How It Was Built, Day by Day
Phase 1 - The 18-day origin (Mar 31 to Apr 17, 2026)
Headline: a non-technical founder went from create-next-app to live at radiusos.ai with
paying users in 18 calendar days, directing Claude Code. ~153 commits on main, ~41,000 lines
of application code, a 30-model Prisma schema. Source: build-log.md.
The teachable beats (not every day, the ones that teach something):
- Day 1 (Mar 31) - the v1 in one session. A single commit: "Build OutreachOS v1", 72 files, 10,190 insertions. The entire first version (schema, CRUD, Kanban, multi-tenant routing, auth, UI) generated in one sitting. The name was OutreachOS. Lesson: AI collapses the cost of a first version to near zero. The hard part starts after.
- Days 2-7 - the infrastructure fights. Prisma 7 + Neon Postgres + Vercel. Roughly 10 commits
across four days just to get the database connection right, because Claude's training data was
stale on Prisma 7. Lesson #1 of the course's "what AI can't do" module: AI writes correct
application logic but cannot see your Vercel dashboard or debug deployment/env issues. The human
bridges that gap. (Every
debug:commit in the whole project lands in Days 7-10 and is about env vars, DB URLs, or OAuth, never application logic.) - Day 7 - the architecture rewrite. Workspace-scoped routing landed: 79 files, +5,690 / -11,460. More deleted than the entire v1 added. Lesson: ship small was not yet the practice; this is the "before" picture that the later discipline corrects.
- Day 11 (Apr 11) - "it clicked." Nine feature batches in one day, ~8,200 lines, minimal fixes: tags, automation engine, Stripe billing, email tracking, marketplace. The inflection point. Once the codebase had enough patterns (server-action shape, plan gating, auth flow), Claude could pattern-match against existing code instead of inventing. Lesson: the codebase itself becomes context; consistency compounds.
- Day 13 (Apr 13) - the rebrand. OutreachOS to RadiusOS, plus the template marketplace and email sequences. Lesson: naming/positioning is a founder call, executed by AI across the whole repo in a day.
- Day 14 (Apr 14) - the fix ratio flips. Highest commit-count day, ~80% fixes. "The product was being used by real people." Mobile audit, Stripe hardening, cron limits. Lesson: real usage is the only thing that surfaces the cross-cutting bugs (e.g. "loosen Clerk auth checks across 27 server actions" on Day 15 - too-strict guards blocking real users).
- Day 16 (Apr 16) - AI goes live. Smart Deal Scoring (Claude Haiku 4.5): context assembly, prompt caching, budget gating, rule-based fallback, nightly batch cron. 10 of the day's commits were prompt/persona refinements - the AI features needed more iteration on the prompt than on the code. Lesson: prompt engineering is its own work stream.
The patterns that become course principles (all from build-log.md):
- The fix ratio (Days 1-7 were 1:3 features:fixes because infra was fighting; Day 11 was
near-pure forward progress once patterns existed) - teaches "consistency compounds."
- Debug commits = deployment, not code - teaches the human/AI division of labor.
- What shipped clean vs. what took many attempts - Stripe/automation/scoring shipped in one
commit; Prisma-Neon, Gmail OAuth, Vercel cron, and auth strictness took many. Maps directly to
"where AI is strong vs. where you have to drive."
Phase 2 - The intelligence + trades era (Apr 24 to June 2026)
Spine to enrich by mining the 40 pass-offs (next job). At the arc level, this is where the product went from "a solid CRM built fast" to a differentiated, repositioned product:
- MCP server - one protocol layer, two front doors (public server + in-product "Ask RadiusOS" chat via in-process bridge). Distribution shipped with the feature.
- AI intelligence layer - real scoring, drafts, enrichment, semantic search, the Morning Digest, voice-matched drafts, inbox-driven auto-update.
- The "operating system for the trades" pivot - from "CRM for solopreneurs" to a vertical, service-business product. Walk & Talk (talk through a job site, AI writes the record) and phone-signable quotes became the marquee features.
- Quote approval + signature + job costing - the walk-to-paid lifecycle.
- The honesty discipline hardened into build gates - Principles #7/#11 plus prebuild scripts
(
check-marketing-coverage,check-help-articles,check-no-em-dashes) that fail the build. - The error/feedback system - friendly error boundaries, typed error codes, toasts, feedback widget.
(Each of these has a pass-off or several; the mining job turns them into dated entries with the decision and the scar attached.)
Phase 3 - The proof (admin dashboard, June 2026)
Real numbers, exactly as shown, no rounding (Principle #7). All PII excluded.
Business: - 26 real users / 24 real orgs · 6 paid orgs - MRR $204 · ~$427 lifetime revenue · 25% free-to-paid conversion · 0 cancellations all-time - Plan mix: Free 18 · Pro 3 · Business 2 · Team 1 - Live and revenue-generating roughly 2-3 months after the first signups (Apr 2026)
The cost story (the standout teachable asset): - Total AI cost to run the entire product, all-time: $27.36 (2,731 calls). ~$12.58/month now. - ~43% of input tokens served from cache (1.74M of 4.09M) - prompt caching, proven in production. - Free-tier spend pool: $0.05 of $100 used. The cap + killswitch were prudent and never stressed. - Deal scoring is ~half the AI spend; the marquee Walk & Talk features cost pennies.
The honest scars (better teaching than the wins): - Walk & Talk is the marquee feature and has 0% 30-day adoption (0 of 25 workspaces). What you market is not what users do. - Conversion and churn look great but on a small sample (6 paid of 24). Teach with N attached.
The defensible headline for the whole course: a non-technical founder built a real, paying-customer SaaS solo with AI, in about three weeks to first revenue, for $27 in AI costs. The believability is the asset. Do not inflate it.
This is one chapter of the operating playbook.