Start with a process someone can finish
Choose one trigger and one observable end state. For the returns example below, the trigger is a customer return request. The end state is a closed support ticket after Finance has recorded a refund ID. “Handle returns” would be too broad if the same document also tried to define international shipping, damaged-item decisions, and every refund policy. Those rules need an authorized source before they can become instructions.
Collect the current policy or approved checklist, the forms or system instructions people actually use, and a walkthrough from the operator who handles the handoffs. Ask a newer operator where they have to guess. Label each source with its owner and date, and separate a proposed change from the rule currently in force. If records contain customer data, use an approved, redacted copy under your organization's AI-use rules.
| What you have | What ChatGPT can draft | What remains open |
|---|---|---|
| Current approved policy and operator walkthrough | A procedure with steps tied to the policy and observed handoffs | Any difference between the policy and current practice goes to the owner |
| Interview notes only | A draft of the observed sequence, with interview-only rules marked | Authority, timing, and controls that no approved source confirms |
| Two conflicting versions | A comparison of the conflicting lines and affected steps | Which version governs; do not ask the model to choose |
Use a prompt that exposes missing rules
Paste only material you are allowed to share. Give each source a short ID, such as P1 for an approved policy and I1 for an interview. OpenAI's prompting guidance recommends supplying the task, relevant context, and the form of the answer. For an SOP, the important extra constraint is to show where a step came from and where the source stops.
Draft an SOP for [one named process]. The start is [trigger]. The end is [observable completion condition]. The users are [roles].
Use only the numbered source notes below. For each procedure step, give the role, action, decision condition if any, expected result, and source ID. Name a completion record only if the source names one. If a field or next action is missing, write [VERIFY: specific question] instead of inventing it.
Include purpose and scope, prerequisites, numbered steps, exceptions, required records, open owner decisions, and a blank approval/version block. Keep a proposed rule separate from an approved rule. Do not create deadlines, system fields, thresholds, permissions, forms, or approval outcomes.
After the draft, return a short audit table: each source ID, the SOP step that uses it, and any unresolved conflict or gap.
Sources:
[P1] [Paste dated, approved policy excerpts]
[I1] [Paste dated, approved-for-sharing operator notes]
[F1] [Paste current form or system instructions]The prompt is a starting structure, not a guarantee. Check the returned audit table against the originals yourself. If the model supplies a step without a source ID, ask where it came from or remove it. A plausible completion record is still an invented record when the source only describes an action.
Worked example: a returns and refunds procedure
This example uses fictional process notes. They were submitted in a ChatGPT web test on August 18, 2026. The retained test summary says the response preserved the three timing rules and left the damaged-item and international-return gaps open. The screenshot shows only the beginning of the response, so the complete procedure below is an editorial reconstruction from the notes, not a verbatim ChatGPT output or an approved returns policy.
Here is the source packet used for the reconstruction. Keeping it next to the draft lets a reviewer check every role and deadline without trusting the model's summary.
R1 Support confirms the order number and 30-day eligibility. Missing receipts go to the support lead.
R2 Support creates an RMA and emails a return label. The warehouse logs arrival within one business day of receipt and grades the item as unopened, opened, or damaged.
R3 Finance refunds unopened or opened items within five business days after grading. Damaged items require operations-manager review before any refund. Support closes the ticket only after Finance records the refund ID.
R4 The international-returns path and its service level are unresolved.
A reviewer can turn those notes into the following bounded sequence. The bracketed questions are part of the draft; they show where the source is too thin to direct a worker.
