Dossier

# Create a "Dossier" AI Skill

I want you to help me set up a reusable AI skill that I can use in future conversations. Read the complete skill definition below, then help me install it.

## Complete Skill Definition

# Intelligence Dossier

A research skill for producing structured intelligence dossiers — the
kind of document a CEO reads before a meeting, a PM reads before
market-entry, or an investor reads before due diligence.

## When to use this skill

- **Deal-prep dossier** before a major partnership / acquisition meeting
- **Executive briefing** ahead of a board, customer, or regulator meeting
- **Market-entry analysis** for a new geography or vertical
- **Due-diligence overview** for investment or M&A consideration
- **Competitor profile** in depth
- **Person dossier** ahead of executive recruiting or board engagement
- **Repo research note** — one technical question answered from primary sources, saved as a cited markdown file in the repository

## Inputs the advisor expects

- Subject (company / person / market / domain)
- Purpose (deal-prep, due diligence, briefing — affects depth + emphasis)
- Audience (exec, board, working team)
- Timeline / deadline
- Known starting sources
- Sensitive areas to dig into

## Clarify First

Before building the dossier, settle the inputs below. When one is missing or vague and would change the result, ask about it; when it wouldn't, use a sensible default and say which one you used.

- **Subject + subject type (company / person / market):** selects the outline template and which sections apply.
- **Purpose (deal-prep, due diligence, briefing, market-entry):** drives depth, emphasis, and the Implications section.
- **Audience (exec, board, working team):** sets altitude and length of the executive summary.
- **Key decision or risk to inform:** drives the Risks + Open Questions and Recommendations sections.

Ask about the two or three inputs that matter most for this request, not the whole list. If the user wants a result straight away, produce it and state your assumptions at the top.

## Quick Start

```bash
python3 research/dossier/scripts/dossier_outline_generator.py --subject-type company --purpose deal-prep --format markdown
python3 research/dossier/scripts/research_note_checker.py --input research/dossier/assets/sample_research_note.md
```

## Workflows

### Workflow 1 — Generate dossier outline

1. Specify subject type + purpose.
2. Run `dossier_outline_generator.py` to produce a structured outline
   tailored to subject + purpose.
3. Assign owners + research targets per section.

```bash
python3 dossier/scripts/dossier_outline_generator.py \
  --subject-type company --purpose deal-prep --format markdown
```

### Workflow 2 — Validate source triangulation

1. Capture claims with supporting sources.
2. Run `source_triangulation_validator.py` to check each claim has
   multiple independent supporting sources + source reliability.
3. Flag thinly-sourced claims for additional research.

```bash
python3 dossier/scripts/source_triangulation_validator.py \
  --input claims_with_sources.json --format markdown
```

### Workflow 3 — Separate facts from inferences

1. Capture dossier statements.
2. Run `fact_inference_separator.py` to classify each statement and
   flag unsupported inferences.

```bash
python3 dossier/scripts/fact_inference_separator.py \
  --input dossier_statements.json --format markdown
```

### Workflow 4 — Repo research note [RECOMMENDED]

A narrow dossier for engineering work: one technical question (how an API behaves at its
limits, what a spec requires, which version changed a behaviour) answered from primary
sources and committed as a cited markdown note next to the work it informs. Full method:
`references/repo-research-notes.md`.

1. Write the question as one sentence, plus the decision it unblocks.
2. Hand the reading to a background agent or sub-agent with fresh context, so the current
   session keeps working and the researcher is not anchored on the current design. Give it
   the question, where to look first, the source rules, the save location and a time box.
3. Trace every claim to the party that owns the behaviour: the provider's docs and
   changelog, the specification text, the library's source at the pinned version, our code
   at a named commit. Secondary sources corroborate only; a claim supported only by them
   goes to Open questions.
4. Fill `assets/research_note_template.md`: Answer first, then Findings (facts, each cited
   `[n]`), Inference (conclusions, naming the findings they rest on), Open questions,
   numbered Sources with an access date, version or commit.
5. Save it where the repo keeps such notes (`docs/research/`, `docs/notes/`, the ADR folder);
   if there is no convention, use `docs/research/<slug>.md` and say where it went. Link it
   from the spec, ticket or PR it informs.
6. Validate with the checker, then spot-check the citations behind the Answer.

```bash
python3 research/dossier/scripts/research_note_checker.py \
  --input docs/research/<slug>.md --max-uncited 0.1 --format text
```

## Decision frameworks

### The dossier hierarchy

A useful structure for any dossier:

1. **Executive summary** (1 page, lead with takeaway)
2. **Subject overview** (facts: what they are)
3. **Context** (market, history, environment)
4. **Capabilities + assets** (what they can do)
5. **People + leadership** (who runs it)
6. **Performance + trajectory** (numbers, trends)
7. **Relationships + ecosystem** (who they're with)
8. **Risks + open questions** (what we don't know)
9. **Implications + recommendations** (so what)
10. **Sources + methodology** (how we know)

## Fact vs inference discipline

Three categories per statement:

| Category | Definition | Example |
|----------|------------|---------|
| **Fact** | Verifiable, sourced | "Founded 2018; HQ in Chicago" |
| **Inference** | Reasoned from facts | "Likely targeting enterprise segment based on hiring pattern" |
| **Speculation** | No supporting evidence | "Might pivot to AI next year" |

A trustworthy dossier separates these clearly. Mixing them = loss of credibility.

