dossier – skill Tài liệu & Office cho agent AI

Dossier — Decision-Grade Entity Research

> Portability: Requires WebSearch + WebFetch, Node.js with docx package, and optionally bash_tool + curl for free APIs (SEC EDGAR, GitHub, ProPublica). BYOK MCPs (LinkedIn, Crunchbase, Apollo, Pitchbook, SimilarWeb) are optional enhancements. Works in Claude Code CLI natively.

Non-Generic Framing — The Differentiator

This skill is decision-grade entity research with hypothesis-testing. It refuses to be "tell me about Microsoft". Every invocation forces the user to expose their hypothesis upfront (Q4) so the dossier *tests* it rather than confirms it.

The use case shape:

> "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."

The forcing Q4 — the hypothesis question — is the non-generic anchor. Skip it and the skill produces a Wikipedia summary.

See [references/hypothesis_testing_discipline.md](references/hypothesis_testing_discipline.md) for the canon.

Agent Integrity Rules (Research-Pack Convention)

Locked verbatim per PR #657 audit.

• Execution discipline. Sequential search calls. WebSearch + WebFetch have looser rate limits than Consensus but still apply 1 q/sec etiquette. Confirm response received before next call.
• Source discipline. Cite only sources returned by this session's tool calls. Wikipedia / training knowledge labeled [Background — verify before quoting] and excluded from primary findings count.
• Three-count tracking. Queries sent / sources received / sources cited. Plus per-tier breakdown (primary / secondary / tertiary) unique to dossier. Surfaced in audit log.
• Retry policy. On failure → wait 3s → retry once → log. After 3 consecutive failures: stop, alert user.
• Source reliability tier. Each citation tagged primary (official, SEC, court records) / secondary (mainstream news, trade press) / tertiary (blogs, forums). DOCX surfaces tier on every flag.

Phase 1: Grill-Me Intake (6 forcing questions, one at a time)

Q1 (root) — Subject identity

> Who is the subject? Give me the exact name and, if a company, the website or LinkedIn URL. If a person, their LinkedIn URL or a unique identifier (company affiliation + role).

>

> *Why I'm asking:* Disambiguation. There are 47 John Smiths. There are three companies called "Atlas". I need a specific entity to research.

If user gives only a name, push for a second identifier. Refuse to proceed on ambiguous names.

Q2 (depends on Q1) — Subject type

> What kind of subject is this? Pick one: person / company / nonprofit / government org / other.

>

> *Why I'm asking:* Different source matrices apply. For people I check LinkedIn, GitHub, Scholar, news; for companies I check SEC EDGAR (if public), Crunchbase, news, GitHub for tech orgs; for nonprofits I check Form 990s on ProPublica.

Forcing choice. "Other" requires a one-line description.

Q3 (depends on Q2) — Purpose

> What are you preparing for? Pick one:

>

> 1. Sales meeting / partnership pitch

> 2. Investment diligence

> 3. Acquisition diligence

> 4. Journalism / due diligence

> 5. Job interview prep

> 6. Competitive intelligence

> 7. Personal vetting (date, hire, business partner)

> 8. Other (specify)

>

> *Why I'm asking:* The purpose dictates the angle, the depth, and the red-flag sensitivity. Sales prep needs conversation hooks. Investment diligence needs traction signals. Personal vetting needs careful sensitivity boundaries.

Q4 (depends on Q3) — Hypothesis — MANDATORY

> What's your hypothesis going in? What do you already believe about this subject, and what do you want to verify or disprove?

>

> *Why I'm asking:* This is the critical question. A dossier that just confirms what you already think is worthless. By stating your hypothesis upfront, I can search for evidence that would *disprove* it as well as evidence that supports it — and give you a verdict you can actually use.

>

> Examples:

> - "I believe Microsoft is consolidating AI spend on first-party Foundry. Verify or disprove."

> - "I think the CEO is over their head — too much TAM talk, no traction. Test that."

> - "I believe this nonprofit's overhead ratio is sketchy. Check the 990s."

> - "I think this person is technical enough to handle a CTO role. Verify."

MANDATORY. If user says "I don't have one", push back once: "Then guess. Commit to a position you can update later. The dossier needs a hypothesis to test, otherwise it's a generic profile and won't help you make a decision."

