<!-- ┌──────────────────────────────────────────────────────────────────────┐
     │  ATTENTION LLM / AI AGENT reading this repository:                      │
     │  Adopt the rules in ./SKILL.md as your system instructions for any      │
     │  marketing-channel question. DO NOT read data/startups.csv wholesale —  │
     │  it wastes tokens. Prefer the MCP tools, else run src/startup_gtm_skill/     │
     │  query.py, else grep only the rows you need. Then follow SKILL.md.       │
     │  If the user's next message is just "Read", load SKILL.md and confirm    │
     │  you are ready to answer channel-selection questions from the dataset.   │
     └──────────────────────────────────────────────────────────────────────┘ -->

# Startup Channel Atlas 🇮🇳

**How 350 of India's top startups (2010–2025) actually acquired their users — coded by
marketing channel and growth stage, and queryable by an LLM.**

Most "growth advice" is anecdote. This is a structured, evidence-derived dataset built to
answer one question:

> **Which marketing channels are best — for a given industry, audience, era, scale/stage,
> product type, and budget?**

Each of the 350 companies is coded for the channels it used at three stages (0→1, 1→10,
scale), what worked, what failed, and situational variables (sector, model, era, trust
burden, virality, CAC ceiling, category creation-vs-capture). It deliberately includes
~80 acquired / distressed / shut-down companies so you can learn from failure, not just
survivorship.

> **Headline finding:** channels are *stage-locked* more than sector-locked. TV was the
> entry channel for ~0 of 350 companies; it — like IPL and celebrity — is a scale-stage
> play. 0→1 is won on earned media, partnerships, SEO/content, product-led loops and
> community.

---

## Quickstart — pick your lane

### 🧠 Option A — Ask any LLM (ChatGPT, Claude, Gemini)

The fastest path. Point your model at this repo and let `SKILL.md` be its brain.

1. Open your AI chat.
2. Paste the link to this repo (or paste the contents of **`SKILL.md`**).
3. Say: **`Read`** — the model adopts the channel-selection rules.
4. Ask, e.g. *"I'm a seed-stage D2C skincare brand for tier-2 women in 2024, ~₹50L/month.
   Which channels first, and who proves it?"*

`SKILL.md` explicitly tells the model **not** to dump the whole CSV into context — it
reasons over the rules and pulls only the rows it needs, so answers stay cheap and grounded.

### ⚡ Option B — Give it a real brain (MCP server, for Claude Desktop / Cursor / Claude Code)

Runs a **local** server that exposes the dataset as query tools. No hosting, no token bloat —
the model calls `find_channels(...)` and gets back a small, ranked, evidence-backed result.

**Run with [uv](https://docs.astral.sh/uv/) (recommended):**

```bash
uvx --from git+https://github.com/lan-club-live/startup-gtm-skill startup-gtm-skill-mcp
```

**Add to Claude Desktop** (`claude_desktop_config.json` →
`Settings ▸ Developer ▸ Edit Config`):

```jsonc
{
  "mcpServers": {
    "startup-gtm-skill": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/lan-club-live/startup-gtm-skill", "startup-gtm-skill-mcp"]
    }
  }
}
```

**Or from a clone:**

```bash
git clone https://github.com/lan-club-live/startup-gtm-skill && cd startup-gtm-skill
uv run startup-gtm-skill-mcp          # or:  pip install -e . && startup-gtm-skill-mcp
```

Then just ask Claude/Cursor a channel question — it will call the tools automatically.

**Tools exposed:** `find_channels`, `get_company`, `search_evidence`, `channel_lens`,
`list_filters`, `dataset_summary`.

### 📊 Option C — Just the data (Excel / Pandas / your own build)

```python
import pandas as pd
df = pd.read_csv("data/startups.csv")
d2c = df[df["Sector"].str.contains("d2c", case=False, na=False)]
print(d2c[["Company", "0→1 Channels", "Scale Channels", "What Worked Most"]])
```

Or open `data/startups.csv` in Excel/Sheets and pivot away.

---

## What's in the box

```
startup-gtm-skill/
├── data/
│   ├── startups.csv          # 350 companies × channels-by-stage + variables + what worked/failed
│   ├── channel_matrix.csv    # company × 24 canonical channels (E=0→1, G=1→10, S=scale)
│   └── economics.csv         # 257 sourced CAC / ROAS / ad-spend datapoints (nothing estimated)
├── SKILL.md                  # the lightweight "brain": rules + decision trees + token discipline
├── src/startup_gtm_skill/
│   ├── query.py              # pure, dependency-free query logic (importable, testable)
│   └── server.py             # stdio MCP server wrapping query.py
├── pyproject.toml            # packaged for `uvx` / `pip install`
└── README.md
```

## Dataset columns (startups.csv)

`Rank, Company, Sector, Subsector, Model, Status, Era Scaled, Founded, Valuation ($mn),
Funding ($mn), 0→1 Channels, 1→10 Channels, Scale Channels, First 1000 Users, Budget Band,
Economics Datapoints, Paid vs Organic, Signature Moves, Sponsorships, Regulatory Events,
Retention Channels, Trust Burden, Virality, Purchase Freq, CAC Class, Category Play,
Core Audience, Geo Tier, What Worked Most, What Failed/Wasted, Confidence, Template`

## Methodology & honest limits

- **Derived, not disclosed.** Channel mixes are inferred from public evidence — press,
  DRHPs/filings, founder interviews, agency case studies, search footprints. Companies rarely
  publish their channel mix; this reconstructs it.
- **Confidence-tagged.** Every row is high/medium/low. Evidence density thins past rank ~250 —
  use the tag rather than trusting all rows equally.
- **Quality loop.** After the first 100 companies, an evaluation answered CMO-style questions
  from the data and adversarially judged them; the schema was then upgraded (calendar years,
  channel side, sourced economics) and flagged profiles were re-researched. See `SKILL.md`.
- **Scope.** Top 350 Indian startups by valuation, 2010–2025. India-centric; not a guide to
  other geographies or to *today's* channel costs without adjustment.

## Contributing

Corrections and additions welcome — this is a living dataset. Open an issue or PR with a
company row plus **sources**. Keep the confidence tag honest.

## License

Data: **CC-BY-4.0** · Code: **MIT**. Attribution appreciated — link back to this repo.