promptimize

Designs and diagnoses the control flow of agentic loops - pattern selection, supervisor-enforced caps and terminal states, no-progress detection, external verifier gates, context and compaction budgets, crash recovery. Use when an agent loop stalls, drifts, repeats itself, or never stops, or when building agent orchestration. Not for goal or prompt wording, ordinary code loops, or scheduled jobs.

Builds standalone prompt evaluations - test sets from real failures, paired A/B comparisons with sample-size math, deterministic checks before LLM judges, judge calibration and bias controls. Use when asked which prompt is better, to A/B test or benchmark prompts, or to set up LLM-as-judge scoring. Not for rewriting prompt text, agent goals, or loop runtime design.

Rewrites and tunes existing prompt text - removes contradictions, applies reasoning-vs-standard regime rules, sizes few-shot examples, fixes structured output, follows current Fable 5 / GPT-5.6 guidance, then checks the rewrite. Use when asked to optimize, improve, rewrite, shorten, tune, fix, or migrate a prompt or system prompt. Not for agent-run goal text, prompt benchmarks, or loop design.

Drafts or fixes the goal text for an autonomous agent run (/goal, unattended sessions) - checkable success conditions, evidence requirements, scope and permission limits, retry and budget exits. Use when writing an agent goal, run definition of done, or run success criteria, or when a goal gets gamed or marked done early. Not for generic project goals, prompt rewrites, or loop runtime design.