Loop it

Release Check License: MIT

Loop it is a categorized Loop goals library and Agent Skill for Codex, Claude Code, and Cursor. It turns an open-ended request into one complete, command-free prompt that works toward a review-ready deliverable and returns rubric evidence or a clear blocker, with an iteration cap, stop conditions, and approval gates.

Users start with a normal-language goal. They do not need to copy or run terminal commands.

Open the Loop It product page

Start with a goal

Choose a loop goal, add the source material it needs, and open the generated prompt in your agent.

Turn these numbered customer comments into traceable themes and recommend one evidence-backed next action. Keep observations, interpretation, and assumptions separate. Return the brief and score it against the supplied proof rubric.

Create three landing-page message directions for this audience using only the approved product evidence. Select the strongest direction, mark unsupported claims as assumptions, and return one review-ready concept.

Inspect this repository, repair the highest-confidence issue with the smallest scoped change, run the project’s safe local verification inside the agent workflow, and return proof.

Loop goals library

The starter library contains 12 experimental loop goals across six categories.

Category Loop goals Review-ready deliverable
Product & UX First Value Sprint; Mobile Journey Rescue Focused journey improvement with before-and-after evidence
Design & Prototyping Visual Consistency Pass; Clickable Flow Prototype Reviewable interface change or local prototype
Research & Decisions Feedback Theme Synthesis; Competitor Evidence Brief Cited research brief with facts and inference separated
Content & Messaging Landing Page Message Pass; Source-to-Content Pack Review-ready message or draft content set
Data & Evaluation Data Quality Snapshot; Experiment Results Brief Reconciled quality or decision report
Operations & Support SOP From Messy Notes; Support Knowledge Gap Audit Usable procedure or prioritized coverage brief

Every loop goal declares:

  • Required inputs
  • Expected deliverable
  • Proof rubric and evidence
  • Iteration cap
  • Required capabilities
  • Reliability status
  • Approval gates

Each loop goal compiles one complete natural-language prompt that works toward a review-ready local deliverable and returns criterion-by-criterion rubric evidence or a clear blocker. Loop goals never publish, send messages, contact people, deploy, purchase, access private systems, or mutate production data without explicit approval.

How the loop works

Creative and analytical loop goals use one bounded workflow:

UNDERSTAND → CREATE → CRITIQUE → REFINE → PROVE

The agent stops when the rubric is satisfied, the iteration cap is reached, the same weakness repeats twice, required context is missing, or approval is needed. A completed rubric means the artifact is ready for review; it does not mean a human automatically approved it.

Advanced repository work uses:

DISCOVER → PLAN → EXECUTE → VERIFY → ITERATE

The agent runs safe project checks internally and reports the evidence. The user receives the prompt and proof, not a command-generation walkthrough.

Advanced loop library

The original 20-loop library remains available for engineering and operational work:

  • Turn-based explanation, review, debugging, copy, and small-edit passes
  • Goal-based CI repair, ticket repair, security hardening, and release readiness
  • Time-based PR, CI, dependency, documentation, and production-smoke watches
  • Proactive bug, dependency, review-comment, feedback, and code-health routines

Time-based and proactive loops still require an approved scheduler or connector. Loop It does not claim to run in the background unless that heartbeat is actually configured.

Prompt-only contract

Generated user prompts must:

  • Be normal-language messages that work without a slash command
  • Name the goal, artifact or verifier, proof, cap, stop conditions, and approval gates
  • Tell the agent to handle safe local verification itself
  • Avoid terminal commands and shell snippets
  • Keep external writes and irreversible actions behind approval
  • Return evidence, assumptions, blockers, remaining risks, and the next safe action

What is inside

  • skills/loop-it/SKILL.md: canonical Agent Skill instructions
  • skills/loop-it/references/library/goals.json: categorized Loop goals library
  • skills/loop-it/references/library/goals-schema.json: goal-library contract
  • skills/loop-it/references/library/goals-evals.json: goal-routing evaluation set
  • skills/loop-it/references/library/loops.json: advanced execution patterns
  • skills/loop-it/scripts/goal-library.mjs: goal validation, routing, and prompt compiler
  • skills/loop-it/scripts/select-loop.mjs: advanced loop selector
  • skills/loop-it/scripts/start-loop.mjs: prompt-only host launch compiler
  • skills/loop-it/scripts/run-loop.mjs: guarded repository runner and loop-goal prompt router
  • tests/: install, routing, prompt, readiness, and run-proof checks

Install

Install Loop It from the Codex plugin interface, or use the product page to open a generated goal in Codex. Project-local installs are also supported for Codex, Claude Code, and Cursor; see installation notes when maintaining a local development copy.

Reliability

The 12 new loop goals begin as experimental. Promote a goal to tested only after repeated real runs produce the declared deliverable, satisfy the rubric, preserve safety boundaries, and avoid confident misrouting.

The advanced loop recipes are verifier-gated patterns, not guaranteed outcomes. Loop It reports whether proof passed, failed, was blocked, or still needs review.

Contributing

See CONTRIBUTING.md for development workflow and RELEASE.md for maintainer-only release checks.

License

MIT