Tell Claude what you want once. It works until the job is verifiably done, and it gets smarter every time.

Raw agent sessions make you the quality control. ultragoal puts that discipline in the system instead: a rubric gates progress, a separate verifier re-runs every check itself, and memory compounds across runs.

The loop

01
Consult
Reads memory and scans your repo before asking anything.
02
Interview
Asks only the decisions that steer the outcome.
03
Spec
Builds a rubric of exact, checkable commands.
04
Arm
Gates the loop to a goal file that survives restarts.
05
Work
Feeds the rubric back each turn, so it keeps going.
06
Verify
A fresh-context agent re-runs every check to refute it.

What's inside

Rubric-gated loop

Progress is blocked until every item is checked with evidence from a real command, never vibes.

Fresh-context verifier

An independent subagent re-runs the checks and tries to prove the work wrong, not confirm it.

Persistent goals

State lives in a git-committed file, so a goal survives sessions, /clear, and days away.

Shared team memory

Verified facts, working patterns, and dead ends are committed to git. Every run starts smarter.

Use it for

Optimization that can't regress

Baselines the metric, keeps only real improvements, and reverts anything that makes it worse.

› cut checkout p95 under 200ms without breaking contract tests

Migrations, end to end

Survives /clear and restarts, and won't call it done until every call site is gone and checks pass.

› migrate every class component to hooks; keep tests and types green

Bugs fixed for real

A fresh-context verifier re-runs every check and tries to refute the fix, so done isn't just claimed.

› make the failing repro test pass without breaking the suite

Backlogs while you sleep

Works turn after turn unattended within a set budget, gated so it can't fake progress.

› clear the typed-error backlog in src/, one module at a time

Get started

Install for Claude Code, Codex, or both.

Start a goal inside Claude Code. A messy ramble is fine.

Autopilot, end to end.