capped at 1 cycle. Zero overhead when no HIGH flags exist.
• Critique-to-Iteration Bridge (after results): Remaining MEDIUM/LOW flags from Layer 1 and Layer 2 are mapped to specific technique re-runs and evidence collection focuses, presented as ready-to-run /analyze --iterate commands. Only fires when actionable flags exist beyond what auto-remediation already addressed. All flags and their statuses are written to next-steps.md in the analysis root — a standalone ledger tracking OPEN, REMEDIATED, RESOLVED, and DEFERRED items across iterations. The --iterate handler reads this file as its primary input.

Citation Requirement

Every claim in every artifact must be cited. No exceptions. Citation methods:

• OSINT: [Source](URL) — Retrieved: YYYY-MM-DD
• FILE: [filename:line_range]
• USER: [User-provided, session context]
• ANALYSIS: [Derived via technique_name]
• PRIOR-ITERATION: [PRIOR-v{N}: technique_name]

OSINT is never presented as fact — always "according to [source]".

Reference Library

For deep background on any technique, read docs/library/00-prime.md and the specific library files referenced in each protocol. The library contains the full theoretical foundation, axioms, selection matrices, and empirical critiques underlying this skill.