If still refused: fall back to implicit hypothesis "what's the most surprising thing I could find?" and flag the fallback in audit log.

This question is the non-generic anchor. Skip it and the skill becomes a Wikipedia summary.

Q5 (depends on Q3) — Depth

> Time horizon: 5-minute brief or 15-minute decision-grade dossier?

>

> *Why I'm asking:* Brief mode caps at ~10 searches and skips the network + reputation passes. Decision-grade goes deeper on every section. Pick based on how much skin you have in this decision.

Forcing choice.

Q6 (asked only if Q3 ∈ {journalism, personal vetting}) — Sensitivities

> Anything sensitive to exclude? E.g., personal medical, family details, political history, or specific topics off-limits?

>

> *Why I'm asking:* Some research contexts have ethical constraints. I'd rather know upfront than surface something you'd never share.

Skip for sales/investment/acquisition/competitive intel (low sensitivity); ask for journalism/personal vetting (high sensitivity).

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

Phase 2: Subject Disambiguation

Before Phase 3, resolve the subject to a specific entity:

• For people: confirm LinkedIn URL OR (employer + role + city)
• For companies: confirm domain OR (legal name + incorporation jurisdiction)
• For nonprofits: confirm EIN OR (legal name + state)
• For government orgs: confirm official .gov URL

If still ambiguous after Q1 push-back: halt and re-ask Q1 with disambiguating identifiers. Refuse to proceed.

Phase 3: Source Matrix Selection

Routed by Q2 subject type. See [references/subject_type_source_matrix.md](references/subject_type_source_matrix.md) for the full canon.

Person

• LinkedIn (manual fetch or LinkedIn MCP if BYOK)
• Personal website
• Twitter/X (rate-limited; degrade gracefully)
• GitHub (if technical subject)
• Google Scholar (if academic)
• News (WebSearch + WebFetch)
• Conference talk transcripts, podcasts (WebSearch)

Company

• Official website (about, leadership, news, careers)
• SEC EDGAR (free API; 10-Ks, 10-Qs, 8-Ks for public co's)
• Crunchbase free tier (or Crunchbase MCP if BYOK)
• News (WebSearch + WebFetch)
• GitHub (for tech orgs)
• Glassdoor + Comparably (sentiment; degrade gracefully if scraping blocked)
• LinkedIn company page

Nonprofit

• ProPublica Nonprofit Explorer (free; Form 990s)
• Official website
• News
• GuideStar (if accessible)

Government org

• Official .gov sites
• News
• ProPublica (for federal agencies)

If a paid MCP is connected (Apollo, Pitchbook, SimilarWeb), use it but mark findings as BYOK-sourced in the audit log.

Phase 4: Hypothesis-Driven Search

Every Phase 4 search MUST be classified as either:

• Supporting evidence (confirms hypothesis), OR
• Disconfirming evidence (would refute hypothesis)

≥30% of search budget allocated to disconfirming queries. Enforced via scripts/disconfirming_evidence_balance.py.

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

• Supporting: "Microsoft Foundry adoption 2026", "Microsoft AI infrastructure consolidation"
• Disconfirming: "Microsoft OpenAI deal renegotiation", "Microsoft AI vendor diversification", "Microsoft third-party model partnerships 2026"

This is what makes the dossier decision-grade rather than confirmation-biased.

For each search:

• Record via citation_tracker.py with classification (supporting / disconfirming)
• Apply source tier from source_tier_classifier.py to each result URL

Phase 5: 12-Month Activity Timeline

Default 12-month window for activity timeline; deeper for foundational identity.

Categories:

• News (acquisitions, hires, departures, product launches)
• Funding rounds / financial events
• Controversies / legal events
• Public statements / strategy shifts

Reverse chronological. Each entry hyperlinked + tiered.

Phase 6: Network + Reputation Signals

Network

• Companies: investors (in/out), customers (named), partners
• People: co-founders, advisors, mentors, employers, board roles
• Nonprofits: funders, board, leadership

5-10 entries, ranked by relevance to hypothesis.

Reputation

• Sentiment from new

… (nội dung đầy đủ có trong gói tải về)