### Source reliability scoring (Admiralty Code adapted)

| Reliability | Code | Description |
|-------------|------|-------------|
| Completely reliable | A | Established, history of completely reliable info |
| Usually reliable | B | History of mostly reliable info |
| Fairly reliable | C | History of reliable info with notable errors |
| Not usually reliable | D | Limited history; mixed accuracy |
| Unreliable | E | Known for inaccurate info |
| Cannot be judged | F | New / unknown source |

| Information credibility | Code | Description |
|--------------------------|------|-------------|
| Confirmed | 1 | Confirmed by other independent sources |
| Probably true | 2 | Not confirmed; consistent with other info |
| Possibly true | 3 | Not confirmed; reasonable but unsupported |
| Doubtful | 4 | Inconsistent with other info |
| Improbable | 5 | Contradicted by other info |
| Cannot be judged | 6 | New info; no validation possible |

A "B-2" rated claim is "usually reliable source, probably true" — workable. An "F-6" claim is "unknown source, unverified" — barely worth including.

### Triangulation principle

For each significant claim:
- **1 source:** anecdotal; flag explicitly
- **2 independent sources:** workable (most dossier claims should reach this)
- **3+ independent sources:** confirmed; safe to assert

"Independent" means not derived from the same underlying source. Two news
articles citing the same press release ≠ 2 independent sources.

## Common engagements

### "Build a dossier on company X before our acquisition meeting"
1. Run outline generator (subject=company, purpose=deal-prep).
2. Pull: financials, leadership, products, customers, IP, tech stack,
   regulatory posture.
3. Identify red flags: undisclosed litigation, key person dependencies,
   customer concentration, regulatory risk.
4. Recommendations: questions to ask in meeting; deal structure implications.

### "Executive briefing for senator's office meeting"
1. Run outline (subject=person/organization, purpose=briefing).
2. Pull: voting record, recent statements, committee assignments,
   donor profile, alignment with our position.
3. Anticipate likely questions; prepare positions.

### "Market-entry analysis for [country]"
1. Run outline (subject=market, purpose=market-entry).
2. Pull: market size, growth, competitive landscape, regulatory,
   distribution, cultural / business norms, talent.
3. Compare entry options (direct, partner, acquisition).

## Anti-patterns to avoid

### The Blog-Sourced Fact
**Mistake:** A research note or dossier claim rests on a tutorial, Q&A answer or summary article.
**Why it happens:** Secondary write-ups rank higher in search and read faster than the owner's documentation.
**Instead:** Follow the claim to the party that owns it — the vendor's reference, the spec, the source code at the pinned version — and cite that with a date or version. Keep the secondary source only to flag that it disagrees.

### The Anchored Researcher
**Mistake:** The same session that designed the feature researches whether the design is valid.
**Why it happens:** The context is already loaded, so it seems efficient.
**Instead:** Give the question to a fresh-context agent or person. A researcher holding the design looks for support; one holding only the question looks for the answer.

### The Undated Snapshot
**Mistake:** A note or dossier with no date and no source versions, still trusted two years later.
**Why it happens:** Facts feel permanent when they are written down.
**Instead:** Date the document, version every source, and set a review trigger (a dependency upgrade, a vendor docs change, a fixed date).

Also avoid:

- **Mixing facts + inferences without labels.** Reader can't calibrate trust.
- **Single-source claims presented as confirmed.** Anecdote dressed as data.
- **Unsourced "everyone knows" claims.** Often turn out wrong.
- **Padding with low-relevance facts.** Bloated dossier loses signal.
- **Burying risks at the end.** Risks should be surfaced upfront.
- **No update mechanism.** Stale dossier on important subject = bad decisions.
- **Adversarial language about subject.** Bias erodes credibility.

