dossier

---
name: dossier
description: Use when the user needs decision-grade background research on a specific company, person, nonprofit, or gov org (meeting prep, due diligence, competitor research) that tests a stated hypothesis.
category: Research
version: 1.0.0
tools: []
---

# Dossier — Decision-Grade Entity Research

Produce hypothesis-tested research on a specific company, person, nonprofit, or government org — never a generic profile. Use `web_search` and `knowledge_base_search` as the workhorses. Deliver the finished dossier as a formatted chat report (headings, tables, bullets) rather than a file.

## The Differentiator: Hypothesis-Testing, Not Profiling

This skill refuses to be "tell me about Microsoft." Every invocation forces the user to expose their hypothesis upfront so the dossier *tests* it rather than confirms it:

> "I'm pitching Microsoft Tuesday. My hypothesis is they're consolidating AI spend on their first-party Foundry platform. Validate or disprove, and give me three conversation hooks tied to what you find."

**Not:** "Tell me about Microsoft." Skip the hypothesis question and the output degrades into a Wikipedia summary — that failure mode must be actively avoided.

## Integrity Rules

- **Source discipline.** Cite only sources actually returned by `web_search`/`knowledge_base_search` this session. Anything from background/training knowledge is labeled `[Background — verify before quoting]` and excluded from the primary-findings count.
- **Source reliability tiers.** Tag every citation primary (official filings, court records), secondary (mainstream news, trade press), or tertiary (blogs, forums). Surface the tier next to every flag and finding.
- **Tracking.** Keep a running count of searches run / sources found / sources actually cited, broken out by tier, and report it in the closing audit section.
- **On a failed or empty search**, try once more with a reworded query before moving on; after repeated failures across the board, tell the user what's missing rather than fabricating.

## Phase 1: Intake (6 questions, one at a time)

1. **Subject identity.** Exact name plus a second identifier (website, LinkedIn URL, or company+role) — refuse to proceed on an ambiguous name ("there are 47 John Smiths").
2. **Subject type.** Person / company / nonprofit / government org / other — this selects the source matrix below.
3. **Purpose.** Sales/partnership pitch, investment diligence, acquisition diligence, journalism/due diligence, job-interview prep, competitive intelligence, personal vetting, or other — purpose sets the angle, depth, and red-flag sensitivity.
4. **Hypothesis (mandatory).** What does the user already believe, and what do they want verified or disproven? This is the non-generic anchor. If the user has none, push back once ("commit to a position you can update later"); if still refused, fall back to the implicit hypothesis "what's the most surprising thing I could find?" and flag that fallback in the audit section.
5. **Depth.** A 5-minute brief (roughly 10 searches, skips network/reputation passes) or a 15-minute decision-grade dossier (deeper on every section).
6. **Sensitivities** — asked only when purpose is journalism or personal vetting: anything off-limits (medical, family, political history).

Stop after question 6 (or earlier via the dependency skips) and start research — never reopen intake mid-run.

## Phase 2: Disambiguation

Before searching, resolve the subject to one specific entity: for people, confirm a LinkedIn URL or employer+role+city; for companies, a domain or legal name+jurisdiction; for nonprofits, an EIN or legal name+state; for government orgs, an official .gov URL. If still ambiguous, halt and re-ask question 1 rather than guessing.

## Phase 3: Source Matrix by Subject Type

- **Person:** LinkedIn, personal website, X/Twitter (degrade gracefully if rate-limited), GitHub (if technical), Google Scholar (if academic), news, conference talks/podcasts.
- **Company:** official site (about/leadership/news/careers), SEC EDGAR filings (10-Ks, 10-Qs, 8-Ks for public companies), Crunchbase, news, GitHub (tech orgs), Glassdoor/Comparably sentiment, LinkedIn company page.
- **Nonprofit:** ProPublica Nonprofit Explorer (Form 990s), official site, news, GuideStar.
- **Government org:** official .gov sites, news, ProPublica for federal agencies.

## Phase 4: Hypothesis-Driven Search

Classify every search as either **supporting** (would confirm the hypothesis) or **disconfirming** (would refute it), and hold **at least 30% of the search effort to disconfirming queries** — this discipline is what makes the dossier decision-grade rather than confirmation-biased. Example, for the hypothesis "Microsoft is consolidating AI spend on Foundry": supporting queries look like "Microsoft Foundry adoption 2026"; disconfirming queries look like "Microsoft OpenAI deal renegotiation" or "Microsoft AI vendor diversification."

## Phase 5: 12-Month Activity Timeline

Default to a 12-month window (go deeper for foundational identity facts). Cover news (acquisitions, hires, departures, launches), funding/financial events, controversies/legal events, and public strategy statements. Present reverse-chronologically, each entry cited and tiered.

## Phase 6: Network + Reputation

**Network:** for companies, investors/customers/partners; for people, co-founders/advisors/employers/board roles; for nonprofits, funders/board/leadership. Keep 5-10 entries ranked by relevance to the hypothesis. **Reputation:** recent-12-month news sentiment, Glassdoor rating + a few representative reviews for companies, peer mentions for people — flag that reputation signal is noisy and tier it accordingly.

## Phase 7: Red-Flag Pass

Surface without sensationalizing: litigation (primary tier, court records), regulatory actions (SEC/DOJ/agency — primary), unusual departures (key personnel exiting within 90 days), financial red flags (going-concern notes in filings — primary), sustained negative coverage (secondary). Tier every flag.

## Phase 8: Conversation Hooks

Produce 3-5 hooks tied to actual findings, never generic talking points:

| Generic (reject) | Finding-tied (target) |
|---|---|
| "Ask about their roadmap" | "Mention their recent acquisition of X — it signals investment in vertical Y. Framing: 'Saw the X announcement — how does that change your roadmap on Y?'" |
| "Ask about hiring" | "Their VP Engineering left 3 weeks ago (LinkedIn). Framing: 'I noticed [name] moved on — what's the eng leadership plan?'" |

Each hook needs: the hook itself, the finding it's tied to (with tier), and a suggested verbatim framing the user can adapt.

## Delivering the Report

Structure the chat report as: (1) Executive Summary — one paragraph on who/why-they-matter plus the **verdict on the hypothesis** (SUPPORTED / PARTIALLY SUPPORTED / DISPROVEN / INCONCLUSIVE) and three key bullets; (2) Identity Facts; (3) Hypothesis Test — the hypothesis stated verbatim, supporting evidence (3-5 bullets), disconfirming evidence (3-5 bullets), and a 2-3 sentence verdict rationale; (4) 12-Month Timeline; (5) Network Signals; (6) Reputation Signals; (7) Red Flags (tiered); (8) Conversation Hooks; (9) Source list with tiers plus the search/source/citation counts. If a teammate has domain context that would sharpen the dossier, @mention them rather than guessing.

## Error Handling

| Situation | Response |
|---|---|
| Ambiguous subject name | Refuse to proceed; re-ask for a disambiguating identifier. |
| User won't state a hypothesis | Push back once; if still refused, use "most surprising finding" as an implicit hypothesis and flag it. |
| Subject has no public footprint | State this explicitly rather than padding with weak tertiary sources. |
| Repeated search failures | Stop, tell the user what's missing, share what was found. |