deep-dive

Invoke the deep-dive skill with a request like "audit this codebase" and it deploys 4–6 specialist agents in parallel across lanes such as core logic, architecture, risk controls, and tests. A synthesis agent then cross-checks findings, a red-team attacks them, and you receive a structured evidence package plus an executive briefing.

It is a structured investigation where specialist agents each own one lane of a problem, a synthesis agent reconciles their outputs, follow-up agents verify load-bearing claims, and an adversarial reviewer tries to break the conclusions. The result is sequential markdown research files with severity-tiered findings and honest confidence ratings.

The method runs in any capable agent runtime. Where parallel subagents are unavailable, such as Codex, the same lanes run serially with identical prompts and deliverables, though wall-clock time increases and the final confidence rating reflects the reduced cross-agent independence.

Skip it for single-file code reviews, simple factual questions, tightly scoped tasks like fixing one bug, or anything under roughly 15 minutes of investigative work. For tiny scopes, use the Quick mode with 1–2 lanes instead of a full Standard or Exhaustive run.

No. The skill defaults to pure research mode and produces markdown files only. Code edits happen solely in the patching phase after fresh explicit approval, a git status check, and a named-files patch plan, so concurrent work is never silently overwritten.

Every output carries a 1–10 rating with explicit reasoning plus a ground-truth tally of how many load-bearing conclusions are externally verified through code runs, git, tests, or cited sources. The headline number is capped by that ratio, so mostly judgment-based conclusions cannot exceed the 4–5 band.