Strategy notes · Expert network · 2026
Expert-network agency, autonomous growth system
Decks turned into evidence, the website audited against them, then an agent pipeline rebuilt with tests and a cutover plan.
Context
An expert-network agency building an agentic operating system. The work had three parts: turn the pitch decks into structured material, audit the live website against it, then run the brand's traffic and content as an agent pipeline and rebuild that pipeline as a tested web app.
Starting point
Two client decks, one of them 13 slides. A live website whose copy had been written separately from the decks. A content pipeline already running as n8n workflows, described as fully autonomous, with no human gates.
Decision
Extract first, then audit, then build. Each step had to produce something a person could check.
- Deck to evidence. The company deck became 14 verbatim rows, each tagged as a definition, a claim, a positioning line or a commitment.
- Copy audit. Every website string was checked against the deck. The fixes came as four ready-to-run prompts, one per group of files.
- Pipeline. An AI supervisor confirms topics, but a deterministic dispatcher, not the model, calls the publisher. A fast rule check with no model runs before a model judge. The thresholds are written down: a signal is confirmed at 0.6, content is approved at 0.8, and a draft gets at most 2 revisions.
- Rebuild. A Next.js dashboard, a Postgres store and a worker behind a feature flag. The migration runs in four phases, with a dry run that compares outputs for 3 to 5 cycles before switching, and n8n kept as a manual fallback.
Proof
The audit: 38 strings, 25 fine, 9 drifted, 4 stale. The rebuild has a 17-table schema and 9 test files. I have no record that the cutover ran in production, and no figure for traffic.
Where AI was used
The extraction tool reads decks with Claude, through the API, the Agent SDK or native vision. In the pipeline, Claude scores signals for relevance and a Claude Haiku guardrail judges content. The coding assistant had a written rule: look up each n8n node schema before changing anything, and check real execution output, because "success means nothing".
Time saved: not measured.
Deliverables
- Deck extraction pipeline with a web UI and CLI
- Verbatim extract tagged at source, keyword intel table
- Website content audit with fix prompts
- Autonomous traffic pipeline on n8n, documented
- Dashboard and worker rebuild, architecture and migration runbook