PTPT: Prompting Techniques for People in Tax

Most prompting guides are written for everyone, which means they are written for no one in particular. When I started using AI seriously in my own tax work, I kept translating generic marketing examples into tax ones before I could tell whether a technique was any good. PTPT is the result of doing that translation properly, once.

Every technique below uses the same worked scenario: a SaaS company wants to know the tax treatment of its business promotion methods, things like free trial periods, reward credits and referral bonuses. Keeping the scenario constant matters. When only the technique changes, you can see exactly what each one buys you.

This is the full set, free, in the order I would teach it: six foundations, eight more that earn their place once the foundations are habits, and an honest list of what I have retired because the models moved on.

Foundations

1. Zero-shot prompting

Ask directly, with no examples. You are relying on what the model already knows. This is the baseline every other technique improves on, and for standard, well-defined questions it is often enough.

As a tax manager, analyse the tax treatment of various business promotion methods for a SaaS company, like free trial period, reward credits, referral to friends etc.

If the answer is good, stop here. If it is vague or generic, the techniques below tell you which lever to pull.

2. Role prompting

Give the model a professional identity. It shapes the perspective, the assumed knowledge and the register of the answer.

You are a senior tax manager with 10 years of SaaS industry experience. Analyse the tax treatment of various business promotion methods for a SaaS company, like free trial period, reward credits, referral to friends etc.

Be specific about seniority and specialism, and say what communication style you want: peer review, client memo, or a mentor walking you through it.

3. Few-shot prompting

Show two to five examples of the input and output pattern you want, then hand over the real question. The model follows the pattern rather than guessing at one.

Analyse tax treatment of SaaS promotions. EXAMPLE: Free software licences given to influencers = marketing expense, deductible when incurred. EXAMPLE: Cash rebates to customers = reduction in revenue, not separate expense. Now analyse: free trial periods, reward credits, and referral bonuses.

Pick examples that cover the edge cases you actually meet, and keep the formatting consistent across all of them. The model imitates whatever you show it, including your sloppiness.

4. Specific instructions and output format

Define exactly what the output should look like: structure, length, required elements. This is the difference between an essay you have to mine and an answer you can drop into a workpaper.

As a tax manager, analyse the tax treatment of various business promotion methods for a SaaS company, like free trial period, reward credits, referral to friends etc. Provide exactly 3 bullet points per promotion method: tax classification, timing of deduction or recognition, and key compliance requirement.

Standardised output has a second benefit: it makes answers comparable across runs, which matters when you are reviewing rather than composing.

5. Contextual prompting

Tell the model about the situation, not just the question. Business stage, upcoming events and reporting priorities all change what a good answer looks like.

Context: This is for a Series B SaaS company with £40m ARR, preparing for audit and considering IPO in 2 years. High focus on revenue recognition accuracy. As a tax manager, analyse the tax treatment of various business promotion methods for a SaaS company, like free trial period, reward credits, referral to friends etc.

The same technical question gets a materially different answer for a pre-IPO company than for a five-person startup, exactly as it would from a human adviser.

6. Chain of thought

Ask for the reasoning, step by step, before the conclusion. Reasoning models now do much of this on their own; the reason to ask anyway is that it puts the workings where you can review them.

As a tax manager, analyse the tax treatment of various business promotion methods for a SaaS company, like free trial period, reward credits, referral to friends etc. Let’s think through this step by step, considering each promotion type’s economic substance and applicable tax principles.

Ask for numbered steps and visible workings. In tax the reasoning is the product; a conclusion you cannot trace is a conclusion you cannot sign off.

Going further

7. Step-back prompting

Ask for the governing principles first, then apply them to the specific problem. It grounds the answer in principle rather than pattern-matching, which is what you want on novel or contested questions.

First, what are the fundamental tax principles that govern customer acquisition costs and promotional activities? [Wait for the response] Now apply these principles to analyse the tax treatment of various business promotion methods for a SaaS company, like free trial period, reward credits, referral to friends etc.

8. Iterative refinement

Start broad, then narrow through follow-ups. This is how working with these tools actually feels day to day: the first answer is raw material, not the deliverable.

As a tax manager, analyse the tax treatment of various business promotion methods for a SaaS company. [After the response] Focus specifically on the revenue recognition timing differences. [After the response] Now add the key HMRC positions on each method.

Keep the intermediate answers. In a review file, the path to the conclusion is worth as much as the conclusion.

9. Planning before execution

Ask for the plan, approve it, then let the work happen. On anything multi-part this catches scope problems before they cost you an hour, and it matters more as tools become more autonomous.

I need to analyse SaaS promotion tax treatment. Before providing analysis, create a research plan covering: 1) Applicable tax law 2) Revenue vs expense classification 3) Timing considerations 4) Documentation requirements. Then execute this plan for free trial periods, reward credits and referral bonuses.

