The method: your documents, your clauses, your week

Two weeks before a cohort we send a short questionnaire. We ask which tools are already licensed, whether an enterprise tenancy exists, and how documents actually arrive on the desk — scanned PDFs with handwritten amendments are the hard case, not tidy exports from a single system. We ask for the three heaviest recurring tasks, not a wish-list of everything a model might one day touch. Flag-state and class constraints on data handling are recorded at this stage, because they decide whether work stays inside an on-premise workspace or can use a named enterprise tenancy.

The answers shape the exercise pack and the format. A desk that cannot release samples is taught from our sanitised corpus; a desk that can supply redacted recaps and SOFs works on its own paper. Nothing in the questionnaire is used to profile the company for advertising. It exists so that Monday morning’s session matches the week the desk actually runs, including the ugly files, not a demonstration document that never appears in post-fixture or technical correspondence.

When a client cannot release samples, we teach from an in-house synthetic corpus of 48 exercises. The pack is built from publicly available form documents — recap layouts, NOR wording, SOF grids, noon-report sheets, bunker delivery notes, PSC deficiency lists — and from invented voyages with invented counterparties. Rates, names and vessel identities are fictitious. The point is not to impersonate a particular fleet; it is to give attendees a file that behaves like the one they will open on Monday, with awkward columns, clause cross-references, and the same places a model will invent a time.

Clients who can supply redacted samples replace the matching exercises so the cohort works on its own paper. Redaction happens before any file is shown in the room: counterparty names, freight and hire figures, crew personal data, and incident detail that may sit under legal privilege are stripped or replaced. The trainer works only from those client-supplied redacted samples or from the synthetic corpus. Public chat tools never see a live commercial document. That rule is part of the method, not an optional extra for cautious offices.

A session is not a lecture with a demonstration at the end. We move through five stages so that people leave able to run the workflow without the trainer at their shoulder. Everyone works at a laptop. Outputs are read aloud and challenged against the source. Paper and screen stay together, because the source of truth is still the document, not the chat window. The sequence is the same in a half-day clinic and a two-day cohort; only the number of workflows changes.

1. 1

   Model-behaviour briefing

   We show how current tools behave on maritime paperwork: fluent paragraphs that hide invented shifting times, weak arithmetic on laytime totals, OCR drift on scanned SOFs, and the gap between a confident summary and a checkable extract. Numbers and dates are treated as hostile until proved. This hour is meant to make over-trust uncomfortable before anyone pastes a real clause.

2. 2

   Pattern library

   Attendees receive prompt patterns tied to the document types on their desk — recaps, SOFs, noon reports, bunker notes, deficiency lists, circulars — including the standing instruction to demand clause citations, a dated source list, and a separate working for any total. Patterns are written in the desk’s vocabulary, not in generic “summarise this email” language.

3. 3

   Guided run

   The trainer and the desk work through one redacted document together. We pause wherever the model invents a time, misreads a column, drops a reservation of rights, or tidies away an ambiguity that the charterparty still needs. The guided run is slow on purpose. The aim is to make the failure visible while there is still someone in the room who can name the clause.

4. 4

   Unassisted run

   Each person runs the same class of task on a fresh document at their own laptop, without the trainer prompting the next step. The pattern has to stand on its own. People are expected to refuse an output that cannot be traced, and to re-run rather than to polish a fluent error. This is where a cohort finds out whether the morning’s notes were understood.

5. 5

   Verification drill and debrief

   Every output must be traced to a clause, a log entry or a reading. Totals are re-derived by hand once. The debrief records what the model got wrong and which check caught it, so the desk leaves with a short routine rather than a pile of screenshots. Nothing leaves the room as correspondence until a named human has signed it.

The teaching sequence does not change with the calendar. What changes is how much of a programme fits in the room, how many seats we will take, and whether the work is a single clinic or a six-week embedding while a licensed tool is rolled out. Scoping decides the format; the format does not decide the industry.

Redaction is not a courtesy; it is the condition of the work. Before any document is used in a session we remove counterparty names, freight, hire and bunker rates, crew personal data, and incident specifics that may sit under legal privilege. Those fields are never pasted into a consumer chatbot. A public tool and an enterprise tenancy are not interchangeable: the first is treated as an external party with no place in live commercial files; the second is used only where the company already holds that tenancy and has set retention. On-premise options are recorded at scoping when flag or class rules require them.

Retention settings are the client’s, not ours. We do not keep client documents after the cohort unless the company asks in writing for a redacted exercise to remain in its own playbook. Course materials describe general practice. They do not replace charterparty terms, class rules, flag requirements or the company’s SMS. Anything that leaves the office as correspondence, a claim narrative, an observation response or a circular still needs a human signature. The course enforces that rule in the room: if an output cannot be traced to a clause, a log or a reading, it is not finished work, however fluent it looks.

A two-day cohort does not end at the door. We offer a six-week reinforcement window with two office-hour sessions, a written playbook for the desk, and a short review of what stopped being used and why. The review is more useful than another demonstration. Tools change; desks drop patterns that were awkward, and they keep the ones that survived a real SOF or a real verifier query. We recut materials when a model’s behaviour changes enough to break an exercise, and we say so rather than pretending last quarter’s screenshots still apply.

Office hours are for the cases that did not fit the room: a scanned SOF with column drift, a recap that mixes two forms, a noon-report set that will not reconcile with the BDN. They are not a claims service and not an audit. If the desk wants another workflow added, that is a scoping conversation, not an extra hour tacked onto a clinic. Start with the call if you do not yet know whether a half-day or a two-day seat is the right size.