# UltraPlan

Verified Chunks for Agent Plan Mode

## What It Is

UltraPlan is a small addition to native Plan Mode for long or difficult agent work. Native Plan Mode lists what the agent intends to do. UltraPlan defines what counts as done, splits the work into verified chunks, and refuses to advance without proof from the real artifact. It is the planning sibling to UltraGoal: UltraGoal turns a lazy goal into a working contract before the agent starts moving; UltraPlan turns a rough planning request into a plan that has to prove itself. Invoke it as `/ultraplan migrate the login flow to OAuth`; everything after `/ultraplan` is raw planning intent, and the skill sharpens it into an Intent Grading Contract, chunk scopes, evidence requirements, and review gates before work begins.

## More Context

Plan Mode keeps an agent on rails, but long tasks still create a familiar kind of fake work: a demo called complete, a bug papered over, dashboard analytics that turn out to be placeholders. UltraPlan is built for that specific failure mode.

Everything after `/ultraplan` starts as raw intent, not the final plan. UltraPlan turns it into an Intent Grading Contract, a pass/fail definition of done. Work that looks useful but misses what was actually asked for fails.

The plan breaks into verified chunks. Each chunk has its own scope, steps, and required evidence, then runs an Evidence Loop: the agent inspects the real thing it made, runs the command, hits the endpoint, screenshots the page, and revises from what it finds.

A separate sub-agent reviews each chunk adversarially, trying to reject it by hunting for fake progress, paper-over fixes, and placeholder behavior. If the host cannot provide real sub-agent review, UltraPlan can draft the plan but cannot complete the execution chunks. It fails closed instead of pretending self-review is the same thing.

## Media

- Primary image: https://www.justinwetch.com/ultraplan/banner.png

## Facts

- **Year:** 2026
- **Published:** 2026-06-27
- **Last updated:** 2026-07-01
- **Category:** AI, Open Source, Developer Tools
- **Tags:** AI, Agents, Skills, Codex, Claude, OpenAI, Open Source, Dev Tools, Automation, R&D
- **Canonical URL:** https://www.justinwetch.com/projects/ultraplan/

## Links

- [GitHub repo](https://github.com/justinwetch/UltraPlan) - source

## Tags

- AI
- Agents
- Skills
- Codex
- Claude
- OpenAI
- Open Source
- Dev Tools
- Automation
- R&D