Drafting an SOP From a Recorded Tech Walkthrough

Why this matters

Your best tech knows how to do the job and cannot write it down. Not because they are bad at writing, but because the parts that matter most have become automatic. Ask them for a procedure and you get eight lines, none of which mention the thing they do every single time without thinking: the photo of the valve positions before closing anything, the bucket set lower than the unit, the glance at the date code on the cartridge. That automatic knowledge is called tacit knowledge, and it is exactly what leaves the shop when the senior tech does.

Recording the tech narrating a real job captures most of it. AI transcription and drafting turn an hour of talk into a structured draft in minutes. But the draft is the halfway point. A transcript has gaps where the tech worked silently, and an AI draft fills gaps with generic, confident text that the tech never said. The step that turns the draft into something you can trust is a ride-along where someone else performs the job from it while the author watches and corrects. This card walks the four phases using one real example carried all the way through.

Phase 1: Record a real job, narrated

Pick a representative job, not a showcase. A routine visit of the type the SOP will cover, at a typical house, with the usual problems. A demonstration at the shop on a clean unit skips the conditions that make the job hard.

Give the tech three things to say out loud, and only three, so narration does not slow the work: what I am doing, what I am looking at, and what would make me stop. The third one is the gold. "If this valve does not turn with hand pressure I leave it alone and note it" is a stop rule, and stop rules are what SOPs most often lack.

Bring a second person to ask "why". Not constantly, but whenever the tech does something without saying anything. "Why did you photograph that?" gets you a step that would otherwise be invisible on the recording.

Do not interrupt at the hazardous moments. Agree beforehand that during isolation, opening, and anything with pressure, heat, chemicals or energy, the tech says what they are doing and the questioner stays quiet until it is done. The person filming stays out of the discharge path of any hose or valve and out of the tech's working space; a phone on a chest mount or a fixed tripod works better than a handheld camera at arm's length.

Get the customer's permission first, say what will be filmed, and keep their faces, documents and the inside of their home out of frame. Audio recording of people's conversations is governed by state law, and some states require every party's consent. See related: AI Meeting Notes, Call Recording and Consent.

What you lose by skipping the "why" questions: a transcript that records the actions and loses the reasons, which means an SOP whose steps a newer tech will drop the first time one seems pointless.

Phase 2: Transcribe, then fix the trade vocabulary

Run the recording through an AI transcription tool and read the result against the audio for the first few minutes. Transcription errors cluster in two places, and both matter:

• Trade terms. Abbreviations and part names are often rendered as ordinary words that sound similar. Search the transcript for every term you know the job uses and fix them.
• Numbers. "Fifteen" and "fifty" sound close, especially over a running pump. Every number in the transcript gets checked against the audio, with its timestamp noted.

Keep the timestamps in the transcript. They are how anyone checks a step later.

Phase 3: Draft the steps with AI, and make it flag its own additions

The prompt has to defend against the model's most predictable behavior, which is filling silence with plausible procedure. A prompt that works: "Turn this transcript into numbered SOP steps. Use only what the tech said. Keep every number, every condition, and every statement with 'always', 'never', 'before' or 'if' in it. Put the timestamp each step came from in brackets. Where the transcript is unclear, write [UNCLEAR] and a question. If you think a step is missing, you may add it, but mark it [ADDED] and say why."

The [ADDED] label is the important part. A model asked to write an SOP will add steps it has seen in other procedures, and some of them will be wrong for this equipment. Forcing it to label them makes them reviewable instead of invisible. Then go through the draft with the tech who was recorded, answering the [UNCLEAR] questions and accepting or striking each [ADDED] step, before anyone tries to work from it. The general reason this matters, that a language model generates what is likely rather than what is known, is covered elsewhere. See related: What a Language Model Is Actually Doing When It Answers.

Phase 4: The verification ride-along

This is where the draft becomes an SOP. A different tech, ideally someone who has done the job a few times but not many, performs a real job from the draft while the author watches. The rules:

• The performer follows the draft literally, even where they would normally do it differently, and says so when the draft surprises them.
• The author does not take over unless something is about to go wrong. They note every point where the performer did what the draft said and the result was not what the author would have done.
• Every correction is logged by type: something the draft omitted, something the draft says that is wrong, something in the wrong order, or something the AI invented.

