# Build an Evidence Packet Before an AI Agent Drafts Social Posts [Read the original article]() By Garry · Founder Published: 2026-09-29T20:10:03.000Z Updated: 2026-09-29T20:13:22Z 6 min read Categories: Content creation An evidence packet gives an AI agent a small, checkable basis for every factual sentence it drafts. ![An ivory archive envelope opens into a precise fan of mint paper sheets, with a pale blue glass magnifier over one sheet.]() ## Key takeaways - Include the reader's question, the exact approved facts, the source for each fact, the product version, and the limits the draft must preserve. - Ask for a draft accompanied by a claim ledger that maps factual sentences to the packet. - Save the selected text after its material claims and intended use have been checked. An evidence packet gives an AI agent a small, checkable basis for every factual sentence it drafts. It connects a social post to approved material before the agent saves anything in Caroush. The deliverable is a claim ledger and a draft, with a visible distinction between what the source says and what the writer proposes. Consider a fictional appointment-software business explaining its cancellation settings. Its source material includes a current help page, a screenshot from an approved demonstration account, and a product owner's note about plan availability. An agent could turn that into a useful explanation. It could also accidentally promise automatic refunds, even though the sources only describe a cancellation window. The packet makes that jump easier to catch. ## What belongs in an agent's evidence packet? Include the reader's question, the exact approved facts, the source for each fact, the product version, and the limits the draft must preserve. Keep writing preferences separate so that a request for a friendlier tone cannot quietly rewrite the offer. Give each fact a short working label in your editorial document. A cancellation window, its applicable appointment types, and its exceptions are separate facts. Attach a checked date and a responsible person to details that change. A link alone is insufficient when a page contains several products or revisions; quote or summarize the relevant passage and retain its location. The packet should fit the assignment. A single text post does not need a full export of your knowledge base. [Anthropic's context-engineering guidance]() treats context as a finite resource that needs deliberate selection. In editorial work, select the material that helps answer this reader's question and preserve the facts that constrain the answer. Start with your [existing brand voice guidance](), then supply two or three approved examples of tone. Label those examples as style references. Otherwise an agent may copy their old offer details into a new post. Include prohibited inferences such as “do not describe a refund unless a source explicitly confirms it.” ## How should the agent connect claims to sources? Ask for a draft accompanied by a claim ledger that maps factual sentences to the packet. Unsupported statements should become questions for the fact owner, rather than fluent guesses inserted into the caption. For the fictional software example, the ledger might identify “set a cancellation window” as supported, “applies to all bookings” as unverified, and “refunds are automatic” as excluded. This is an editorial side document; it does not need to appear in the social post. Its job is to keep the reasoning inspectable while the caption remains readable. A useful instruction is: > Draft one explanation for an appointment coordinator using only the supplied evidence. Preserve the eligibility limits. Under the draft, list each factual claim and its source label. Put missing information in a separate questions list. Do not create a Caroush draft until I select the wording. Check the ledger against the source yourself. A model can attach a plausible citation to an unsupported sentence. A correct URL is not proof that the page supports the precise claim. Inspect numbers, negations, comparisons, availability, and cause-and-effect language especially carefully. [Google's helpful-content guidance]() asks whether content adds original value rather than merely rewriting sources. That guidance concerns search content, but the editorial test is useful here: the social post should explain the decision a reader can make. A list of correctly cited features may still fail to answer their question. ## When should the agent create a Caroush draft? Save the selected text after its material claims and intended use have been checked. Creating a draft is a separate action from approving publication, so preserve those two decisions throughout the handoff. The documented create\_text\_post tool saves text in Caroush and does not publish or schedule it. Before using a supported connection, confirm the authorized workspace and choose the reviewed caption. Consult the current [Caroush tool catalog]() for required inputs and available scopes. A connection must be functioning and authorized before a tool-based workflow can proceed. Store the returned post identifier in your production record. Keep the evidence packet and its version beside that identifier in your own editorial system. Do not assume Caroush stores every source note merely because the caption was saved. The