auto-diligence

Auto-diligence (full end-to-end)

Single-trigger orchestrator that runs the entire dudu chain on one deal starting from just a company name. Everything else (slug, founders, pitch, deck) is either auto-discovered or filled in from public sources. Read lib/deal.md, lib/research-protocol.md, and lib/playwright-auth.md first.

This skill is the convenience surface for "I have a company name, give me the report.html." For deeper control over each stage, invoke the layered skills directly: dudu:background-check, dudu:pmf-signal, dudu:customer-debrief.

Inputs

Required:

• Company name (free-form string)

Optional (skips the discovery step when supplied):

• --slug <kebab> — explicit deal slug; auto-derived from company name otherwise
• --website <url> — company homepage; auto-discovered otherwise
• --founders "Alice Lin, Bob Chen" — comma-separated; auto-discovered otherwise
• --pitch "<one-liner>" — one-line pitch; auto-derived from website otherwise
• --deck <path> — pitch deck file; uses manifest.pitch + L1 artifacts otherwise

Optional flags:

• --auto-call — after Layer 2, drive callagent against the top-N candidates from outreach.md and save results to deals/<slug>/calls/. Default off.
• --call-n <int> — number of candidates to call when --auto-call is set (default 3, max 10)
• --no-render — stop after MEMO.md; skip report.html
• --force — propagate --force to every sub-skill, overwriting existing artifacts

Pre-flight

1. Refuse to run if the working directory is not a dudu repo root (no lib/deal.md present).
2. If --slug was given, validate it: kebab-case, lowercase, no leading underscore. Reject the run with a clear error otherwise.
3. If deals/<slug>/report.html already exists and --force was not passed, print Auto-diligence already complete at deals/<slug>/report.html. Pass --force to re-run. and stop.

Stage 1 — Discovery (skip when all overrides supplied)

Goal: produce a manifest.json seed with company, slug, pitch, founders[], and optionally website.

If --slug, --website, --founders, and --pitch are all supplied, skip discovery entirely and proceed to Stage 2.

Otherwise:

1. Resolve the company. Web search the company name with the WebSearch tool. Fetch budget: 1 search.
2. Pick the canonical site. From the top hits, identify the official domain (prefer .com/.io/.ai apex over LinkedIn / Crunchbase / press articles). If two or more candidates look equally plausible, pause and ask the user to disambiguate — do not guess.
3. Fetch the homepage. WebFetch budget: 3 (homepage, /about, /pricing). Extract:
   • The one-line value proposition (becomes pitch if --pitch not supplied)
   • Founder names from /about or footer (becomes founders[] if --founders not supplied)
4. LinkedIn / Crunchbase fallback (only if founders not found on the site). Authenticated browser via Playwright — see lib/playwright-auth.md. Resolve the company's LinkedIn page; read founder names from "People → Founders" or the company "About" block. Fetch budget: 2 navigations.
5. Slug derivation. If --slug was not supplied, kebab-case the company name (lowercase, ASCII only, drop punctuation, max 30 chars). Examples:
   • "LedgerLoop" → ledgerloop
   • "Twenty-Three Capital" → twenty-three-capital
   • "Müller & Söhne GmbH" → muller-sohne
6. Confirmation. Print the resolved seed back to the user (slug, company, founders, pitch, website) and stop only if anything is uncertain — flag specific fields with ? and ask for confirmation. If everything looks confident, proceed without pausing.

Total Stage-1 budget: ~6 web fetches + 1 search. Pure secondary research; no LLM-heavy synthesis (Layer 2 owns that).

Stage 2 — Initialize the deal

1. If deals/<slug>/ does not exist, create it. Write manifest.json per the schema in lib/deal.md. Include the discovered (or supplied) website under a top-level website field if known.
2. If --deck <path> was supplied, copy the deck to deals/<slug>/inputs/deck.<ext>. If the deck is plain text (e.g., pasted), write to inputs/deck.md.
3. No deck is required; Stage 0 of dudu:pmf-signal falls back to manifest.pitch + L1 artifacts.

Stage 3 — Layer 1 (dudu:background-check)

Invoke dudu:background-check with the resolved slug and supplied/discovered inputs. Propagate --force if supplied. Wait for completion: deals/<slug>/background.md must exist on success.

If background-check fails, surface its error verbatim and stop. Do not proceed to Layer 2.

Stage 4 — Layer 2 (dudu:pmf-signal)

Invoke dudu:pmf-signal on the same slug. Propagate --force if supplied. Wait for completion: deals/<slug>/pmf-signal.md, outreach.md, and personas/verdicts.yaml must all exist on success.

Stage 5 — Optional automated screener calls

If --auto-call was passed:

1. Read the top --call-n (default 3) candidates from deals/<slug>/outreach.md (sorted by warm-path quality).
2. For each candidate, follow the callagent recipe documented in skills/pmf-signal/SKILL.md Stage 5: write task-screener-<id>.md, context.md, and screener-schema.json under deals/<slug>/calls/.
3. Iterate once with callagent simulate against each task to sanity-check the brief; abort the auto-call stage if any simulate run produces a bad transcript. Surface the issue and let the user adjust manually.
4. Run callagent run for each candidate. Save audio recordings to deals/<slug>/calls/<id>.<ext> (callagent's native output) and result JSON to deals/<slug>/calls/<id>.json.
5. If callagent is not on PATH, print callagent not found — skipping auto-call. Install callagent or run real interviews and save transcripts to deals/<slug>/inputs/. and continue to Stage 6 with whatever calls/ already contains.