Using a mid-experience performer matters. The author cannot test the draft, because they fill its gaps from memory without noticing. A brand-new tech is the wrong test too, because they stall on things the SOP should not have to teach.

Worked example: a tankless water heater descaling flush

A plumbing shop records its senior tech doing a scale flush on a gas tankless water heater in a garage, using a submersible pump, a bucket, hoses on the unit's service valves and a descaling solution.

Recording. The narrated job runs 38 minutes. The questioner asks "why" seven times. Two of those answers become steps that would otherwise be missing: photographing the service-valve positions before closing anything ("so I put them back the way the last guy left them, and if something was wrong before, I can prove it"), and setting the bucket on the floor below the unit's outlet ("so it does not siphon back into the heater when the pump stops").

Transcription. The transcript runs about 5,400 words. Reading it against the audio finds two trade terms mangled and one number garbled: the tech said the circulation time from the manufacturer's manual for that model, and the transcript rendered it as a figure a third of that. The corrected transcript notes the timestamp and that the governing figure is the one in the unit's manual, not the tech's memory.

Draft. The AI produces 16 numbered steps, with 4 [UNCLEAR] flags and 2 [ADDED] steps. Reviewing with the senior tech, one [ADDED] step is kept (closing the cold and hot service valves is stated twice in the transcript but was merged out). The other is struck: the model had added a step from a different type of water heater that this unit and this method do not use, and that would have released hot water at the tech's working position. All four [UNCLEAR] flags are answered.

Ride-along. A tech with about a year on the job performs a flush at another customer's home from the reviewed draft. The author logs seven corrections:

The order corrections are the ones that would have hurt somebody. Isolating the water first and leaving the burner live means the unit can try to fire while its water path is being opened up.

The hazard clauses the final SOP carries, because the ride-along exposed them. Shut the gas supply to the unit and remove its electrical power before any service valve is touched, and confirm the unit does not respond to a hot tap opened elsewhere. Let the unit and its lines cool, or treat any water from the purge ports as scalding. Wear chemical splash goggles and gloves rated for the descaling solution, read its safety data sheet before opening it (OSHA's Hazard Communication standard, 29 CFR 1910.1200, requires the supplier to provide one), and never mix an acid descaler with any chlorine-based cleaner, which releases chlorine gas; do the flush with the garage door open. Keep the discharge hose secured in the bucket so it cannot whip out under pump pressure. And the restore step: reopen the valves to the photographed positions, purge air from the hot side at a fixture until flow is steady, and only then restore gas and power, staying clear of the unit until it has fired and run normally.

The result. 16 draft steps, minus the [ADDED] step struck at review and the duplicated rinse removed at the ride-along, plus the 3 omitted steps, makes 17 steps in the final version. The author signs it, the shop files it with its version date, and the ride-along log goes with it.

What would have changed the outcome. If the shop had skipped the ride-along and published the reviewed draft, the two order errors and the missing filter check would have shipped, because the author had already read the draft and approved it. Every one of the seven corrections had survived the author's own read of the draft and only showed up when someone else worked from it, which is the whole case for the ride-along.

How to know the SOP is done

• Every step traces to a timestamp in the transcript or to a logged ride-along correction.
• No [UNCLEAR] or [ADDED] marks remain.
• A second ride-along, if the first one produced more than a handful of corrections, comes back with none that change order or hazard handling.
• The author has signed it, it carries a version date, and the recording and transcript are kept with it for the next revision.

Once it is done, the same source can be turned into a field checklist, carefully. See related: Turning a Trusted Procedure Into a Checklist With AI.

References

• Manufacturer installation and maintenance manual for the specific water heater, for circulation time, solution type and quantity
• Safety data sheet for the descaling product in use (OSHA Hazard Communication, 29 CFR 1910.1200)
• See related: What a Language Model Is Actually Doing When It Answers
• See related: Turning a Trusted Procedure Into a Checklist With AI
• See related: AI Meeting Notes, Call Recording and Consent