How to write a standard operating procedure with ChatGPT

Start with the work, not a blank template

It is tempting to type 'write an SOP for order processing' and accept a neat document. But order processing means different things in different businesses. One shop reserves goods before payment. Another does not. One person can approve substitutions. Another cannot. If you leave those facts out, the draft may supply sensible-looking rules that are wrong for your team.

Choose a job small enough that someone can point to its finished result. 'Manage the shop' is too broad. 'Prepare a paid collection order' has a beginning, a finish, and a visible handoff. You can join several procedures later. Starting with a narrow job makes it easier to spot missing steps and prevents the procedure from becoming a long description of everything the business does.

Ask the person doing the work to talk through a recent normal case. Write down actions in their order, even if the notes are messy. Then ask: What do you check before starting? What usually makes you pause? Who can decide what happens next? What do you leave behind so the next person knows the job is done? Those questions uncover knowledge that experienced staff often skip because it feels obvious to them.

Build a small evidence pack

Your evidence pack is simply the material the draft must follow. It can contain rough notes, a blank form, an approved message, and a description of the tool states. It does not have to be a huge file. In the collection example, the meanings of PAID and READY matter more than a screenshot of the whole account. Use harmless sample IDs instead of real orders when you only need to explain the structure.

Separate a fact from a suggestion. 'The assistant may not offer a refund' is a current boundary. 'It would be faster to let assistants approve small refunds' is an improvement idea. Label the second statement as a proposal outside the SOP. ChatGPT should not choose between the two because the newer-looking idea sounds easier.

Look for missing cases before drafting. The shop notes explain a missing item, but do not explain a damaged item or a payment status that changes after packing starts. Do not let the draft invent those paths. Ask the shift lead what the current rule is, or state that the worker stops and contacts the lead when the case is outside the documented scope. This is a real stopping rule, not a vague instruction to 'use your judgment.'

• Confirmed: process only PAID orders; match item codes and quantities; notify only after the approved condition is met.
• Boundary: no substitution or refund decision by the packing assistant.
• Unknown: how long held items stay reserved; who covers when the shift lead is absent.
• Proposal outside the procedure: add a shelf location to the order sheet to save searching.

A usable SOP from the collection notes

The following worked draft uses the confirmed training notes. It does not claim to be a procedure used by a real shop. Notice how the exception appears where the worker needs it, instead of being hidden in a final warning. A worker should not have to read the entire document again to decide what to do about one missing item.

The purpose is to prepare complete paid orders for collection while avoiding unapproved substitutions and early messages. The scope is packing only. Payment disputes, refunds, and changes to customer orders stay outside the assistant's role. The process starts with an order in PAID and finishes with either a complete READY order or a held order passed to the lead for a decision.

SOP: Prepare a paid click-and-collect order. Owner: shift lead. Role performing the task: packing assistant. Prerequisites: access to the paid order queue, order printer, collection tray, holding tray, and approved collection message. Step 1: Open the paid order queue and select an order marked PAID. If the order is not marked PAID, do not pack it under this procedure. Step 2: Print the order sheet. Keep it with the goods during packing. Step 3: For each line, match the shelf item code to the order sheet and count the requested quantity. Step 4: If every item and quantity is available, place the complete order in the collection tray. Record the order ID, packing initials, time, and number of packages. Step 5: Mark the complete order READY and send the approved ready-for-collection message. Exception: If any requested item is missing, leave the order in PAID. Place any found items in the separate holding tray and contact the shift lead with the order ID and missing item code and quantity. Do not substitute goods, offer a refund, or send the ready message. The shift lead records the decision and approves any different customer message. Stop rule: For an unclear order state or a case not covered here, stop and contact the shift lead before changing the order or messaging the customer. Completion evidence: the packing record and the correct order state, or the lead's recorded decision for a held order.

Find instructions that sound clear but are not

Words such as check, handle, verify, promptly, and escalate often hide the actual job. 'Check the order' could mean count items, check the payment, confirm the address, or all three. Replace it with what the worker must compare. In the example, matching the item code and counting the quantity are observable actions. A colleague can say whether those actions happened.

'Escalate the issue' means ask someone with the authority to decide, but it still leaves questions. Who is that person? What information do they need? What should remain untouched while the worker waits? The held-order path answers those questions. It names the shift lead, calls for the order ID and missing item details, and keeps the state in PAID. If the lead's backup is unknown, flag it rather than inventing a manager name.

Do not force every action into one sentence if that makes the instruction harder to follow. A main action can have a short explanation or condition. What matters is that the worker can see the sequence. Use a table for several decision paths, a checklist for repeated checks, or a picture of the actual form when that explains the work better. Format serves the job.

Walk through both branches before using the document

For a normal training case, give the worker order TRAIN-104: two blue A5 notebooks, item NB-A5-BL, and one black pen, item PN-BK. Both items are available. Ask them to describe each step using the procedure. The expected finish is a complete order in the collection tray, a record containing the required fields, READY status, and the approved message. The order sheet and the package count should agree with what was packed.

For an exception, use TRAIN-105 with three green folders, item FD-GR, when only two are available. The correct result is not 'pack two and mark ready.' It is a held order, unchanged PAID state, missing quantity reported as one, no substitution, and no ready message. That example checks the boundary that a generic draft is most likely to lose.

This article supplies those expected outcomes as a worked example; it does not report a live workplace test. In your workplace, ask a colleague to use the actual draft with a harmless practice case. If they need you to explain an unwritten rule, add the rule where they got stuck. The goal is to remove dependence on the author's memory, not to see whether the colleague can guess correctly.

Repair a specific failure instead of rewriting everything

Suppose the first draft says, 'If stock is short, offer an alternative product and continue packing.' That is not just a tone issue. It changes authority and could lead to the wrong goods being supplied. Tell ChatGPT exactly which source boundary it broke: the assistant must not substitute products, and the shift lead decides what happens next.

A useful correction request is: 'Replace the shortage path using only the confirmed notes. Preserve PAID, hold found items, name the shift lead, include the missing quantity, and block both substitution and the ready message. Do not change the normal path.' Then compare the revised path against the notes. This narrow request makes the important change visible and avoids accidentally changing steps that already work.

If ChatGPT adds a five-minute response deadline, a new ticketing system, or a refund threshold, ask where each rule came from. A blank answer means it is not part of your documented process. You can keep worthwhile ideas in a separate improvement list, with someone responsible for deciding them. Do not hide new business decisions inside better wording.

Keep the working copy easy to trust

Give the finished SOP a name, owner, version, and effective date. These details solve a practical problem: two colleagues may otherwise use different copies. Store the current version where the work happens, and mark older copies as superseded. A long approval history is usually less useful to the worker than a short statement of which version to use and whom to contact.

Define the events that should trigger an update. For this example, a change to order states, who can approve substitutions, the collection message, or the packing record would affect the procedure. A new logo on the order sheet probably would not. Review the affected steps when the process changes instead of rewriting the document on a schedule with no reason.

A procedure is ready when the worker can follow it, its decisions match the actual business rules, and its completion evidence shows what happened. Keep the wording plain enough to use on a busy day. If the document only makes sense after the owner gives a long explanation, it has not yet captured the process.

Your next useful step

Use the completed procedure to build a short training exercise: one normal case, one blocked case, and the records a worker should produce. For the wider task list and ownership, continue with using ChatGPT at work.