If --auto-call was not passed: skip this stage. The user is expected to either run their own interviews and save transcripts to deals/<slug>/inputs/, or invoke this skill again later with --auto-call.

Stage 6 — Customer debrief (when material exists)

Check whether interview material is present:

• deals/<slug>/inputs/ contains any *.md, *.txt, or *.vtt file beyond deck.*, OR
• deals/<slug>/calls/ contains any *.json callagent result file

If yes: invoke dudu:customer-debrief on the slug. It writes customer-discovery.md. Propagate --force.

If no: skip this stage. The downstream MEMO will note that the customer-discovery section is pending.

Stage 7 — Stitch MEMO.md

Read every artifact under deals/<slug>/ and produce deals/<slug>/MEMO.md using the same shape as the deprecated dudu:diligence wrapper:

# Due diligence report: <Company>

**Deal:** <slug>
**Generated:** <ISO timestamp>

## Founder background

[For each founder, summarize from `founder-<name>.md` in 3-5 bullets. Include prior managers, coworkers, collaborators, co-founders, board members, or senior operating partners when publicly identified. Show only intentionally public professional/social contact links such as LinkedIn, X/Twitter, GitHub, personal site, or company bio. Do not include private emails, phone numbers, or inferred/private relationship data. Link to the full dossier.]

## Problem and product

[Summary from `market-context.md` (4-6 sentences) + the strongest pattern + the most valuable contradiction. Link to the full file.]

## PMF signal & claim verification (calibrated prior + cross-artifact + external)

[Headline read from `pmf-signal.md`. Top 5 rows of the consolidated claim ledger (worst-news-first ordering). The 1 strongest cluster pattern. Explicit Stance B disclaimer for the persona-reaction rows; cross-artifact and external-evidence rows do not need the same disclaimer because they triangulate against actual evidence. List of `requires-data-room` flags for the VC to follow up on.]

## Customer signal

[Summary from `customer-discovery.md` if present; otherwise note "Pending — no interviews recorded yet." Quote 2-3 strongest verbatims when present. Flag any persona contradictions explicitly.]

## Competitive landscape

[Summary from `competitive-landscape.md`. Top 3 direct competitors. Incumbent verdict. Moat verdict.]

## Market sizing

[Wedge TAM range, expansion TAM range, comparison to founder claim.]

## Cross-artifact synthesis

[Surfaces contradictions ACROSS artifacts. e.g.: "Founder claims engineering teams are the buyer (deck p.3), but customer interviews showed product managers driving the purchase (interview-2)."]

## Decision inputs

[Neutral synthesis of the evidence gathered. Do not recommend Pass, Watch, Pursue, invest, decline, or any other decision. State the strongest confirming evidence, the strongest contrary evidence, unresolved information gaps, and specific facts the user may want to verify before making their own decision.]

## Source artifacts

- founder-<name>.md
- market-context.md
- pmf-signal.md
- outreach.md
- customer-discovery.md (if present)
- competitive-landscape.md
- market-sizing.md

Stage 8 — Render report.html

Unless --no-render was passed:

1. Run python3 scripts/render-report.py deals/<slug>.
2. The renderer auto-discovers any deals/<slug>/calls/*.{mp3,wav,m4a,webm,ogg} and deals/<slug>/inputs/*.{mp3,wav,m4a,webm,ogg} files and embeds them as <audio controls> tags inside the Customer Signal section. Audio uses relative paths — report.html and the calls/ / inputs/ subdirectories must travel together.
3. If the script exits non-zero, surface the stderr but do not block — MEMO.md is still useful on its own.
4. If python3 is not on PATH, print report.html skipped — install Python 3 or run python3 scripts/render-report.py deals/<slug> manually.

Stage 9 — Print final paths

Print:

Auto-diligence complete.

  deals/<slug>/MEMO.md
  deals/<slug>/report.html   (if rendered)

Layered artifacts:
  deals/<slug>/background.md
  deals/<slug>/pmf-signal.md
  deals/<slug>/outreach.md
  deals/<slug>/customer-discovery.md   (if interviews ran)

If any sub-skill was skipped or failed, list the missing artifact and the recommended next command (dudu:customer-debrief, etc.).

Re-runnability

Each underlying skill checks its own artifacts and skips when they exist (unless --force). Re-running auto-diligence after adding transcripts to inputs/ or audio to calls/ will:

• Skip Stages 3–4 (already done, unless --force)
• Run Stage 6 if customer-discovery.md is missing and material is now present
• Re-stitch MEMO.md and re-render report.html

What this skill does NOT do

• Does not write any per-deal artifact directly. Every file under deals/<slug>/ comes from a sub-skill or the renderer.
• Does not emit personas — that namespace is owned exclusively by dudu:pmf-signal.
• Does not bypass dudu:pmf-signal's preflight gate. If Layer 1 is incomplete, Layer 2 will refuse and this skill stops.
• Does not run real human interviews. --auto-call drives callagent (a separate tool); for real interviews the VC saves transcripts to inputs/ and re-invokes.