HOW-TO · SOFTWARE FACTORY
Most agent collisions are two features on one branch. A factory fixes order of operations — isolate, build to house style, prove with evidence, ship through an outside reviewer — as markdown workflow and skills, not a special model.
Model-agnostic · harness-agnostic · runs on agents.md
“A software factory is literally just a bunch of markdown files.”
01
Custom order station
Skill new feature
Fresh git worktree from
origin/main; never build on main. Worktree = a copy
of the app; merge back later; delete the worktree after merge.
Parallel agents in separate tabs stay out of each other’s way.
Why: ~95% of collisions = two agents, two features, one branch.
02
Assembly line
Skill code structure
service-layer; readable cold
Models “get it done” but can be sloppy. Pattern: one model writes, another reviews (dupes, dead code) — Mickey: Sol write → Fable review.
“Just because it works doesn’t mean it’s written well.”
03
Quality control
Skills evidence driven testing +
before and after
Record before (bug) then after (fix); screenshots if you can’t record. Invisible changes → measure (e.g. 850ms → 60ms). Agent checks the after-shot against criteria; if half-done, it walks itself back to Step 02.
“Agents can’t pinky promise.”
04
Loading dock + merge / deploy / release notes
Skill grep loop
Greptile / Code Rabbit / Macroscope
Open a PR with before/after proof in the description; present the
PR URL. Reviewer confidence /5. Loop
3/5 → build → prove → ship → … → 5/5 →
human merges. The agent runs the loop without
being re-told “grep loop”.
Factory is markdown workflow + skills + domain knowledge; model and harness barely matter.
“A software factory is literally just a bunch of markdown files.”Back to FakeNerd.ai
The Startup Ideas Podcast · Greg Isenberg · Mickey’s setup as described in the article. Not Fortibus original research.