10. Multi-step workflow

Decompose a big task into numbered stages with clear hand-offs. Related to planning, but here you design the stages yourself because you know how the work should flow.

Step 1: Classify each promotion method by economic substance. Step 2: Identify applicable tax treatments. Step 3: Assess compliance requirements. Step 4: Provide strategic recommendations. Now execute this workflow for SaaS promotion methods: free trial period, reward credits, referral to friends etc.

11. Variables and templates

Turn your best prompts into reusable templates with placeholders. This is where prompting stops being personal craft and becomes team infrastructure: a shared library beats twenty people improvising.

You are a {ROLE} at a {COMPANY_TYPE}. Analyse the tax treatment of {PROMOTION_METHODS}. Focus on {TAX_JURISDICTION} implications.

Name the placeholders clearly and test each template on a few different fillings before you circulate it.

12. Constraint versus instruction

Say what you want, not what to avoid. Positive instructions steer better than lists of prohibitions, which models handle unevenly.

As a tax manager, analyse the tax treatment of various business promotion methods for a SaaS company. Focus on providing actionable guidance rather than theoretical discussion. Prioritise practical compliance steps over academic analysis.

13. Comparative analysis framework

Ask for a systematic comparison across named dimensions instead of a narrative. For choosing between treatments, a matrix you can challenge beats prose that flatters everything.

As a tax manager, analyse the tax treatment of various business promotion methods for a SaaS company. Create a comparison matrix showing how free trial periods, reward credits and referral bonuses differ across: tax classification, timing of recognition, documentation needs, and audit risk levels.

14. Competing reasoning paths

This began life in the research literature as Tree of Thoughts, and reasoning models have since internalised the formal version. What survives, and what I still use on any contested position, is explicitly requesting independent competing analyses before any synthesis.

Analyse SaaS promotion tax treatment by exploring multiple reasoning paths independently: Path A) revenue reduction approach, Path B) marketing expense approach, Path C) customer acquisition cost approach. Evaluate each path for free trials, reward credits and referrals, then synthesise the best elements.

If a position matters, you want to see the losing arguments too. This is the prompting version of asking a colleague to argue the other side.

Retired, and why

PTPT originally taught twenty techniques. Six are not above, on purpose: the models moved, and keeping them would be pretending otherwise.

• System prompting. Serious tools now have projects, custom instructions or system prompt slots. Set your standing rules there once instead of repeating them in every prompt.
• Self-consistency. Running the same prompt five times and taking a vote was an accuracy patch for earlier models; reasoning models largely do this internally now. The habit worth keeping: get a second independent check, from another model or a colleague, on anything that matters.
• ReAct. Interleaving reasoning with search actions is built into every tool-using assistant now. You switch research or tools on; you no longer write the scaffold by hand.
• Subagent delegation. Pretending one chat contains three specialists was a workaround. Real delegation now lives in agentic tools that actually spawn subagents; within a chat, role prompting plus decomposition covers it.
• Automatic prompt engineering. “Improve this prompt, then answer it” is now a one-line request rather than a technique, and most tools quietly optimise your prompts anyway.
• Course correction. Interrupting and redirecting mid-answer stopped being a technique when chat interfaces matured. It is just talking now.

One caveat, because this is a tax site: AI output on a tax position is a draft for professional review, never a conclusion, and nothing on this page is tax advice.

Done the patterns? Take the Tax Technology Readiness Scorecard next: eleven questions, three minutes, an honest next step.

Worth reading

A short list, chosen because they hold up for tax work rather than because they are famous.

• Prompt engineering overview (Anthropic). The clearest reference for the fundamentals, and where several of the techniques above are documented properly.
• Prompt Engineering whitepaper (Lee Boonstra, Google). Structure, formatting patterns and pitfalls, in more depth than most teams will need on day one.
• Prompting guide 101 for Gemini in Google Workspace (Google). Practical if your firm lives in Workspace; the examples transfer easily to tax admin work.
• AI in the Enterprise (OpenAI). Useful when the conversation moves from techniques to rollout, ROI and governance, which in tax functions it always does.

The structured version

If you would rather be taught this ground than self-serve it, the CIOT runs two courses I helped build: the Diploma in Tax Technology, the profession’s qualification for exactly this territory, and the shorter AI for Tax course for a faster start. Both are the CIOT’s courses; my part in creating them is voluntary. I would point you at the DITT even if I had never touched it.

Why this page is free

Prompting well is worth learning, and it should be free to learn. This page is the set I would teach any tax team first. But technique gets you a better answer; it does not give you permission to rely on one. The questions that decide whether AI helps or hurts a tax function are not prompting questions: which work AI should never touch, what review an AI-drafted position needs before it goes near a return or an audit response, how client and company data is handled, and how any of it stays consistent across a team rather than living with one enthusiast. That is judgement, governance and workflow, and it is where I spend my time now. A prompt is advice; a skill is infrastructure. If that is the conversation you need, start on the Work with me page.