| Draft step | Role and action supported by the notes | Source / check |
|---|---|---|
| 1 | Support confirms the order number and 30-day eligibility. | R1. [VERIFY: Which date starts the 30-day window, and what happens if the request is ineligible?] |
| 2 | If the receipt is missing, Support sends the case to the support lead. | R1. [VERIFY: Does the RMA wait for a lead decision, and where is that decision recorded?] |
| 3 | Support creates an RMA and emails the return label when the case can proceed. | R2. “When the case can proceed” depends on the unresolved decision in step 2; do not silently treat escalation as approval. |
| 4 | Warehouse logs arrival within one business day of receipt and grades the item as unopened, opened, or damaged. | R2. [VERIFY: Which system holds the arrival and grade records?] |
| 5 | Finance refunds unopened or opened items within five business days after grading. | R3. The clock begins at grading, not at the customer's request or warehouse arrival. |
| 6 | For a damaged item, Operations Manager reviews the case before any refund. | R3. [VERIFY: Review criteria, decision record, outcome, refund owner, and timing after review.] |
| 7 | Support closes the ticket after Finance records the refund ID. | R3. [VERIFY: Where can Support confirm that the ID was recorded?] |
The source does not say whether a missing receipt makes a return ineligible. It also says nothing about an international label, carrier, owner, or service level. Those omissions should stay in an owner-question list. Writing “follow the usual international process” would hide the fact that no such process was supplied.
Check the draft against the source, not its fluency
Read the draft with the source packet open. First compare the order of work and every number. Then check each branch and handoff. The August test summary reported that its ChatGPT response kept the 30-day eligibility rule, one-business-day warehouse logging, and five-business-day refund timing. That observation applies to that one synthetic run. It does not verify a real company's policy or show that a new run will make the same choices.
| Check | Pass only if | Reject or return to owner when |
|---|---|---|
| Timing | The 30-day rule stays attached to eligibility, the one-business-day rule to arrival logging, and the five-business-day rule to grading. | The draft changes a number or starts the clock at a different event. |
| Authority | Support lead handles the missing-receipt escalation; Operations Manager reviews damage before a refund. | ChatGPT treats either review as an automatic approval or gives that authority to Finance. |
| Closure | Support closes only after Finance has recorded the refund ID. | Ticket closure moves before the recorded refund ID. |
| Unknown path | International returns and the result of damaged-item review remain explicit questions. | The draft invents a carrier, refund deadline, decision rule, or approver. |
Make one correction at a time and recheck the affected step against its source. If a corrected paragraph also mentions another role or deadline, check those fields again. Keep the source packet and the correction record with the draft so the approver can see why a rule appears.
Walk through normal work and the branches that stop it
A readable draft can still fail when someone tries to use it. Give an operator a safe fictional case and ask them to follow the SOP without verbal hints. The walkthrough sheet contains seven returns scenarios with expected actions grounded in R1 to R4. Its observed-result cells are deliberately blank: no operator walkthrough was performed for this editorial example.
| Scenario | What the draft should let the reader do | Stop point or missing decision |
|---|---|---|
| Unopened item with a receipt | Follow the order check, RMA, warehouse grade, Finance refund, refund ID, and ticket closure. | Confirm the actual system records and the precise eligibility calculation before live use. |
| Receipt missing | Escalate to the support lead. | The notes do not say whether the RMA may be created while the lead reviews. |
| Damaged item | Send for Operations Manager review before any refund. | No review criteria, approval result, post-review refund path, or timing was supplied. |
| International return | Identify it as outside the documented path. | No owner or service level is supplied. Do not reuse the domestic label step by guesswork. |
Also try an ineligible request, a duplicate RMA, and a system outage. The notes provide no handling rule for those cases, so a passing draft must expose the gap, name the process owner who needs to answer it, and refrain from directing an invented action. The owner can then approve a rule in the controlled source before it appears in the SOP. A reviewer should record where the test operator hesitated, which branch they chose, and what source or instruction needs to change.
Approve and maintain the finished SOP
Move a draft to live use only after an authorized process owner resolves the open decisions, a second operator can complete the relevant safe walk-through, and the required approvals are recorded. The EPA's SOP preparation guidance describes review by people experienced with the process, testing by someone other than the writer, approval, document control, and revision when procedures change. It is guidance for EPA quality-system work, so use your organization's actual approval rules for this procedure.
Keep the SOP's owner, effective date, version, approver, next review date, and linked controlled forms visible. When a rule changes, update the authoritative policy or instruction first, revise the affected steps, repeat their walkthroughs, and record the change. If the owner cannot answer who may authorize a damaged-item refund or how international returns proceed, leave those branches marked [VERIFY] and do not release them as usable instructions.
