Loop Engineering: How to Get Better AI Outputs

Giving an AI a vague task and hoping it will work out what you meant is not a strategy. It is delegation without a brief. **Loop engineering** is an emerging term for a more disciplined approach: set a clear goal, define the required output, check the result against explicit criteria, then revise or stop. The name is new and not yet a formal technical standard. The underlying practices are well established in AI evaluation and software workflows. For an SME, the useful lesson is simple. Better AI work rarely comes from finding one clever sentence to type into a chat box. It comes from designing a repeatable way to judge whether the answer is actually fit for purpose. ## Why a better prompt is only the beginning Prompt engineering usually concentrates on the instruction given to the AI. That matters, but an instruction is only one part of a reliable process. Consider the difference between these two requests: - “Write a customer update about our delayed delivery.” - “Write a 150-word email for existing customers whose orders will arrive three days late. Acknowledge the inconvenience, give the revised date, avoid blaming the courier, and end with the customer-service contact details.” The second request gives the AI a goal, an audience, boundaries and an output format. There is less room for it to guess. That principle applies well beyond writing. A useful request for research might specify which claims need sources. A request for analysis might require a table containing evidence, risk and recommended action. A request for meeting notes might require decisions, owners and deadlines rather than a general summary. This is why [AI problem definition matters before the tool](/insights/ai-problem-definition-explain-the-problem/). If you cannot describe the business outcome, the AI cannot reliably infer it for you. ## The five parts of a practical loop A useful loop engineering workflow has five parts. It does not need specialist software. You can use the same pattern in an ordinary AI conversation. ### 1. State the goal Start with what the work is meant to achieve, not merely what the AI should produce. “Create a report” describes an artefact. “Help the management team decide whether to renew this supplier contract” describes a goal. The second version gives the model a reason to prioritise cost, performance, risk and alternatives. A clear goal also helps you spot unnecessary output. If the decision only needs a one-page comparison, a polished 20-page report is not better. It is more work to create and more work to review. ### 2. Specify the output Tell the AI what shape the answer should take. That might be: - a table with named columns; - five bullet points ranked by commercial impact; - a draft email under 200 words; - a list of claims, each with a source and confidence note; - a structured record containing fixed fields. Structure improves usability because the result arrives closer to the way your team needs to use it. In technical systems, schema-constrained output can ensure that required fields are present and correctly formatted. That does **not** prove that the contents are true. Correctly shaped data can still contain a bad decision. ### 3. Define what good looks like A rubric turns an implicit judgement into explicit criteria. For a proposal, the criteria might be: - answers the client’s stated problem; - uses only evidence supplied in the brief; - distinguishes facts from assumptions; - gives costs in pounds and excludes VAT; - contains no unsupported performance claim; - ends with one clear next step. This does not magically remove bias. It makes the basis of the review visible and repeatable. It also makes disagreement more useful: instead of saying “I don’t like it”, a reviewer can identify which criterion the draft missed. We covered the practical use of criteria in [AI rubrics: ask what good looks like](/insights/ai-rubric-what-good-looks-like/). Loop engineering adds what happens next: checking, revising and knowing when to stop. ### 4. Check with evidence The strongest checks do not rely on the AI simply declaring that its own answer is correct. Research on AI self-correction is mixed. A critical review published in *Transactions of the Association for Computational Linguistics* found that unaided self-correction is unreliable, while correction works better when the model receives dependable external feedback. The [review explains the conditions under which self-correction can work](https://direct.mit.edu/tacl/article/doi/10.1162/tacl_a_00713/125177/When-Can-LLMs-Actually-Correct-Their-Own-Mistakes). Use the most objective check available: - verify a total with a calculator; - test code with an automated test; - check a link by opening it; - compare a quotation with the original document; - check mandatory fields against a list; - ask a qualified person to review a high-consequence decision. A test passing is evidence against a particular failure. It is not mathematical proof that the whole output is correct. ### 5. Revise, stop or escalate Feed specific failures back into the next attempt: “The draft has no evidence for claims two and four. Remove them or provide an authoritative source.” That is more useful than “try again”. Set a stopping rule before the loop starts. Stop when every required criterion passes, when the maximum number of attempts is reached, or when the work needs human judgement. Without a stopping rule, an automated process can keep polishing an answer without making it more useful. ## Where automated loops help — and where they do not Loop engineering is most useful when success can be checked. Formatting records, testing calculations, finding missing fields, checking links and applying an agreed house style all provide relatively clear feedback. Anthropic describes “evaluator-optimiser” workflows in which one model creates a response and another evaluates it against defined criteria. Its guidance also recommends starting with the [simplest effective agent or workflow](https://www.anthropic.com/engineering/building-effective-agents), rather than adding complexity for its own sake. The caution matters. A second AI is not automatically an independent expert. Two instances of the same model can share blind spots. Research has also found that iterative model-based evaluation can produce higher automated scores without a corresponding improvement in human judgement. External checks remain important. Use human approval where an error could affect someone’s rights, safety, employment, finances or reputation. The US National Institute of Standards and Technology’s [Generative AI risk-management profile](https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence) treats measurement, evaluation and human oversight as ongoing parts of managing AI risk, not as a one-off prompt-writing exercise. The practical boundary is similar to deciding [which jobs are worth handing to AI](/insights/jobs-worth-handing-to-ai/). Automate repeatable checking. Keep accountable judgement with a person. ## A template you can use this week Choose one recurring task where AI currently produces inconsistent work. Use this short brief: 1. **Goal:** What decision or action should this output support? 2. **Audience:** Who will use it, and what do they already know? 3. **Inputs:** Which documents or facts may the AI use? 4. **Output:** What format, length and fields are required? 5. **Rubric:** What five conditions must a good answer meet? 6. **Checks:** Which claims, figures, links or fields can be verified independently? 7. **Stopping rule:** When is the work accepted, rejected or sent to a person? Run the task once. Record the first failure you find. Improve the instruction or the check that should have prevented it. That is a useful loop: not endless AI self-reflection, but a working process that gets clearer after a real mistake. If you want to turn a promising AI experiment into a dependable business workflow, our [automation and workflow service](/services/automation-workflows/) starts with the process, controls and ownership rather than the tool. Or [get in touch](/contact/) and we will happily talk through the task you have in mind. ## FAQs ### What is loop engineering? Loop engineering is emerging shorthand for designing an AI workflow that creates an output, checks it against criteria, revises it when necessary and stops under defined conditions. It is not yet a formal or standardised technical discipline. ### How is loop engineering different from prompt engineering? Prompt engineering focuses mainly on the instruction and context given to a model. Loop engineering covers the wider process: the goal, output format, evaluation criteria, external checks, revision and stopping rule. ### Can AI reliably check its own work? Not on its own in every task. AI self-critique can help, but research shows it is more dependable when grounded in external feedback such as source documents, calculations, tests or human review. ### Do rubrics remove bias from AI outputs? No. A rubric can make expectations more explicit and reviews more consistent, but the criteria themselves can contain omissions or bias. Important rubrics still need human scrutiny and evidence. ### Does structured output make an AI answer accurate? No. A specified table or data structure can make an answer easier to process and can prevent missing fields. It cannot guarantee that the facts, reasoning or recommendation inside those fields are correct.

This article was written by Peter Lowe. The ideas and opinions are his own; AI was used to assist with drafting and editing.