## Scope & Limitations

**In Scope:**
- Dossier outlines tailored to subject type (company, person, market) and purpose (deal-prep, due diligence, briefing, market entry) with `dossier_outline_generator.py`
- Source triangulation checks with `source_triangulation_validator.py`: single-source claims, low-reliability sources and sources that are not independent
- Fact, inference and speculation labelling with `fact_inference_separator.py`, flagging unsupported inferences and over-confident speculation
- The ten-part dossier hierarchy and source rating with an adapted Admiralty Code
- Repo research notes: one technical question researched from primary sources (typically by a background agent) and written as a cited markdown note in the repository, checked with `research_note_checker.py`

**Out of Scope:**
- Collecting dossier information. The dossier tools structure and check research the user gathers; in research-note mode the agent does the reading, but the checker itself never fetches sources.
- Academic literature search (see `research/litreview/`)
- Legal due diligence and legal risk opinions (see `c-level-advisor/general-counsel-advisor/`)
- Competitor teardowns for positioning and sales use (see `business-growth/competitive-teardown/`)

**Important Caveats:**
- The tools do not search the web or any database. Claims, sources, reliability codes and independence links are all entered by the user.
- Source independence is only as accurate as the `independent_from` links supplied. Two articles that repeat one press release count as independent unless the user marks them.
- Classification of statements uses the declared category, the evidence attached and wording patterns. It is a check, not a verification of the facts.
- A dossier is a snapshot. Facts about companies, people and markets go stale and need a set review date.

## Files

| File | Purpose |
|------|---------|
| `scripts/dossier_outline_generator.py` | Outline tailored to subject type and purpose |
| `scripts/source_triangulation_validator.py` | Checks claims for independent, reliable sources |
| `scripts/fact_inference_separator.py` | Classifies statements as fact, inference or speculation |
| `scripts/research_note_checker.py` | Checks a repo research note: claims cited, citations resolve, primary sources dominate, required sections present |
| `assets/research_note_template.md` | Template for a repo research note |
| `assets/sample_research_note.md` | Complete sample note that passes the checker |

## Tools Overview

`research_note_checker.py` exits 0 when the note passes, 1 on any error (uncited claims over
`--max-uncited`, a citation with no source, mostly secondary sources, a missing Question,
Answer or Sources section), and 2 when the file cannot be read. It never fetches links.

## Reference Documentation

- `references/repo-research-notes.md` — when to write a research note, running it in the background, primary-source rules, note structure, where to save it, keeping it current
- `references/dossier-frameworks-and-structure.md` — outline patterns per subject type
- `references/source-triangulation-and-reliability.md` — source assessment, triangulation
- `references/fact-vs-inference-discipline.md` — categorization + writing patterns

## Related skills

- `research/litreview` — academic literature search
- `c-level-advisor/ceo-advisor` — strategic briefing patterns
- `c-level-advisor/general-counsel-advisor` — legal due diligence overlap
- `business-growth/competitive-teardown` — competitive intel angle
- `business-growth/customer-success-manager` — account research patterns

---

## What I Need You to Do

First, detect which platform I'm using (Claude.ai, ChatGPT, etc.) and follow the matching instructions below.

### If I'm on Claude.ai:

Walk me through these exact steps:

1. **Create the Project:** Tell me to go to **claude.ai > Projects > Create project** and name it **"Dossier"**

2. **Add Project Knowledge:** Give me the COMPLETE skill definition above as a single copyable text block inside a code fence. Tell me to click **"Add content" > "Add text content"** inside the project, then paste that entire block. Do NOT say "paste from above" -- give me the actual text to copy right there.

3. **Set Custom Instructions:** Tell me to open project settings and paste this exact instruction:
   "You are an expert Dossier in the Research domain. Use the project knowledge as your expertise. Follow the workflows, frameworks, and templates defined there. Always provide specific, actionable output."

4. **Test It:** Give me a specific sample prompt I can use inside the new project to verify it works. Pick a real task from the skill's workflows.

### If I'm on ChatGPT:

Walk me through these exact steps:

1. **Create a Custom GPT:** Tell me to go to **chatgpt.com > Explore GPTs > Create**
2. **Configure it:**
   - Name: **"Dossier"**
   - Description: "Structured intelligence dossiers on companies, people, markets, or domains, with source triangulation and fact/inference discipline. Use when preparing a deal-prep dossier, executive briefing, or due-diligence overview."
   - Instructions: Give me the COMPLETE skill definition above as a single copyable text block inside a code fence to paste into the Instructions field. Do NOT say "paste from above."
3. **Test It:** Give me a sample prompt to verify it works.

### If I'm on another platform:
Ask which tool I'm using and adapt the instructions accordingly.

## Important
- Always provide the full skill text in a ready-to-copy code block -- never tell me to "scroll up" or "copy from above"
- Keep the setup steps simple and numbered
- After setup, test it with me using a real workflow from the skill

Source: https://github.com/borghei/Claude-Skills/tree/main/research/dossier/SKILL.md