Loop Patterns
The Loop Pack
Self-improving loops are the engine behind every serious AI workflow. These 10 prompts give you the core loop patterns — self-refine, red-team, ReAct, reflexion, variant breeding, auto-eval, plan-execute-verify, compounding improvement, and safe harness — ready to drop into Claude and run.
10 promptsFree to use
ClaudeLoopsAgentsSelf-Improving
01The Self-Refine Loop
/Loop Do [task]. Then critique your own output: list its 3 biggest weaknesses against [criteria]. Rewrite it fixing all 3. Repeat this generate -> critique -> rewrite cycle 3 times. Show only the final version plus one line on what improved.
02The Optimization Loop
/Loop Goal: maximize [metric]. Baseline: [current]. Propose ONE specific change and predict its effect. I'll give you the result - keep it if better, revert if not - then propose the next. Track 'best so far' each round. Stop after 3 rounds with no gain.
03The ReAct Agent Loop
/Loop Loop: Thought -> one Action -> Observation, using each result to decide the next step, until the goal is met. Goal: [X]. Tools available: [list]. Print Thought / Action / Observation every round.
04The Reflexion Loop
/Loop Attempt [task]. If it falls short, write a short LESSON on exactly what went wrong and what to change. Retry applying every lesson so far. Keep a running Lessons list. Repeat until it succeeds (max 5 tries).
05The Red-Team Loop
/Loop Produce [solution]. Now attack it as a skeptical expert - find how it breaks or where it's wrong. Patch every flaw. Attack again. Repeat until you can't break it. Show the hardened version and the attacks it survived.
06The Variant Loop
/Loop Generate 10 variations of [thing]. Score each 1-10 on [metric] with a reason. Keep the top 2. Generate 10 NEW variations from those. Repeat for 3 rounds. Show the single best result and why it won.
07The Auto-Eval Loop
/Loop Produce [output] AND a strict 5-point rubric to grade it. Grade your own output against the rubric. If it scores below 4/5, fix the weakest point and re-grade. Loop until it passes - then show the output and its scorecard.
08The Plan-Execute-Verify Loop
/Loop Break [goal] into clear steps. Execute step 1, then VERIFY it actually worked before moving on. If a step fails, diagnose, fix, and re-verify. Continue to the end. Never skip verification.
09The Compounding Loop
/Loop Take [current version]. Make the single highest-leverage improvement - show the change and the new version. Every time I say 'again', improve it once more without repeating a past change. Keep a log of every improvement.
10The Loop Harness
/Loop Run the loop autonomously, but add safe exits: stop when the goal-check passes, OR after [N] iterations, OR if two rounds in a row don't change the result. Before stopping, summarize what improved and what's still weak.
More packs on the way
I drop new prompt packs on Facebook and Instagram regularly. Follow to get them as soon as they land.