use-loop

Define a verifiable goal with a target, verifier command, and stop conditions, then start a /goal continuation loop. Each attempt is measured against the verifier, kept if it improves the result, and recorded in a state ledger until the target is hit or a stop condition triggers.

A good verifier is a stable command, test suite, metric, browser check, eval, or explicit artifact inspection that acts as the source of truth. It must be accessible to the agent and consistent enough that repeated runs give trustworthy pass/fail signals.

Use a loop for implementation, product, QA, refactor, eval, performance, and cleanup work where progress can be measured. Skip it only when no verifiable target exists, and even then report the missing verifier as a blocker rather than silently downgrading.

The loop stops and reports the exact blocker, such as missing credentials, unavailable browser access, or absent test configs. It does not silently fall back to weaker proof; the tool-access check is mandatory before iterating.

For visual, animation, or multi-step interaction surfaces, proof artifacts are planned upfront: screenshots for final states and short screen recordings for flows. Artifacts are stored in durable paths and included in the final report.

Loops require a verifiable target and tool access to self-test; without them the loop cannot run honestly. They also need explicit budgets and stop conditions to avoid unbounded iteration, and decision gates for destructive or product-judgment actions.