Case study
Hierocode
Hierarchical coding tool that splits planning (frontier models) from drafting (local models) to keep incremental cost near zero.
Highlights
- Frontier planner + local drafter split
- $0 incremental cost on subscription paths
- QA loop: accept/revise/split/escalate with capped revisions
- Diffs printed, never auto-applied
Hierocode splits coding work by role to keep cost near zero on a bring-your-own-subscription setup. A frontier-tier planner, a Claude Pro or ChatGPT Plus subscription, or an Anthropic API key, breaks a task into a structured plan and reviews the result. A local Ollama model does the actual drafting for each bounded unit of work, entirely on your own hardware.
The economics are the point: planner calls stay short (roughly 5-20k input tokens, 1-3k output), so two or three of them barely touch a subscription's soft quota, while the drafter runs at $0 per call. A typical task costs four planner round-trips: one planning call plus three QA rounds when the plan cache misses, and zero when it hits.
The pipeline itself is a QA loop between drafter and planner: the reviewer can accept a diff, ask for a revision (capped), split the unit into smaller pieces, or escalate it to the planner directly (also capped). Every diff is printed to stdout for manual review and application; nothing is applied automatically. The wizard detects your environment (RAM, installed CLIs) and writes a working config in one step.