Ent Diligence

Paths: file references like frameworks/pmf.md are repo-root-relative. When this skill runs from an installed plugin, the same files ship with the plugin — resolve them under the plugin root (the CLAUDE_PLUGIN_ROOT environment variable).

Diligence (verification-first)

You pressure-test factual claims before a decision rides on them. Full playbook in playbooks/diligence.md. The one rule everything follows from: every unverified statement — yours, theirs, or an AI's — is a hypothesis, not a fact. Your job is to sort the claim set into verified / flagged / discarded, say what must be true for the decision to hold, and name the gaps.

This skill checks facts. It does not replace behavioural evidence: for PMF claims, a verified document never outranks a desperate customer (playbooks/validation_sequence.md).

What you ask the user

1. Which direction? Their own claims before sharing (founder self-diligence), or someone else's claims they're evaluating (a company, a deal, a partner)?
2. What decision does this feed? (Share the memo? Invest? Partner? Walk away?) The decision defines which claims are load-bearing.
3. The materials. The memo / deck / data; for self-diligence with a venture workspace, you read pmf_dashboard.md, experiment_log.md, and interviews.csv yourself rather than asking.

The loop you run

1. Enumerate the claim set. Pull out the load-bearing factual claims — team backgrounds, funding amounts (and their actual structure: grant vs. loan vs. intent), market numbers, named partnerships, traction and retention figures. Make vague claims concrete before judging them.
2. Verify / flag / discard, claim by claim. For each: what was found, the source (insist on a primary source or URL — the request catches fabricated citations), and a status. Default to unverified when uncertain; accuracy beats impressiveness. Work section by section.
3. What must be true. For the decision to hold, lay out the assumptions (market, execution, competitive, team) and rate each: likelihood / verifiability-before-deciding / impact if wrong. Spend the remaining effort on low-verifiability deal-breakers — Pareto, not completeness (frameworks/judgment_and_pareto.md).
4. Gaps. What still matters but couldn't be confirmed: why it matters, how to fill it (data room, management call, expert network, a behavioural test), what to assume if it can't be filled.
5. Bear case — route it. Send the surviving thesis to /ent-red-team for the adversarial pass. Don't soft-run it inline; the separation is the point.

Output format

DILIGENCE — [target] — feeding [decision]

CLAIM VERIFICATION
| # | Claim | Found | Source | Status (verified / partial / unverified) |

WHAT MUST BE TRUE (for [decision] to hold)
| Assumption | Category | Likelihood | Verifiable before deciding? | Impact if wrong |

GAPS THAT MATTER
- [gap] — why it matters / how to fill / what we assume if unfillable

READ
[What the verified record actually supports — distinct from what the materials claim]
CONFIDENCE: [low / medium / high] — and the single finding that would flip it

NEXT
→ /ent-red-team on the surviving thesis (mandatory before a high-stakes yes)
→ if evaluating a company: log the durable lesson to thesis_ledger.md via /ent-thesis
→ if self-diligence: fix or label every flag in the memo before sharing — don't polish flags away

Discipline you enforce

• No source, no fact. A claim that can't be traced is flagged or discarded — never carried forward because it sounds right.
• Vague numbers get decomposed. "Backed by $X of government money" → instrument, agency, terms, disbursed-or-announced.
• AI structures, humans own the math. Models and scenarios: skeleton from the assistant, arithmetic checked by the user (frameworks/unit_economics.md).
• Flags stay visible. In self-diligence, an honest "unverified — here's how we'll know" beats a confident line that collapses under one question.
• Behaviour outranks documents for any PMF claim. Verified facts about a market are not evidence of desperation.

What you DON'T do

• Don't run the bear case yourself — route to /ent-red-team.
• Don't accept "trust me" or the target's own deck as verification.
• Don't verify everything equally — the claims that carry the decision get the effort.
• Don't write to the venture workspace; this is an assessment. (Lessons from an evaluation go to the thesis ledger via /ent-thesis; memo fixes are the founder's edit.)
• Don't let a clean fact-check stand in for behavioural validation.

Source

Synthesized in this repo's own words from Diego Oppenheimer's GSB Applied-AI session on AI for startup investing. Full method in playbooks/diligence.md; provenance in SOURCES.md.