deep-dive

What it does

The Deep-Dive skill conducts a rigorous, multi-lane investigation for complex tasks such as auditing codebases, evaluating strategies or systems, validating designs, or open-ended research. It deploys 4–6 specialist agents in parallel across distinct lanes, then synthesizes their findings, runs an adversarial red-team review, applies fixes (only with explicit approval), and delivers a structured evidence package plus a plain-English executive briefing with a 1–10 confidence rating.

How it works

• Phase 0: Setup (one-time per run): mark a session chapter, create a research output directory, set up task tracking, and perform a 30-second initial sweep of repo/files.
• Phase 1: Parallel specialist deployment: dispatch 4–6 specialist agents in a single message. Each agent gets a clearly defined lane, a specific deliverable (Markdown file path, target word count 2500–6000, heading structure), source-verification requirements (at least 2 independent sources for load-bearing numerical claims), severity tiers (Blocker/High/Medium/Low/Note), file:line references when reviewing code, and an end-of-turn 250-word executive summary with an honest 1–10 confidence rating.
• Phase 2: Synthesis: a single synthesis agent reads all outputs, cross-checks load-bearing claims, resolves contradictions, identifies gaps, produces a unified deliverable, lists prioritized recommendations with file:line refs, and provides honest combined confidence with what would change it.
• Phase 3: Follow-up verification (if needed): commission 2–6 focused single-claim verification agents to verify or falsify specific claims with explicit verdicts and revised effect sizes.
• Phase 4: Red-team review: an adversarial reviewer reads the synthesis, attempts to break it, and produces a blocker/high/medium/low list with file:line refs, recommends surgical fixes, and assesses safety to ship.
• Phase 5: Patching (optional, requires explicit approval): only on fresh explicit approval to edit code; run git status --short pre-patching; write a patch plan; apply mechanical edits to approved files if allowed.
• Phase 6: Executive briefing: deliver a final markdown briefing with TL;DR, critical findings by severity with file:line refs, honest confidence and ground-truth tallies, plain-English translations, and a prioritized action list (Tier 0–3) plus a clear proceed decision.

When to use it

Invoke when the user says phrases like "deep dive", "thorough audit", "rigorous analysis", "comprehensive review", "audit this codebase", "evaluate this strategy / design", or related triggers. Also recommended for open-ended investigative tasks involving codebases, strategies, systems, or research questions expected to require substantial analysis (30+ minutes).

What it can touch

The skill uses the tools: "claude-code", "codex" for parallel agents and execution flows. It prescribes delivering Markdown artifacts and an executive briefing, with explicit file:line references in code reviews.

Caveats

• Patching requires explicit user approval and safety checks (git status) before edits. Only applies to named files with a patch plan.
• The workflow relies on parallel agents and a synthesis step; some phase outcomes depend on cross-agent cross-checks and may flag follow-ups for verification.
• External web sources are used only when load-bearing claims require verification; unverified external claims should be labeled as such in the briefing.