Most accountants who have used AI for more than a fortnight have arrived at the same conclusion. The model is only as good as the prompt, and a sloppy prompt produces a sloppy output that is then unsafe to use. The fix is not more powerful AI; it is a small library of carefully written prompts that the team uses repeatedly, refines over time, and treats as part of the firm's working methodology.
This piece sets out fifteen prompts that earn their place in a UK practice in 2026. They are written for use against a docs-only or business-grade AI tool - not a consumer chatbot with client data pasted into it - and they assume the documents in question have already been uploaded to the tool. Each prompt is paired with a short note on how to review the output before it lands in a working paper file.
Why a prompt library beats freestyle prompting
A freestyle prompt is the team member improvising at the keyboard, getting a passable answer, copying it into a file note, and moving on. The output is unreviewable because nobody else knows what was asked. The next person asking the same question gets a different answer because they phrase it differently. The firm cannot improve because there is no shared artefact to improve.
A library fixes all three. The prompt is written down, reviewed, and standardised. The output is consistent across the team. When something goes wrong, the prompt is what gets corrected, not the individual. Over a year a firm with twelve people will run thousands of AI interactions; the prompts that drive them deserve to be a managed asset.
Prompt 1 to 5 - document review
The first cluster of prompts handles document review across contracts, leases, accounts, and correspondence. The discipline in each is that the prompt forces the model to cite sources, refuse on absent information, and surface the things the reviewer most needs to see:
- 1. "Summarise this engagement letter in 200 words for a partner review. List the scope, the fee basis, the term, the termination provisions, and any unusual clauses. Cite the clause number for each point. Do not infer anything not present."
- 2. "Extract from this lease the following data points into a table: landlord, tenant, premises, term start, term end, break dates, rent, rent review dates, service charge basis, repair obligation, alienation provisions. Cite the clause number. Leave blank and flag if absent."
- 3. "Compare these two sets of statutory accounts. List every line item that has moved by more than 10% year-on-year, with the prior year and current year figures and the variance. Do not interpret the variance."
- 4. "Read this client email chain and produce a chronological note of the points raised, the questions asked, the answers given, and the open items. Quote the relevant sentences. Do not summarise away anything that looks like an instruction."
- 5. "Read this shareholders' agreement and list every clause that gives a shareholder a right of veto, pre-emption, drag-along, or tag-along, with the clause number and a one-line plain-English explanation."
Prompt 6 to 9 - drafting
The next cluster covers drafting. The output is always a first draft, never the final version, and the prompt makes that explicit. The reviewer rewrites for tone and accuracy before the document goes to the client:
- 6. "Draft a client email confirming that we have received the documents listed below, setting expectations that the first draft of the accounts will be ready in three weeks, and asking for confirmation of the directors' loan position as at the year end. Keep it under 150 words and use a neutral professional tone."
- 7. "Draft a one-page file note recording the partner's review of the AML risk assessment for [client], referencing the risk factors in the assessment, the screening result, and the partner's conclusion. Mark places where the partner needs to fill in their judgement with [TO COMPLETE]."
- 8. "Draft the management commentary section for these accounts using only the variances shown in the prior year comparison. Do not introduce new explanations. Mark each variance over 10% as [REQUIRES EXPLANATION]."
- 9. "Draft a checklist of every working paper we would expect to see on file for a small-company audit under ISA (UK) for [client] in [sector]. Group by FRS 102 area. Do not invent client-specific facts."
Prompt 10 to 12 - tax research framing
These prompts deliberately stop short of asking the AI for a tax answer. They use the AI to frame the question, identify the points of law to look up, and produce the structure of the research note that a qualified tax person will then complete using primary sources:
- 10. "I need to research the corporation tax treatment of [scenario]. List the specific sections of CTA 2009 and CTA 2010 I should read, the HMRC manual references most likely to be relevant, and any leading cases that are commonly cited on this point. Do not state the answer."
- 11. "Produce the structure of a tax research note on [scenario] with sections for: facts, issues, applicable law, analysis, conclusion, and risks. Leave each section blank for me to complete. Add a footer that says 'This note must be reviewed by a qualified tax advisor before reliance.'"
- 12. "List the questions I should ask the client about [scenario] before I can give a complete tax answer. Group them by relevance to entity, transaction, timing, and counterparties."
Prompt 13 to 15 - internal operations
The final cluster is for internal operations - the meeting notes, the team handovers, the management reporting that consumes hours of senior time and is well suited to AI assistance:
- 13. "Read the transcript of this client meeting and produce a structured note with: attendees, agreed actions with owners and dates, decisions made, open questions for follow-up. Quote the relevant lines. Leave anything ambiguous as [TO CLARIFY]."
- 14. "Read the last 30 days of activity for this client and produce a one-paragraph status summary suitable for a partner who has not looked at the file recently. Highlight anything that needs partner attention this week."
- 15. "Convert this rough handover note into a structured handover document with sections for: client background, current engagement status, recent issues, contacts, and outstanding tasks with owners. Do not introduce new information."
How to review the output before it enters the file
Each of the fifteen prompts is designed so the output is reviewable in two to three minutes rather than fifteen. The standard discipline is: check the citations exist and say what the model claims they say, check that the model has not filled in anything that should have been marked as missing, rewrite for tone where the output is going to a client, and only then save it against the client record.
A working paper that has been produced with AI assistance should record the prompt used, the documents available to the AI, the reviewer's checks, and the firm's final position. That last item is the partner's view, not the AI's view. The AI is a draftsperson, never the signatory.
Maintaining the library
A prompt library is a living document. The firm reviews it quarterly, retires prompts that have stopped working, adds new ones for use cases that have appeared, and revises the wording of any prompt where the reviewer found themselves regularly rewriting the output. Ownership should sit with a named individual - typically the operations lead or the senior manager responsible for technology - rather than diffused across the team.
A short companion document explaining which prompt to use for which task, and which tool the prompt is approved for, prevents the library from becoming theoretical. The team should be able to find the right prompt for the situation in under a minute.
What not to include in the library
Equally important is what does not belong. Any prompt that requires the team to paste client-identifying data into a consumer chatbot account should not be in the library. Any prompt that asks for a definitive tax or audit opinion to be produced by the model should not be in the library. Any prompt whose output goes straight to a client without partner review should not be in the library. The library is a productivity tool, not a substitute for professional judgement or regulatory discipline.
How Accupe helps
Accupe's AI document analysis is designed for exactly this kind of library use. The three modes - Fast for quick first-pass review, Planning for structured drafting, Ultra-Detailed for deep document interrogation - and a docs-only mode that refuses to answer beyond the uploaded files give the team a consistent surface to run a prompt library against. Each output cites its source passages, sits inside the client record, and feeds straight into the working paper file. Per-firm pricing from £20/month.