Episode Details

In episode 3 of [REDACTED], we (David and Taylor) run a live demo of an agentic CRM-cleanup pipeline (and finds out, on air, what it costs). We also walk through a landing page workflow that compresses what used to take five people and a month of meetings into something one person can do in an afternoon.

The bulk of episode 3 is a live demo, just what we wanted. Plus, the best part of Redacted is that nothing is finished. Every episode is a midstream demo of something that might break, might cost more than it should, or might be the thing that quietly changes how a whole category of work gets done. Enjoy the conversation.

What We Cover

• $1.98 to clean a brand: Taylor’s CRM pipeline ran live on air with one restaurant with two locations, 64 turns, 95% cache hit, total cost under two dollars. 
• The code-vs-agent slider: Taylor built an interactive slider with pros and cons on each side. Fully deterministic code can’t handle ambiguity, fully agentic can’t be tested or priced. 
• The “no brand graph” problem: There is no canonical source of truth for the restaurant industry. Google Places, SERP, and Yelp all return ranked top-20s — you can’t reconstruct the full graph locally. This is what makes the cleanup problem agentic by necessity.
• n8n’s flowers, n8n’s thorns: Taylor’s running joke is that every local agent he builds eventually looks like n8n. But debugging n8n at scale meant pulling down a million-token JSON dump every morning just to grep it. Local won on debuggability, not capability.
• Confidence tiers for copy extraction: “Act on it” = recurring in 6+ meetings. “Pattern” = 4–5. “Emerging” = 1–3. The model can suggest copy at any tier; the operator decides what gets shipped.
• The headline no human wrote: “Get found by foodies who are paying to find you.” This sidesteps the #1 objection in the space (is this a free deals site?) in one line. David says no one on his team has produced anything like it in ten years of writing copy.
• Claude Design as a brand harness: David came in skeptical that it was just a Claude Code wrapper. He left convinced: same model, but the design harness around it makes the outputs materially better.
• 15–20 hours, one person: Total wall-clock on the landing page from raw transcripts to live wireframes. Historically the same work needed five people (founder, sales, copy, brand, design) and weeks of calendar time.
• Skill folder structure: David’s pattern for non-trivial skills: a tiny top-level skill file that orchestrates, plus subfolders for context (inputs the skill ingests), data (outputs by run), and prompts (exposed so they can be QA’d independently).

Timestamps:
 00:00 — Intro + choosing the Redacted logo live on air
 03:22 — Taylor demos a “sentient HubSpot” CRM cleanup agent
 04:37 — The messy CRM problem: duplicate restaurant brands & locations
 05:23 — Running the AI pipeline live
 06:07 — Why they moved away from n8n workflows
 07:15 — Live cost tracking for AI agents
 09:31 — “Architect agents” creating before/after CRM graphs
 10:29 — The agent fixes HubSpot records autonomously
 12:27 — The vision: a fully AI-maintained CRM
 13:07 — Why every company’s CRM eventually becomes chaos
 16:16 — Why HubSpot workflows can’t fully solve this problem
 17:20 — The missing “brand graph” problem in restaurants
 18:23 — Live demo success: AI cleaned the CRM in real time
 19:52 — Human time vs AI time: CRM cleanup economics
 20:27 — Code-driven vs agent-driven systems
 21:56 — The tradeoffs of n8n vs local AI infrastructure
 24:05 — When n8n still makes sense
 26:57 — David’s AI-powered investor update workflow
 29:17 — AI-generated shareholder updates with minimal edits
 29:53 — Building AI-generate