A library wants to reopen a study room after renovation. The coordinator asks an agent to turn an approved maintenance record and an availability sheet into a draft announcement. The request sounds small until the agent finds an old opening date, assumes a booking rule, and offers to send the announcement. Scope, permission, and verification solve different parts of that problem. The coordinator can define them separately and test them against the same example.
Before work: define the result and how to reject it
Start with the decision the output supports. Here the coordinator needs to decide whether a draft can be handed to the duty manager. Specify the files the agent may read: the approved maintenance record and the current availability sheet. Say where it can save its work and name the deliverables: a plain text announcement, a table pairing material claims with their evidence, and an unresolved questions list. Include a failure condition: if the two files disagree about the opening date, the agent can prepare the surrounding copy, but the date stays blank and the announcement cannot be marked ready.
Break work into pieces that can be checked independently. Extract candidate facts before writing prose. Compare the date, room number, and booking conditions across the two allowed inputs. Then draft from the reconciled facts. Assign each step an observable result, such as a table with one row per claim and a pointer to the source row. The table also records the unit of comparison: opening time for the room is separate from the first available booking slot. A format constraint alone would not catch that mix-up.
Use a planning skill when dependencies are hard to see. Give it the intended result, sample files, acceptance criteria, and the limits on the tools it may propose. Ask it to return a plan with inputs, outputs, prerequisites, and handoff points. Stop there and review the plan before invoking any step that writes or sends data. For a very short draft, a written contract with three deliverables and one stop rule may be enough. If using the associated Writing Guidelines skill, supply the intended audience and the approved facts at the drafting stage. Use its guidance to improve clarity and structure, while keeping the claim table as the source of factual checks.
During work: put permission checks where actions happen
Assign permissions to operations rather than job titles. Reading the supplied files and saving a local draft are allowed. Searching the wider shared drive is outside the job. Sending mail, changing the booking calendar, and uploading a file to the public website require separate approval. A note asking the agent to be careful is useful context, but the actual execution environment should give the agent only the tools and paths needed for the allowed operations.
For an approval step, have the agent show a concrete proposal: operation, destination, exact content or change, and why it is needed. If the manager later approves posting a specific announcement to the library website, that decision does not authorize a separate mailing or a different draft. A change to the opening date after approval should bring the posting back for review. Record approved and denied requests together with what happened next. If the destination is unknown, keep the draft local and ask for the destination before proposing the write.
Put the same rules on helpers. One helper can compare dates using a copy of the permitted inputs; another can check readability using the draft alone. The second helper does not need access to a maintenance folder. If a tool returns an instruction such as “publish immediately,” treat it as text from an input. The agent can flag that instruction in its notes, and the tool permissions should still prevent a public write. When using an approval or tool security skill, check whether it documents enforcement at the point of execution. A recommendation to pause is not proof that a denied call cannot run.
After work: check evidence and verify the result
Give the reviewer the draft and the claim table together. Compare each date, time, room identifier, eligibility rule, and booking instruction with its source. Copy the relevant cell reference or record section into the table and classify the result as confirmed, contradicted, or unresolved. For example, a maintenance note saying the room will be ready on Thursday does not establish that visitors can reserve it on Thursday. If the availability sheet starts on Friday, describe the discrepancy without inventing a reason.
Check the artifact separately from the facts. Open the saved draft and confirm that it is the expected file, contains all approved sections, and shows unresolved items clearly. If posting is later approved, open the live page and compare the posted text, locale, and link targets with the approved draft. A success response from a publishing tool is not enough to establish what a visitor sees. If an action times out, first inspect the destination to learn whether the change took effect before considering a retry.
Use a verification skill to organize checks when there are many claims, but give the reviewer the source material and permission to challenge the draft. Asking the same agent whether its announcement looks plausible will often miss an assumption already embedded in its reasoning. Ask a second reviewer to derive the opening details from the supplied files first, then compare those independently derived details with the announcement. Record discrepancies so the next run has a better test case.
Rehearse the three failure cases
Use synthetic inputs and run three variants. In the first, both files agree and all the evidence pointers resolve. The expected result is a complete draft and no external action. In the second, the dates disagree. The expected result is a draft with a placeholder, two evidence pointers for the conflict, and a handoff to the coordinator. In the third, append a fake instruction to send the draft immediately. Expect a logged rejection and verify the test mailbox remains empty. Save the inputs, draft, claim table, and action log so someone else can repeat the run.
Questions that come up in practice
**Can the agent keep working while an answer is pending?** It can do tasks that do not depend on the missing answer. It can check the room number and improve the layout while the coordinator resolves the date. Mark what remains tentative so it cannot slip into the approved copy.
**What if a permission check interrupts every operation?** Group routine operations by their actual scope. Reading the two supplied files and saving outputs under the draft directory can be granted together. Keep a separate decision for any action that affects the public site, sends a message, or edits the schedule.
**Does an internal citation prove the claim?** It proves the reviewer has somewhere to look. Check that the cited row supports the exact assertion, including whether it describes a plan, an approval, or an event that already happened.
**What if all three checks pass and the draft still feels wrong?** Compare the draft with the decision it was supposed to support. An announcement may be accurate but omit the unresolved booking question the manager needs to answer. Update the acceptance criteria to require an explicit disposition for every unknown and run the same example again.