Most practitioners run one prompt, read the output, and move on. For drafting emails or summarizing a document, that is fine. For anything that requires actual analytical depth, it is the wrong approach. A single LLM pass gives you the first reasonable answer, not the best one. Loop prompting is how you fix that.
What It Is
Loop prompting chains LLM calls so the output of one becomes the input of the next. Each pass builds on the last, forcing the model to revisit and refine rather than settle on the first plausible answer.
Four patterns cover most use cases:
- Refinement loop: Run the same prompt N times. Each pass improves on the prior output. Good for writing quality, argument clarity, or sharpening recommendations.
- Critic loop: Alternate between a generator and a critic. The generator produces, the critic tears it apart, the generator revises. Repeat until the critic runs out of objections.
- Pipeline loop: Sequential specialized prompts, each handling one stage. Extract, analyze, synthesize, format. Each stage gets the full prior output as context.
- Validation loop: Generate, then verify against a rubric or constraint set. Loop until the output passes. Best for structured outputs like JSON, compiled code, or compliance checklists.
The pattern you choose depends on the failure mode you are trying to prevent. Refinement loops fix shallow first drafts. Critic loops catch reasoning gaps and omissions. Pipeline loops handle complex problems where a single prompt collapses under its own scope.
One Thing To Watch
Context window bloat will quietly degrade output quality in longer loops. Each pass accumulates prior output as tokens, and the model’s reasoning gets diluted as the window fills.
The fix: summarize between passes. After each major stage, generate a compressed summary of what was established and carry that forward instead of the full raw output. This also keeps cost and latency manageable in automated pipelines.
The Template: Pipeline Into Critic
The highest-value structure combines both patterns. A pipeline loop handles research and problem decomposition. A critic loop stress-tests the output before you act on it. The two prompts below build that combined structure.
Prompt 1: Use when you have a topic and problem but no existing research
You are a structured research and analysis engine. Follow every step in order. Do not skip any phase. TOPIC: [e.g. "sales compensation design"] PROBLEM: [Describe the specific situation — who is involved, what the tensions are, what constraints exist] --- PHASE 1 — RESEARCH LOOP Step 1: List the 5 most relevant frameworks, studies, or bodies of knowledge that an expert would draw on to solve this problem. For each, write 3–4 sentences summarizing what it says and why it matters here. Step 2: Identify 3 real-world case studies or common industry patterns where this type of problem has been solved well or solved poorly. One paragraph each. Step 3: List the 5 most common mistakes people make when trying to solve this type of problem. One sentence each. Step 4: Summarize everything from Steps 1–3 into a single "Context Brief" — a 200-word paragraph that captures the most important knowledge for solving this problem. You will use this as your context for all analysis below. --- PHASE 2 — ANALYSIS (run 3 times using the Context Brief from Phase 1) Label each RUN 1, RUN 2, RUN 3. Vary your angle: - Run 1: people and incentive psychology - Run 2: financial math and cost risks - Run 3: culture and change management For each run, complete all 4 pipeline stages before starting the critic loop. PIPELINE — 4 stages: Stage 1 — Extract the 5 most relevant principles from your Context Brief for this specific problem. Number them, 2–3 sentences each. Stage 2 — Map each principle to the problem. One sentence per principle on how it applies or where the specific context creates an exception. Stage 3 — Write specific, actionable recommendations. Concrete, not generic. Address the core tensions directly. Written for a smart non-expert. Max 300 words. Stage 4 — Top 3 implementation risks, one concrete mitigation each. Max 250 words. CRITIC LOOP — 3 rounds per run: Round 1: 3 strongest flaws in the Stage 3 recommendations. One concrete fix per flaw. Round 2: 3 NEW failure modes not raised in Round 1. One fix each. Round 3: 3 deeper structural risks. One fix each. --- PHASE 3 — FINAL SYNTHESIS After all 3 runs are complete: 1. State which run produced the strongest Stage 3 recommendation and why (one sentence). 2. Write the single best combined recommendation using the strongest elements from all 3 runs: - One short intro paragraph explaining the problem and why this approach - Numbered list of 5–7 concrete actions - "Watch-out risks" section: the 2 most important risks and their mitigations - Plain language, no jargon, max 450 words Do not stop early. Complete Phase 1 entirely before starting Phase 2. Complete all 3 runs before writing Phase 3.
Prompt 2: Use when you already have research or source material to work from
You are a structured analysis engine. Follow these instructions exactly and do not skip any step. TOPIC: [e.g. "hiring a Director of Sales for a $70M construction firm"] CONTEXT / BACKGROUND: [Paste any research, articles, frameworks, or industry knowledge here] PROBLEM: [Describe the specific situation — who is involved, what the tensions are, what constraints exist] --- INSTRUCTIONS: Run the following sequence 3 times (label each RUN 1, RUN 2, RUN 3). Vary your angle each run: - Run 1: emphasize the human/people side - Run 2: emphasize the financial/math side - Run 3: emphasize culture and change management For each run, complete all 4 pipeline stages before moving to the critic loop. PIPELINE — 4 stages per run: Stage 1 — Extract 5 key principles from the context that are most relevant to solving this problem. Number them 1–5, 2–3 sentences each. Stage 2 — Map each principle to the specific problem. One sentence per principle on how it applies or where it creates an exception. Stage 3 — Write specific, actionable recommendations to solve the problem. Be concrete, not generic. Address the core tensions directly. Write for a smart non-expert. Max 300 words. Stage 4 — Identify the top 3 risks in implementing those recommendations. One concrete mitigation per risk. Max 250 words. CRITIC LOOP — 3 rounds per run: Round 1: Find the 3 strongest flaws in the Stage 3 recommendations. One concrete fix per flaw. Max 300 words. Round 2: Find 3 NEW failure modes not raised in Round 1. One fix each. Round 3: Find 3 deeper structural risks in the overall approach. One fix each. --- After all 3 runs are complete, write a FINAL SYNTHESIS: 1. State which run had the strongest Stage 3 recommendation and why (one sentence). 2. Write the single best combined recommendation using the strongest elements from all 3 runs. Format it as: - One short intro paragraph - Numbered list of 5–7 concrete actions - A "Watch-out risks" section with the 2 most important risks and their mitigations - Plain language, no jargon, max 450 words Do not stop early. Complete every stage of every run before writing the synthesis.
Why This Works
A model generating recommendations will anchor on the most obvious answer and the most familiar mitigations. Ask it to self-critique in the same pass and it pulls its punches. Separate the generator and critic roles across calls and the dynamic changes. The critic has no investment in defending the prior output. It just finds problems.
Running three analytical angles through a structured pipeline, each followed by three rounds of dedicated critique, surfaces failure modes that single-pass prompting reliably misses. The final synthesis step then forces prioritization across all that output, which is where the real compression of value happens.
This is not a casual workflow. It is closer to how a competent analyst would actually structure a hard problem: research first, multiple frames, explicit pressure-testing, synthesis at the end. The LLM does the mechanical work at each stage. You design the structure and review what comes out.
Common Questions Asked
Loop prompting chains LLM calls so the output of one pass becomes the input of the next. Each pass builds on the last, which forces the model to revisit and refine instead of settling on its first plausible answer.
Use a single prompt for simple tasks like drafting an email or summarizing a document. Use loop prompting when the work requires real analytical depth, where the first reasonable answer is rarely the best one.
Refinement loops run the same prompt repeatedly to improve quality. Critic loops alternate a generator and a critic to catch reasoning gaps. Pipeline loops break a complex problem into sequential stages. Validation loops generate, then verify against a rubric or constraint set until the output passes.
