Things we wrote while teaching.

How we use this library

We write these materials because a session without a page to take home tends to fade. The sheets are short on purpose. Each one is meant to sit beside a real task, not to replace judgement. We revise them when a cohort shows us a gap: a check that people skip, a pile in the task audit that needs a clearer label, a confidentiality rule that was too vague to act on.

Nothing here is a product catalogue. We do not sell templates as a line of business. If your team is already in a programme with us, you will receive adapted versions written against your own documents. If you are still deciding, the four sheets below are a fair sample of how we think. Ask for them on the contact form and we will send the current versions as plain files.

The notes underneath are dated in 2026, the year we formed as a collective. They are teaching notes, not news. We keep them public so a sponsor can see the kind of argument we make in the room before booking time with us.

The task audit sheet

Before we teach a single instruction, we ask the group to list ten to twenty recurring pieces of work. The sheet sorts those items into three piles: safe to hand over with a review step, worth assisting but not delegating, and not suitable for a general tool at all. The third pile is the one teams try to skip. We keep it on the page because confidential records, one-off judgements and anything that needs a named person to stand behind it should stay in human hands.

The sheet is one side of A4 plus a worked example from a support queue. People fill it in during the first session and we return to it in week six. If you request it, we send the blank version and a short note on how we facilitate the sort.

Output review checklist

Most teams we meet already generate drafts with a model. The gap is the last five minutes before something leaves the building. This checklist is nine checks: source match, numbers, names, quotations, tone against the house voice, confidentiality, invented citations, date freshness, and whether a person who knows the file would still sign it. Each check has a one-line prompt so a reviewer does not have to invent a method under time pressure.

We print it as a card and also keep a version that lives at the top of a shared document. Participants mark which checks apply to their function. Support teams lean on tone and escalation; finance teams lean on numbers and source match. The sheet is useless if it stays generic, so we ask people to strike out lines that do not apply rather than add more.

Context blocks

A model that has not been given the right background will invent a plausible one. This sheet shows how to write a short context block: who the reader is, what document or policy the answer must stay inside, what the output must not do, and which facts are already decided. It is deliberately not a prompt library. The point is to stop people pasting an entire file into a general tool when a paragraph of constraints would have been enough.

We practise this with the team's own material. A bid writer learns to name the client, the lot, and the three claims that must not drift. A people-operations colleague learns to point at the handbook section rather than restating policy from memory. The sheet includes a worked block and a blank frame. We send both when you ask.

Confidential-by-default rules

This is the shortest sheet and the one we refuse to skip. It lists categories of material that should never be pasted into a general assistant: personal data, unpublished financials, live customer records, draft legal advice, passwords, and anything marked internal for a reason your policy already names. It also covers the grey cases teams actually argue about, such as a redacted extract or a question that only makes sense if you attach the source file.

We write it in plain language so it can sit next to an existing information-security note. We do not replace your policy. We help people apply it at the moment they are about to paste. If your organisation already has a usage rule, we annotate this sheet against that rule during the programme rather than handing out a second, competing document.

Notes from the room

These six notes were written in 2026 after sessions, not before them. Each one records something we had to explain more than once, or a change we made to the six-week cohort because the first runs showed a weak spot. They are short on purpose. If a note helps a sponsor decide whether our method fits, it has done its job.

A short glossary in our own words

We use these terms in sessions because participants meet them in product screens and vendor pages. The definitions below are the ones we actually say in the room. They are not a standard, and they are not complete. They are enough to stop a conversation collapsing into borrowed jargon.

Grounding

Grounding means tying an answer to material the team already trusts: a policy, a spreadsheet, a prior filing. If you cannot point to the source, the answer is not grounded, however tidy it looks.

Hallucination

We use this word for invented facts, citations, names and numbers that the model presents as if they were found. It is a failure of truthfulness, not a creative flourish, and it is why review is written into every routine we leave behind.

Context window

The context window is how much text the tool can consider at once. Pasting a whole archive is rarely useful and often unsafe. A short, accurate briefing usually beats a dump of files the model cannot actually hold.

System instruction

A system instruction is the standing brief you give a tool before the day's task: role, limits, tone, what to refuse. We treat it as a written agreement, not a spell. If the team cannot explain it to a new colleague, it is too clever.

Retrieval

Retrieval is the step where a tool searches a set of documents you have allowed and then drafts from what it found. It still requires a person to check that the passage it used is the right one and that nothing important was missed.

Review loop

A review loop is the named moment when a person checks the output against source material before it is sent. Without that moment in the calendar, fluency training becomes a demonstration that nobody repeats.

Task audit

A task audit is the structured sort of recurring work into what may be assisted, what may be drafted with review, and what should stay human. We do it on paper first so the conversation is about the work, not about the interface.

Handover pack

The handover pack is the set of documents a team keeps after we leave: adapted prompts, a review checklist, the written routine, and the list of tasks we recommend leaving alone. If those pages are not maintained, the programme has not finished.

Sending the sheets

There is no download button on this page. We send the four practice sheets when you ask, using the contact form, so we know a person rather than a scraper wanted them. Include your organisation if you wish, and name the sheets. We do not require a programme booking. We do not add you to a list. Enquiry records are kept as described in the privacy notice.

If you already know the format you want — a half-day clinic, the six-week cohort, an embedded sprint or a leadership briefing — say so in the same message. We will reply with the files and, if it is useful, a time for a thirty-minute scoping call. Public open cohorts are not on our calendar; we teach intact teams.