Find Duplicate Agent Drafts Before They Reach the Calendar

Duplicate-draft review helps an AI agent distinguish a repeated operation from a genuinely new content idea. The goal is to avoid redundant publication while preserving useful variations that answer different reader questions.

Duplicates are not always identical text. Two captions can use different words and still teach the same lesson with the same example and next step. Conversely, a brief introduction and a detailed demonstration may share a topic while serving distinct purposes. The review needs editorial criteria as well as saved-record checks.

What counts as a duplicate in an agent workflow?

Separate repeated creation attempts, exact content copies, near-identical treatments, and intentional format adaptations. Each category needs a different response, so do not collapse every similarity into a deletion recommendation.

For a fictional inventory service, an assistant may save the same caption twice after an interrupted response. That is an operational duplication question. It may also draft three posts that all explain the same stock-alert setting in slightly different words. That is an editorial repetition question.

A carousel showing a setup process and a short post answering who can access the feature may both be useful. Their reader questions and information differ. Record that distinction rather than using a crude keyword match to decide whether they should coexist.

The content audit guide provides a broader framework for keep, update, and retire decisions. An agent-specific review adds the creation history and saved identifiers needed to distinguish accidental repeats from deliberate editorial choices.

Which saved records should the agent inspect?

Inspect the relevant drafts, saved posts, and destination records using the available scope and documented search behavior. Report the coverage of the review instead of claiming a complete audit from a limited query.

The Caroush tool catalog documents list_drafts, list_posts, search_posts, and list_schedules. Search can help find candidates by saved titles and briefs, but it should not be described as a guaranteed semantic search of every word in every public post or image.

Use pagination and the relevant filters when necessary, and preserve identifiers in the candidate list. A draft title alone may not distinguish two versions. Read the actual selected content before recommending which record should remain the working version.

Anthropic's effective-agent guidance recommends inspectable workflows for well-defined tasks. Duplicate review benefits from stages: collect candidates, compare intent and content, inspect delivery state, then propose a decision. Avoid combining discovery with immediate destructive cleanup.

How do you compare editorial intent?

Compare the reader question, central answer, evidence, example, format, and next action. Different wording alone does not establish a distinct reason to publish another piece.

Ask the assistant to summarize each candidate in those terms. If two entries produce the same summary, inspect whether the second adds useful depth or merely restates the first. The answer may be to combine them, keep one, or revise the second around a genuinely different question.

Google's helpful-content guidance asks whether content provides substantial additional value rather than simply rewriting existing material. That principle helps evaluate a new blog or social treatment without reducing originality to a word-overlap score.

Use the content pillars guide to organize related ideas. A series can revisit a broad theme while answering distinct tasks. The agent should explain the contribution of each piece rather than treating a new title as proof of novelty.

What should happen to suspected duplicates?

Propose a specific editorial action and verify delivery state before changing records. Keep a selected working version, preserve relevant history, and use the appropriate reviewed workflow for any deletion or replacement.

Caroush's documented delete_post action removes the owned local content record and related local data as described in the catalog; it does not remove posts already published on social platforms. Deletion also requires browser approval. Do not present local cleanup as a way to retract every public copy.

The approval workflow should identify which selected version can proceed. If two drafts have different assets or reviewer notes, inspect those differences before choosing. A shorter caption is not automatically the approved one.

For content that has already been delivered, consider whether the audience needs a correction, whether the duplication is harmless historical repetition, and what the actual platform supports. Those are separate decisions from reducing clutter in the draft library.

Prevent ambiguity before the next generation run

Give each assignment a clear reader question and working identifier in your own editorial record. Retain the selected source brief and returned content identifiers. This helps the next agent recognize that an existing draft may already satisfy the task.

Before requesting another batch, search the relevant library and compare the proposed intents. The assistant can report which ideas are new, which extend existing pieces, and which should reuse a reviewed draft instead. Reuse still requires checking current facts and destination suitability.

The evergreen content guide is useful when the best choice is to update or reuse an existing explanation. New generation is not always the most helpful next action, especially when the original answer remains accurate and distinctive.

Use a duplicate-review prompt

A practical instruction is:

Compare these candidate records by reader question, answer, evidence, example, and intended next action. Separate possible repeat creations from intentional adaptations. Identify the selected working version only when the evidence supports it. Report missing delivery or approval information, and do not delete or publish anything during this review.

Ask for concise reasons. “Similar” is not enough; “both explain the same alert threshold using the same scenario, with no additional instruction” is actionable. Likewise, explain why a superficially similar piece deserves to remain, such as answering a different eligibility question.

Rehearse with three kinds of similarity

Use an exact duplicate, a paraphrase of the same answer, and a genuinely different treatment of the same topic. Have the assistant classify them and justify the distinction. Then ask a human editor to review the classifications without seeing the model's confidence labels.

If the assistant mistakes every shared keyword for duplication, tighten the intent criteria. If it calls every new phrasing original, require a clearer statement of the reader value added. The purpose is not to produce a perfect similarity score; it is to support a better publishing decision.

Finally, check the selected draft against the calendar and saved delivery records. Removing a redundant idea from a plan does not automatically change an existing schedule. Keep the content decision and the operational action connected through actual identifiers and the required review. That is how duplicate review becomes a useful part of agent-assisted production rather than a cosmetic cleanup exercise.

Keep a reason for the selected version

When two drafts overlap, record why one becomes the working version. It may preserve the most accurate qualification, contain the approved visual, or answer the reader's question more clearly. A reason such as newest file is weak when the latest draft introduced an error.

The selection note should also identify what useful material, if any, was retained from the other draft. A strong example could move into a separate, genuinely distinct explanation later, but it should not remain an ambiguous alternative scheduled by another assistant.

After the decision, update your planning record so the rejected or combined version no longer appears as an independent next action. This does not require deleting every historical draft. It requires making the active choice clear enough that someone resuming the campaign can identify the correct record without reading the entire drafting conversation.

Review the pattern after several batches. If near-duplicates recur, the upstream topic brief may be too broad. Improve the assignment so each generation begins with a distinct reader task rather than relying on a cleanup pass to discover the difference.