# The one-prompt company: running sales & marketing from a knowledge graph

> How Improvado turned its entire sales/marketing stack into one company knowledge graph so agents run ICP discovery, sales-script generation, and multi-channel campaigns from a single prompt.
> by Daniel (Improvado)
> group: company
> event: ai-native-startups
> source: cyber•OS — https://os.cyber.fund/case-studies/one-prompt-company-knowledge-graph

## Problem

At a typical company, the context an agent needs is scattered across dozens or hundreds of tools — at Improvado, ~140 vendors serve 90 people. No agent can act end-to-end when the facts about processes, customers, and policies live in 100+ disconnected systems. Answering a cross-stack question, let alone running a campaign, means a human stitching the context together by hand every time.

## Approach

Improvado built **Miras** — "eyes, hands, and memory" for agents — to collapse that sprawl into a single substrate:

- **Ingest the whole stack.** 1,000+ integrations pull the company's sales and marketing data into one place, run through data lineage, governance, and a semantic layer.
- **Model it as one knowledge graph.** The output is a high-level ontology describing every process, entity, and policy in the business — not just raw tables.
- **Optimize for humans, not the database.** After trying graph and SQL databases, they settled on Markdown files (YAML front matter + links), also stored in ClickHouse. The reasoning: humans are the ultimate bottleneck, and what's easy for a human to read is usually easy for an agent to traverse.
- **Discipline the model.** A mix of Kimball dimensional modeling and Basic Formal Ontology (separating time-bound from timeless entities), kept MECE (mutually exclusive, collectively exhaustive) and following the Minto principle — one idea per document, expanded fractally. Daniel calls modeling reality the hardest and most important part.

## Results

Because all context lives in the graph, almost any task collapses to a single prompt — Daniel calls it "the first one-prompt company." Three real (pre-recorded) Claude Code / Codex sessions demonstrated:

1. **ICP discovery → landing page.** Analyzing every sales call surfaced an unknown high-converting segment (gaming), extracted its pain points, then generated and published an on-brand landing page targeting it.
2. **Sales-technique mining.** Scoring discovery calls against famous sales methodologies exposed gaps — the team never asked the "five whys" and rarely used the Challenger sale — and auto-wrote a mid-market/enterprise discovery-call script the team has since adopted.
3. **One-prompt multi-channel campaign.** A single prompt produced a full campaign across Google Ads, Facebook creatives, outbound email, and landing pages, because the brand, audience, and product context were already in the graph.