deep-dive

Rigorous multi-agent deep-dive analysis — for Claude and OpenAI Codex. Point it at a codebase, a strategy or decision system, a design, or an open research question and it runs a structured investigation instead of giving a one-shot answer.

What it does

For a complex, investigative task, the skill runs a multi-phase pipeline:

1. Parallel specialists — deploys 4–6 subagents at once, each in its own lane.
2. Synthesis — one agent reads every specialist output, cross-checks claims, and resolves contradictions.
3. Follow-up verification — single-claim agents re-check load-bearing or single-sourced numbers.
4. Red-team — an adversarial reviewer tries to break the conclusions.
5. Executive briefing — a plain-English verdict with an honest 1–10 confidence rating.

One honest caveat it states up front: every agent is the same model, so the loop catches where independent reasoning diverges, not blind spots they all share — so it weights externally-checkable evidence (code you can run, git, data, cited sources) above claims resting on model judgment, and caps confidence when a conclusion stands only on shared priors. The briefing makes this explicit: every confidence rating carries a tally of how many load-bearing conclusions are externally verified vs. resting on model judgment, and the headline number is capped by that ratio.

It scales down for narrow scope (2–3 lanes) and up for broad, multi-domain work (6 lanes). By default it runs in pure research mode — it writes markdown findings and changes no code unless you explicitly ask.

Four built-in variants live in deep-dive/references/: codebase audit, strategy / system evaluation, design evaluation, and open-ended research.

Quickstart — try this first

You type: "Do a standard design evaluation of . Research-only."

You get back: a research/<topic>/ folder of specialist findings, a synthesis, an adversarial red-team pass, and a plain-English executive briefing with an honest 1–10 confidence rating. No code is touched unless you explicitly ask.

Heads up — this skill is token-hungry by design. A full run fans out 4–6 specialist agents (each writing thousands of words), then synthesis, follow-up verification, a red-team pass, and a briefing — easily 10+ agent calls and tens of thousands of tokens for a single analysis. That's the right trade for a high-stakes call (a ship/no-ship decision, real money on the line), and a great fit on a Claude Max plan or any setup where you're not token-constrained. On a smaller plan, reach for it deliberately: use the Scale heuristics in SKILL.md (2–3 lanes for narrow scope, skip the red-team for low-stakes work), or just ask for a single-pass review instead.

Install

This is an open Agent Skill — the same deep-dive/ folder works in Claude and OpenAI Codex. Pick your tool:

You use…	Install it by…
Claude Code — terminal, the Code tab of the Claude desktop app, claude.ai/code, or a VS Code / JetBrains IDE	dropping deep-dive/ into ~/.claude/skills/ (all projects) or .claude/skills/ (one project)
OpenAI Codex — CLI, app, or IDE	dropping deep-dive/ into ~/.agents/skills/ (all repos) or .agents/skills/ (one repo), then restarting Codex
Claude chat — the Chat tab of the desktop app, or claude.ai	uploading deep-dive.skill (the zip in this repo) under Customize → Skills
Any other agent	pointing it at deep-dive/SKILL.md — it's just instructions

The .skill file is just the deep-dive/ folder zipped, so one unzip drops it into either skills home:

# Claude Code — detected in-session (verify with /skills):
unzip deep-dive.skill -d ~/.claude/skills/

# OpenAI Codex — the current skills path is ~/.agents/skills/; restart Codex after:
mkdir -p ~/.agents/skills && unzip deep-dive.skill -d ~/.agents/skills/

Prefer a clone? git clone https://github.com/nelsonwerd/deep-dive-skill.git, then cp -r deep-dive-skill/deep-dive into whichever skills folder above.

_{Menu names and exact paths shift between versions — the Claude Skills and Codex Skills docs are the source of truth. One caveat specific to deep-dive: its parallel multi-agent orchestration is native to Claude Code; in Codex (and other single-agent runtimes) it runs the same lanes serially — same method, lower cross-agent independence. The methodology is fully portable.}

Runtime support

The method runs everywhere; only orchestration degrades:

	Claude chat	Claude Code	OpenAI Codex	Other agents
deep-dive	Works (degraded: no repo/file access; serial lanes)	Best — parallel subagents + web	Strong — same lanes run serially (lower cross-agent independence, so confidence is capped); external claims labeled unverified if no web	Works (degraded: serial lanes, local-only)

See the Environment & fallbacks section in SKILL.md for the exact fallbacks (no subagents → serial; no web → label external claims unverified; progress tools → skip).

Use it

• Manually: type /deep-dive and describe the target.
• Automatically: Claude invokes it on its own when a task looks like a thorough audit, rigorous analysis, or comprehensive review.

Examples:

• "Do a deep dive on this codebase before I ship it."
• "Evaluate whether this strategy actually holds up under scrutiny."
• "Is this the right architecture? Review it thoroughly."

What's in this repo

• deep-dive/ — the skill itself (SKILL.md + reference playbooks). This is what you install.
• deep-dive.skill — the same folder, zipped, for a one-step download.