a builder's codex
codex · operators · Arvid Kahl · ins_10-80-10-ai-workflow

10-80-10: human direction, AI execution, human polish

By Arvid Kahl · Bootstrapped founder; builder of Podscan · 2026-03-03 · essay · Indie Hacking the Singularity — 10-80-10 AI workflow

Tier B · TL;DR
10-80-10: human direction, AI execution, human polish

Claim

For AI-augmented bootstrapped builds, divide work into three phases: the founder owns the first 10% (strategic direction, problem definition, architectural decisions), AI handles the middle 80% (code generation, data migration, testing, documentation), the founder owns the final 10% (quality review, edge cases, customer-facing polish). The split is what lets a solo operator produce team-scale output without losing the human contribution.

Mechanism

The 80% middle is well-understood, repetitive work where AI is fast and cheap. The bookend 10%s are where taste, context, and customer judgment compound. Kahl applies this at Podscan, where agentic coding tools handled infrastructure migration while he held direction and final review. The split also clarifies what to outsource mentally — anything that smells like the middle 80% goes to AI by default.

Conditions

Holds when:

Fails when:

Evidence

"You provide the first 10% (strategic direction, problem definition, architectural decisions), AI handles the middle 80% (code generation, data migration, testing, documentation), and you apply the final 10% (quality review, edge case handling, customer-facing polish)."

— Arvid Kahl (synthesized from operator's published work)

Signals

Counter-evidence

Cat Wu's 100% automation rule cuts against the bookend model: if the human polish step is needed every run, the workflow isn't really an automation, just a co-pilot. For repeatable production tasks, the right move is to push the final 10% itself into automation.

Cross-references

Open the interactive view → View original source → Markdown source →