A detailed report on complying with the EU AI Act may look polished while leaving points unclear or insufficiently checked. Reflection is a prompting technique that asks a language model to review, critique, or improve its own previous output. It turns a first draft into the start of a revision cycle, rather than treating it as the final answer. section 6.3.3
Review, then revise
After generating a report draft, ask the model to identify what needs clarification, checking, or more detail, then produce a fuller response. For more control, split the work into two prompts: first score the draft for clarity, completeness, and actionability and return feedback as a JSON array; then pass the draft and feedback into a revision prompt. The review identifies work to do—it does not replace checking the report against regulatory texts and other relevant documents. section 6.3.3
Reflection is useful when a task needs more than a simple prompt or example to produce a coherent result. It differs from self-consistency: rather than comparing several generated candidates, it revisits an output and acts on its weaknesses. It can also follow retrieval-augmented generation to polish a document and address inconsistencies. section 6.3 section 6.3.2 section 7.3.3
In an agent workflow, reflection can examine the results of past actions and add conclusions to memory. A product-roadmap agent, for example, can retain lessons about which research sources led to more viable suggestions. This makes reflection part of planning and later improvement, but context limits can constrain how much past experience an agent can use. section 9.2.2 section 9.2.3 section 9.3.1