agent's chat history should not be the only place where your reviewer can find the evidence. If you are using the [AI social media generator]() through the normal application, the same packet remains useful. Paste the selected copy into the reviewed composer and retain the record manually. This lets you preserve the editorial process when your chosen assistant lacks a verified Caroush connection. ## What should a reviewer reject before the next step? Reject drafts that strengthen a claim, hide a condition, attribute an experience to someone who did not provide it, or cite a source that does not support the sentence. A polished caption is still unfinished when those issues remain. Compare the final draft with the packet sentence by sentence. “Lets coordinators set a window” and “eliminates last-minute cancellations” are different claims. The second promises an outcome the first feature cannot establish. Ask whether each changed verb increases certainty, broadens eligibility, or implies a measured result. Also check the reader's likely interpretation. A condition in the evidence ledger cannot rescue a misleading public caption if the audience never sees it. Place the qualification where it helps the reader understand the offer. If the format cannot carry the necessary explanation, choose a different format or a narrower claim. Use your [social media approval workflow]() to assign the final factual decision. The agent can identify possible problems, but its approval of its own copy is not independent evidence. The product owner should resolve product facts; the publisher should confirm the destination and exact saved version. ## Keep the packet useful after the first post An evidence packet should support controlled reuse, not permanent trust. Add an expiry trigger for changing details: a product release, pricing update, permission change, or retired help page. When a trigger occurs, search your editorial record for affected drafts before reusing their language. Do not ask the agent to regenerate every related post automatically. First identify which claim changed and which published or planned pieces actually depend on it. A revised plan limit may affect an offer caption but leave a general educational explanation intact. This smaller review reduces unnecessary rewriting and makes the correction easier to verify. A [content audit]() can supply the larger maintenance process. The agent's useful contribution is a bounded list of potentially affected items, the relevant source version, and the proposed next decision. The owner still checks whether the dependency is real. For your first run, choose a low-complexity topic with a single fact owner. Finish one packet, one claim ledger, and one saved draft before expanding the process. Review how often the agent asked a useful question versus inventing a connection. Improve the packet where ambiguity appeared. The goal is a repeatable chain from evidence to wording to a reviewable content record, with enough context for another person to continue the work. ## A practical packet review Before handing the appointment-software packet to the agent, ask a colleague to find the source for one selected sentence without your help. If they cannot locate it quickly, improve the labels or excerpts. Then remove one source deliberately and run the draft review again. The expected behavior is a visible question or an omitted claim, not a confident reconstruction from general knowledge. This small exercise tests the working process, not the model's general intelligence. Record where the draft lost an exception or where the reviewer could not tell whether a statement was a proposal. Fix that ambiguity in the packet itself. A useful source record makes correct behavior easier for both the assistant and the human who takes over later. ## Sources - [Effective context engineering for AI agents \\ Anthropic]() - [Creating Helpful, Reliable, People-First Content \| Google Search Central \| Documentation \| Google for Developers]() - [Caroush public tool catalog]() ## Frequently asked questions ### What belongs in an agent's evidence packet? Include the reader's question, the exact approved facts, the source for each fact, the product version, and the limits the draft must preserve. Keep writing preferences separate so that a request for a friendlier tone cannot quietly rewrite the offer. ### How should the agent connect claims to sources? Ask for a draft accompanied by a claim ledger that maps factual sentences to the packet. Unsupported statements should become questions for the fact owner, rather than fluent guesses inserted into the caption. ### When should the agent create a Caroush draft? Save the selected text after its material claims and intended use have been checked. Creating a draft is a separate action from approving publication, so preserve those two decisions throughout the handoff. ### What should a reviewer reject before the next step? Reject drafts that strengthen a claim, hide a condition, attribute an experience to someone who did not provide it, or cite a source that does not support the sentence. A polished caption is still unfinished when those issues remain. ## About the author Garry Gaurav Sapkota builds Caroush, a workspace for creating, scheduling, and publishing social content. - [https://x.com/gauravsapkotanp]()