Procedure
A rubric for procedure models
If the model cannot name the stop, it has not written a procedure.
A procedure is a list of actions with a place to stop. I did not learn that from a writing course. I learned it from QA steps that were too short to be wrong in an interesting way and too vague to be done. “Review the as-built against the field packet and update the GIS as needed.” Everyone who already knew the job could nod at that sentence. Nobody could hand it to a new analyst and leave the room.
Models produce that sentence because a procedure prompt asks for clarity, and the model hears “sound like a policy.” Policy language avoids the branch where the work stops. “As needed” is the branch, hidden. Needed by whom, and what happens on the day it is not? The draft looks responsible because it mentions review, update, and quality in one line. It cannot be followed on a morning when a packet is on the desk and the feature is open.
What I require beyond the six checks
I still score a procedure with the six checks I use on a paragraph: source of record, load-bearing words, fake precision, followability, audience, and leftover AI. A procedure has one more demand. The steps must be numbered. One action per step. The tool or the record named in the step. And a stop condition written as its own step, not as an adverb. If the model cannot say when to stop editing, it has not written a procedure. It has written a wish that the editing go well.
Here is the vague form, the kind I give a model and the kind a model gives back:
Review the as-built against the field packet and update the GIS as needed to ensure data quality before closing the work order.
It fails followability. “Review” has no object list. “Update” has no permission. “As needed” is a decision with the decision removed. “Ensure data quality” is leftover purpose, the clause that closes a paragraph when the writer is done thinking. I mark STEP on the instruction and PAD on “ensure data quality.” I do not start the repair by hunting for a warmer verb.
The stop is allowed to be the point
The repair is a sequence.
- Open the work order and the field packet. Confirm the packet contains the LocusView as-built.
- Compare three items only: material, the station or length given on the as-built, and whether the old service was abandoned in place.
- If any of the three disagrees with the feature, stop. Do not edit the feature. Note the disagreeing item and return the packet.
- If all three agree, post the as-built to the feature and record the packet identifier in the note.
The stop is step 3. It is allowed to be the whole point. In gas-utility QA, a useful outcome is often that you did not post. A model that has been rewarded for being helpful will try to end on the update, because an update looks like service. I cut that ending when the records do not match, and I leave the stop sounding like a stop. “Do not edit” is a complete instruction. It does not need a sentence after it about maintaining excellence.
Numbering is part of the rubric. A numbered step can be resumed after an interruption. I also refuse “as discussed above” and “following best practices.” A step that points at an unstated practice is a TERM failure. If the practice is “compare material, station, and abandon status,” those words belong in the step.
Audience, then precision
Two other failures show up once the steps exist. The first is audience. A procedure for a mapper should name the feature and the attributes. A procedure for a coordinator should name the handoff: who receives the packet, and what they are waiting on. I do not mix them to save lines. A mixed procedure trains the model to write for a reader who does not exist, someone who both posts geometry and manages the crew’s schedule. When I see “notify stakeholders” in a step, I mark AUD and replace stakeholders with the role that actually receives the packet.
The second failure is fake precision inside a step that is otherwise sound. If I do not have a real tolerance, I do not write “within 0.5 feet.” I write “as stated on the as-built,” or I name the three items and stop. An invented tolerance will be copied. INV on a procedure is costly because the procedure is repeated.
Label first
When I train someone on this, I ask them to label before they rewrite. I hold myself to the same order. The vague sentence goes on the left. The marks go in the margin: STEP, PAD, sometimes AUD if the draft addressed a crowd instead of the mapper. Only then do we write the numbered version. If we start with the clean list, the learner memorizes my sentences and cannot see why “as needed” had to go. The rubric is the thing worth repeating. My wording of step 2 is not sacred. The rule that step 3 must be a real stop is.
After the steps are clean, a model will often add a closing line about a seamless close-out. That line is PAD. A procedure is finished when the last step tells the person what to do, including the case where the correct act is to stop and hand the packet back.
A packet either matched or it did not. Give the model the records it may mention. Require a number on each step and a real stop. If I can hand the list to someone who has not seen the ticket, and they can post or refuse without calling me, the procedure is done.