bmad-deep-recon

Invoke the skill with your topic and decision, then choose Run for native web research, Draft to get a prompt for an external tool like ChatGPT or Perplexity, or Process to distill an existing report. Run mode holds a plan gate before fanning out researcher subagents.

It ships six type packs: market, domain, technical, competitive, user-voice, and academic literature. Each pack defines prioritized dimensions, source craft, freshness windows, and two-source claim classes, and custom types can be added through override TOML files.

Yes. Process mode files the original report into a run folder's imports directory, extracts claims into digest files with provenance, checks coverage against the type pack, and distills a cited research.md summary that downstream skills consume without reprocessing.

Verification runs as material lands, at normal, high, or max levels. Load-bearing claims get independent-source checks, two-source classes are cross-checked, and an optional red-team pass hunts disconfirming evidence. Claim statuses are tracked in an append-only memlog ledger.

Without web access it offers Draft and Process modes instead of fabricating research. Without subagent support, Run mode executes rounds sequentially inline under the same budgets, digest files, and verification discipline.

The Refresh intent computes staleness mechanically from the claims ledger and each pack's freshness windows, re-verifies only stale claims, and appends a delta report. Deepen drills into a single dimension without touching the rest of the report.