goal-loop-designer

Goal Loop Designer

Use this skill before starting a long agent loop. It turns a raw /goal or task prompt into a portable loop harness that another agent or model can execute and evaluate.

This skill designs the loop. It does not run autonomous tools, log in to providers, change repos, or approve edits.

Intake

Collect or infer the smallest useful context:

• raw goal or task prompt;
• target host: Codex, Claude Code, MiMo Code, Ollama/OpenAI-compatible, or generic CLI agent;
• repo or artifact scope;
• allowed actions and forbidden actions;
• available verification commands or manual checks;
• judge options: deterministic check, human review, Codex, MiMo Auto, local Ollama, or external LLM;
• hard limits: iterations, time, token/tool budget, file scope, network, credentials, destructive actions.

If key limits are missing, choose conservative defaults and mark them as assumptions.

Goal Critique

Before drafting the harness, critique the raw goal:

• ambiguous completion condition;
• missing non-goals;
• missing verification;
• unclear permission boundary;
• no budget or retry cap;
• self-judging without independent evidence;
• broad file or repo scope;
• hidden external dependencies;
• failure mode that would cause repeated retries.

Rewrite the goal so it is specific, bounded, and testable.

Loop Fit

Classify the work:

• single pass: one execution plus verification is enough;
• supervised workflow: multiple ordered steps, but human approval should gate progress;
• bounded loop: iteration is useful and an evaluator can decide whether to continue;
• do not loop: the work is too risky, under-specified, or unverifiable.

Prefer the lowest level that can succeed.

If a loop produces a reusable lesson, do not automatically add it to memory or a skill. Use agent-learning-layer-triage after the run to decide whether the lesson belongs in a context note, durable doc, checklist, SKILL.md, script/tool, eval, golden fixture, or rejected buffer.