ksimback/looper

Design visual, review-gated agent loops for Claude Code before you run them.

What it solves

Looper addresses the problem of "garbage in, garbage out" when using AI agent loops. It prevents users from running vague or unfalsifiable goals that lead to inefficient or incorrect agent behavior by providing a design layer that forces discipline, explicit verification criteria, and independent review before any execution begins.

How it works

Looper acts as a design coach and specification generator. It interviews the user to define a sharp goal, identifies context sources and actions, and establishes a typed verification system (programmatic checks, model-based judges, or human sign-off). It encourages the use of a "council"—a second model from a different vendor to act as a reviewer or judge to avoid single-vendor blind spots.

Once the design is finalized, Looper emits a portable specification (loop.yaml and loop.resolved.json) and a set of artifacts. These can be executed immediately within a Claude Code session via a generated prompt, or run externally using a provided Python runner.

Who it’s for

It is primarily designed for users of Claude Code who want to build robust, repeatable, and auditable agentic workflows rather than relying on ad-hoc prompts or basic goal-setting commands.

Highlights

  • Design-First Approach: Forces the definition of a "definition of done" and termination guards (iteration caps, budget limits, no-progress signals) before running.
  • Cross-Model Review: Supports wiring in different model families to act as independent judges for higher quality assurance.
  • Portable Specs: Generates versionable loop.yaml files that can be linted in CI to prevent anti-patterns.
  • Template Library: Includes pre-designed patterns for common tasks like security scans, code reviews, bug hunting, and documentation syncing.
  • Privacy Controls: Implements two-layer redaction to ensure sensitive files are not sent to external model providers.