GDD — Get Diligence Done

An agentic workflow system for commercial due diligence: scope the deal, lock the taxonomy, run the modules in fresh-context agents, triangulate, red-team, tell the story. A deterministic engagement spine with open-ended analytical leaves, persistent state in markdown plus a machine lock, and a verification gate whose honest claim is consistency and traceability, not truth.

Claude Code, Codex, and Antigravity CLI are supported runtimes; other runtimes are community ports on demand. See docs/design.md for the full design and docs/install.md for every install path.

Install

One command (recommended). Requires Node.js >= 20.

npx get-diligence-done --claude --global

That projects GDD into ~/.claude, so /gdd: works from any deal folder — the tool's normal per-engagement usage pattern. Swap --claude for --codex or --antigravity to target those runtimes. The installer is a single file with no dependencies; there's nothing to build.

From source (to read the code first, or to hack on it):

git clone https://github.com/Navis-Advisory/gdd.git
cd gdd
node bin/install.js --claude --global

npm test verifies the install round-trip and reference integrity. Full details in docs/install.md.

Once installed, GDD runs out of ~/.claude; engagement state lives in a .diligence/ folder inside each deal folder, never in the install. To update: re-run npx get-diligence-done@latest --claude --global, or /gdd:update from inside the runtime.

Claude Code plugin (secondary). /plugin marketplace add Navis-Advisory/gdd then /plugin install gdd@gdd works, but registers at local scope pinned to the project you ran it in — /gdd: won't follow you into other deal folders, and it can't reach Antigravity CLI. Use the CLI install above unless you're staying in one project folder for the whole engagement.

Claude Cowork. Install GDD — Get Diligence Done from the plugin browser once it's listed in the community catalog. The intake commands render as fill-in-the-blank forms.

Why we built this

We run commercial due diligence for a living. Pointing a general-purpose AI at a diligence question rarely fails by being obviously wrong — it fails by being confidently unverifiable: a clean-sounding market size with no traceable source, a thesis nothing ever tried to kill. That's the failure a good engagement manager exists to prevent. GDD encodes that discipline — every number cited, every market sized two independent ways, every thesis red-teamed — as a workflow, so the rigor is structural rather than hoped-for.

It's a scalpel, not an autopilot: trust the execution, verify like the EM.

The loop

scope-deal → hypothesis-tree → workplan
  → size-market / map-competitors / probe-customers / assess-moat / scan-risks
    (modules, fresh-context, independent)
  → triangulate (verification gate) → red-team → storyline

• scope-deal — interview the deal into an engagement brief and a locked taxonomy (the segment definitions everything downstream must obey).
• hypothesis-tree — decompose the investment thesis into a MECE tree of falsifiable hypotheses, one brief per module.
• workplan — lay the modules across the engagement timeline, dependency-ordered.
• modules — the analytical work, each in a fresh-context agent: size-market, map-competitors, probe-customers, assess-moat, scan-risks.
• triangulate — the verification gate: consistency and traceability checks across every finding. A FAIL blocks the storyline.
• red-team — the strongest case against the thesis; surviving kills go CONTESTED.
• storyline — synthesize the surviving findings into a pyramid- principle storyline.

Why it works

Naive AI diligence fails in three predictable ways. GDD is built around closing each one:

• Untraceable claims. Every finding carries a citation to a tiered source; the verifier fails the gate when a load-bearing number doesn't trace back to one.
• Single-source sizing. The market is sized top-down and bottom-up by two agents that never see each other's work, then reconciled — an unreconciled gap is reported and named, never averaged away.
• Unfalsified theses. A red-team stage argues the strongest case against the thesis before the story is told; a killed finding goes CONTESTED and can't carry a key-line claim.

The verifier's promise is consistency and traceability — it confirms the TAM math checks out, top-down and bottom-up reconcile, every number traces to a cited source, and the segment tree is MECE. It never claims the market is actually $X.

Layout

| Path | What | |---|---| | .claude-plugin/ | Plugin manifest (plugin.json) and private marketplace (marketplace.json) — the no-terminal install surface | | bin/install.js | Single-file CLI installer; runtimes data-defined in runtime-catalog.json (Claude Code, Codex, Antigravity CLI) | | commands/ | The 17 slash commands — flat .md files (plugin skills → /gdd:*), thin wrappers that delegate to workflows | | agents/ | The 10 subagents (scoper, planner, researcher, independent sizers, analyst, verifier, red-teamer, storyliner, librarian) | | gdd-core/workflows/ | The real mechanics each command routes into | | gdd-core/templates/ | Engagement state artifacts (brief, taxonomy, ledger, reports…) | | gdd-core/references/ | The verification check registry, source hierarchy, sizing methods | | docs/design.md | Design overview: hierarchy, artifacts, commands, agents, verification semantics |

User engagement state lives in .diligence/ inside a deal folder (never in this repo — gitignored): ENGAGEMENT.md, TAXONOMY.md + state.json taxonomy lock, LEDGER.md, SOURCES.md, WORKPLAN.md, STATE.md, modules/*/, reports/.

Contributing

Issues are open and welcome — bug reports and feature ideas both. Pull requests are restricted to collaborators (this repo is a mirror). See CONTRIBUTING.md and ROADMAP.md.

Inspiration

GDD takes its name and skeleton in analogy with GSD (MIT), whose command-workflow idea transfers cleanly beyond software development. Machinery here is written fresh against that pattern.

License

Apache-2.0. Copyright 2026 Navis Advisory.