THE PROCESS IS THE CONTENT
Build stories
Every build here ships with its raw log — how it actually went, from the timestamped records: the walls I hit, the approaches that died, and what each build handed to the next. No clean success stories; if something failed three times, all three are here.
- 1
Giving the AI hands
April – May 2026Outlook write access via the Lokka MCP server
Three weeks of Windows, file encoding, and licensing walls — none of them about AI — then one test-driven session that fixed six things and sent the email.
← in A talk on TDD at a Dallas AI meetup
→ out Red→green test protocol
→ out The runbook habit
Read the story →~3 weeks stuck · 1 breakthrough session - 2
The fifteen-minute agent
May 2026Monday Morning Briefing Agent
Calendar + KPIs + a draft email on one command. Built test-first in a quarter hour — because three weeks of groundwork had already eaten the pain.
← in TDD protocol — from Lokka
← in Write access — from Lokka
→ out Runbook-as-agent pattern
→ out Paste-back gates
Read the story →8 tests · ~15 minutes red to green - 3
Built twice before breakfast
May 2026Production Attendance Dashboard, V1 + V2
A wall-monitor dashboard for who hasn’t clocked in — synthetic data, leading-zero landmines, an Excel file lock, and one style renamed off the wall.
← in TDD protocol — from the Briefing Agent
→ out Portable build-kit pattern
→ out The dark ops-center look
Read the story →V1 in 23 min · V2 in 15 · 4 visual styles - 4
Three dead architectures in one day
May – July 2026Pixel Agents
Making invisible agents visible — after the plugin route died twice, the log-tailing route died once, and the demo ran out of API credits mid-demo.
← in Dark dashboard look — from Attendance
← in A local model — from an earlier experiment
→ out The pixel office itself
→ out Zero-token demo pattern
- 5
Project zero
August 2026This website
The AI can’t push to GitHub — by design — so I’m the deploy step. First deploy failed, the sunrise was invisible, and my son overruled the spec. Twice.
← in Pixel office + scenarios — from Pixel Agents
← in The memory-file habit — from the runbook lineage
→ out These build stories
UP NEXT ON THE LADDER
Three Models, One Judge
buildingThe same invoice-audit task, given identically to Claude, ChatGPT, and Grok — then graded by a human, with receipts. An experiment in public.
See where it stands →RAG knowledge base
plannedIndex my own notes, ask them questions conversationally. Learning retrieval end to end: chunking, embeddings, vector store.
Small agent chain
plannedThree or four AI agents doing one real task well — small enough to actually debug the coordination problems.
Written from the primary record — timestamped test logs, commit history, and session briefs. Where the record is thin, the story says so instead of guessing. You can also watch these builds orbit in the Lab.