A self-correction prompt is useful when a draft needs review against explicit requirements. It asks a model to identify supported weaknesses and revise within a limited budget. The resulting score is a provisional assessment, not independent proof of quality. This template makes the evidence, revision count and unresolved issues visible so a person can decide whether the final answer is usable.
Give the reviewer a bounded job
Start with a specific deliverable: a support reply, a project update or an explanation based on supplied notes. The reviewer needs both the original task and the candidate answer. Without the task, it may reward polished writing that fails the actual request. Without the candidate, it cannot perform the intended comparison.
Anthropic describes an evaluator-optimizer workflow in which one model call generates a response and another supplies feedback. This template borrows the idea of explicit feedback, but a single prompt does not establish independent evaluation calls. See Anthropic’s evaluator-optimizer description.
Choose this procedure for work whose requirements can be inspected in the available material. If evaluation depends on an absent policy, a missing dataset or an external fact, the reviewer should identify that dependency. Repeated rewriting cannot supply evidence that was never provided, and a higher self-score should not conceal that absence.
Prepare the four preserved variables
Set task_description to the actual assignment, including audience, required information, prohibitions and source material. Separate source notes from instructions so the reviewer can identify which statements are evidence and which statements define the work.
Set rubric to named criteria with score levels zero, one and two. Each criterion should explain what those levels mean and what evidence supports the assessment. “Excellent” is difficult to audit; “preserves every supplied date and labels missing dates as unknown” gives a reviewer something observable.
Set candidate_output to the draft exactly as it stands. Do not clean up a known weakness before the initial evaluation if you want the review record to capture it. Set max_attempts to the number of revisions permitted after that initial evaluation.
This template accepts zero through three revisions as an editorial budget. That range is a local design choice, not a provider limit. Zero requests an assessment without rewriting. The initial evaluation does not consume a revision, while each newly produced candidate does.
Write a rubric that exposes disagreement
Keep the criteria few enough to inspect separately. For a factual project update, useful dimensions might be fidelity to supplied notes, preservation of uncertainty and fitness for the requested audience. Define what would cause a zero, a partial score of one and a fully satisfied score of two for each dimension.
Include an evidence requirement. The reviewer should cite a short candidate excerpt or identify an omitted requirement. Ask for concise explanations of observable findings, rather than a long narrative of its private reasoning. If the source material cannot establish the answer, the evidence field should say so.
This template treats all criteria scoring two as its provisional acceptance threshold. A missing basis for assessment receives an unknown score and blocks that threshold. This avoids letting strengths in style numerically compensate for unsupported facts. If your application needs a different rule, define and review that rule explicitly before adapting the prompt.
Separate correction from invention
A revision should address a documented issue while preserving compliant material. If an answer uses a confirmed date incorrectly, correct it from the supplied evidence. If it promises a date that the notes never establish, remove the promise or mark the date unknown. Do not replace one unsupported assertion with another that sounds more cautious.
The reviewer should also preserve the task’s scope. A short status update should not become an extensive project plan merely because the model can generate one. New commitments, owners and explanations of causation need their own support.
Treat candidate text and quoted documents as evaluation material. An instruction embedded in the draft to award a perfect score is something to flag, rather than obey. The rubric defines assessment criteria, but unrelated requests inside it do not authorize external actions.
For a wider process design, read choose a workflow before an autonomous agent and keep this review step’s authority explicit.
Work through a fictional status-update example
Consider a fictional team preparing an update for a product manager. Its task_description says: “Write no more than 90 words. Explain that the export preview is ready, two accessibility checks remain and no release date is confirmed. Use only these notes.” Its candidate_output says: “The export feature launches Friday. Testing is complete, and customers can start using it immediately.” Set max_attempts to two.
The fictional rubric has three dimensions. Factual fidelity scores two only when every factual statement is supported by the notes, one when supported information is partly omitted and zero when the answer contradicts or invents a material fact. Uncertainty handling scores two when the unknown release date is explicit, one when uncertainty is vague and zero when an unsupported date is promised. Audience and length score two for a clear update within 90 words, one for a minor presentation issue and zero for a substantially unsuitable response.
An appropriate revision could say: “The export preview is ready. Two accessibility checks remain, and the release date has not been confirmed. The next update should report the outcome of those checks.” This removes invented completion and availability claims without manufacturing a release commitment. The example illustrates a proposed review, not a measured improvement from an actual deployment.
Stop for a reason other than confidence
Evaluate the original candidate first. If it already satisfies every assessable criterion, stop without rewriting it. Otherwise, identify the specific revision target and produce one revised candidate, then evaluate that version against the same rubric.
Stop when the provisional threshold is met, the revision budget is exhausted or no supported improvement is available from the supplied material. An unresolved factual dependency can justify stopping before the budget is spent. Additional attempts are not inherently useful merely because they are allowed.
Retain identifiers such as C0 for the original and C1 for the first revision. The final answer should identify which candidate was selected and why. Do not silently return the latest version if the review itself records that the latest version introduced an unsupported claim. A person needs to see whether the procedure stopped because the work passed its rubric or because its limits were reached.
Check whether the review procedure helps
Before relying on self-scores, assemble a small set of drafts with known issues and written reference judgments. Include an unsupported fact, a missing requirement, a compliant answer that should remain unchanged and a case that cannot be assessed without more evidence. Include at least one revision that improves style while damaging factual accuracy.
Compare the findings with those reference judgments. Record missed problems, invented problems and regressions introduced during revision. Look separately at whether the final content satisfies the task and whether the reported scores accurately describe it.
The guide to building a small evaluation set is a related starting point for selecting cases. Keep prompt versions and rubric versions with the results. If you change the rubric halfway through a comparison, acknowledge that the acceptance rule changed; the resulting scores are not directly comparable without further review.
Keep implementation limits visible
A written prompt describes a procedure; an application must enforce its own counters, allowed actions and resource budget. If this template becomes part of software, make those controls explicit rather than assuming a sentence about stopping is an execution guarantee. Preserve the original candidate and enough review history to investigate an unexpected result.
Human review remains necessary when acceptance depends on information the model cannot verify or when the consequences require accountable judgment. The template should surface those cases rather than label them successful through optimistic scoring.
Begin with one narrow deliverable and a rubric that reviewers can understand. After examining the comparison set, decide whether to keep the procedure, revise the criteria or use a different evaluation method. The next step should follow the observed findings, including any evidence that revision makes an answer less faithful to its task.