/cs-dossier

disable-model-invocation: true
name: "cs-dossier"
description: "/cs-dossier <entity> — Decision-grade entity research with mandatory hypothesis-testing. 6-Q grill-me intake (Q4 hypothesis MANDATORY) → ≥30% disconfirming search budget → 9-section .docx with verdict (SUPPORTED/PARTIALLY/DISPROVEN/INCONCLUSIVE) + 3-5 finding-tied conversation hooks."

/cs-dossier — Decision-Grade Entity Research

**Command:** `/cs-dossier <entity>`

The `cs-dossier` persona produces a hypothesis-tested research dossier on a specific company, person, nonprofit, or government org — **NOT** a generic profile.

When to Run

• Sales meeting / partnership pitch (need conversation hooks tied to specifics)
• Investment / acquisition diligence
• Journalism / personal vetting (with sensitivity exclusions)
• Job interview prep
• Competitive intelligence

When NOT to Run

• Generic curiosity ("what does this company do?") → search the web yourself
• Quick lookup → faster to just google
• No hypothesis to test → the skill refuses, by design

Non-Generic by Design

The skill refuses to be a Wikipedia summary. Q4 (your hypothesis) is **mandatory** — without it, the dossier confirms what you already think and is worthless for decisions.

Forcing Intake (6 Questions, One at a Time)

| Q | Asks | Notes | |---|---|---| | Q1 | Subject identity (name + disambiguating identifier) | refuses ambiguous names | | Q2 | Subject type: person / company / nonprofit / gov org / other | forcing choice — drives source matrix | | Q3 | Purpose: sales / investment / acquisition / journalism / interview / competitive / vetting / other | forcing choice — drives angle + sensitivity | | Q4 | **Hypothesis (MANDATORY)** — what you already believe + want to verify/disprove | non-skippable; pushed back once if refused | | Q5 | Depth: 5-min brief or 15-min decision-grade dossier | forcing choice | | Q6 | Sensitivities to exclude | conditional — only if Q3 ∈ {journalism, personal vetting} |

Stop condition: after Q6 (or earlier with skips), commit and start Phase 2. Never re-open.

What You Get

After all phases:

dossier_<entity-slug>_<YYYY-MM-DD>.docx

9 sections:
1. Executive Summary (verdict: SUPPORTED/PARTIALLY/DISPROVEN/INCONCLUSIVE + 3 must-know)
2. Identity Facts Table (founded/born, location, size, role, affiliations; sourced + tiered)
3. Hypothesis Test (verbatim hypothesis + supporting evidence + disconfirming evidence + verdict)
4. 12-Month Activity Timeline (news, hires, departures, products, controversies)
5. Network Signals (collaborators / investors / customers / advisors)
6. Reputation Signals (sentiment, Glassdoor, peer mentions)
7. Red Flags + Hidden Patterns (litigation, departures, financials, tiered)
8. Conversation Hooks (3-5 finding-tied hooks with framing)
9. Source Provenance + Audit Log (per-source tier + search summary + counts)

Hypothesis-Testing Discipline

**≥30% of search budget allocated to disconfirming queries.** This is the non-negotiable differentiator from a generic profile.

Example for hypothesis "Microsoft is consolidating AI spend on Foundry":

| Query type | Example | |---|---| | **Supporting** (would confirm) | "Microsoft Foundry adoption 2026" | | **Supporting** | "Microsoft AI infrastructure consolidation" | | **Disconfirming** (would refute) | "Microsoft OpenAI deal renegotiation" | | **Disconfirming** | "Microsoft AI vendor diversification" | | **Disconfirming** | "Microsoft third-party model partnerships 2026" |

`skills/dossier/scripts/disconfirming_evidence_balance.py` enforces the ratio. Halts at <30% and prompts more disconfirming queries.

Source Reliability Tiering

Every fact in the DOCX tagged with tier (primary / secondary / tertiary):

| Tier | Examples | |---|---| | **Primary** | SEC EDGAR filings, court records, official .gov sites, company official website | | **Secondary** | Mainstream news (NYT, WSJ, Reuters), trade press (TechCrunch, The Information) | | **Tertiary** | Blogs, forums (Reddit, HN), Glassdoor, social media |

`skills/dossier/scripts/source_tier_classifier.py` does this from URL.

Discipline (Research-Pack Convention)

• **One intake Q per turn.** Never bundle.
• **Q4 mandatory.** Push back once; fall back to "most surprising finding" implicit hypothesis with flag.
• **≥30% disconfirming.** Enforced by tool.
• **Sequential search.** WebSearch + WebFetch sequential, 1 q/sec etiquette.
• **Source discipline.** Cite only session results. Training knowledge labeled `[Background — verify before quoting]`, excluded from counts.
• **Three-count + tier.** Sent / received / cited + per-tier breakdown.
• **Subject disambiguation before Phase 3.** Refuse ambiguous names.
• **Sensitivity exclusions honored.** If Q6 excluded "medical history", don't surface even if found.
• **Conversation hooks finding-tied.** Generic hooks ("ask about their roadmap") rejected.
• **BYOK MCP flagged in audit.** Crunchbase / Pitchbook usage surfaced.

Trigger Phrases

• "research [company]"
• "dossier on [person/company]"
• "background check on [entity]"
• "prep me for a meeting with [person/company]"
• "due diligence on [company]"
• "what should I know about [entity]"
• "research [person] before I [meet/hire/invest]"
• "competitor research on [company]"
• "investor diligence [company]"
• "interview prep for [company]"

Anti-Patterns Rejected

• Producing a dossier without forcing Q4 hypothesis
• <30% disconfirming search budget (confirmation bias)
• Batching intake questions
• Accepting ambiguous subject names
• Generic conversation hooks ("ask about their roadmap")
• Sensationalizing red flags (tier them, don't editorialize)
• Skipping source-reliability tier on flags
• Fabricating coverage when LinkedIn blocked
• Using BYOK MCP without flagging in audit
• Including sensitive topics user excluded (Q6)
• Confirmation-biased verdict ("SUPPORTED" without engaging with disconfirming evidence)

Related

• Agent: [`cs-dossier`](../agents/cs-dossier.md)
• Skill: