prompt engineering for tax research — what actually works vs what sounds clever
i've been refining my prompting approach for tax research queries over the past few months and wanted to share what i've learned (mostly through painful trial and error).
what DOESN'T work:
- "what's the tax treatment of X?" → too vague, gets a generic textbook answer
- "as an expert UK tax advisor, tell me..." → role-prompting doesn't improve accuracy on technical questions, just makes it sound more confident
- long chain-of-thought prompts with 10 steps → model loses the thread by step 6
what DOES work:
- "citing specific legislation and HMRC guidance, what is the VAT treatment of [specific scenario] for a UK-registered business?" → forces citations
- breaking complex questions into 3 separate prompts instead of one mega-prompt
- "what are the common errors practitioners make when applying [specific rule]?" → surprisingly useful for self-checking
- feeding it the actual HMRC manual reference first, THEN asking the question about it
biggest realisation: the prompt matters way less than whether you have a skill/context document loaded. a mediocre prompt + a good skill file beats a perfect prompt with no context every single time.
what's everyone else finding?
4 replies
the "common errors" prompt is gold. i use a variant: "what assumptions in this computation would a reviewer most likely challenge?" — basically asking the agent to pre-empt the review.
agreed on skills > prompts. i spent weeks perfecting prompts before discovering that just loading the right skill file made most of my careful prompting unnecessary.
hot take: prompt engineering for tax is a temporary skill. as skills files get more comprehensive, the prompts become simpler. "apply this skill to these facts" is the endgame prompt. we're just not there yet for every jurisdiction.
for German tax research i've found that prompting in German gets measurably better results than English, even with the same model. makes sense — the training data for Steuerrecht is predominantly German.
also: asking for the Fundstelle (legal citation) in the prompt is essential. if the model can't cite a specific paragraph of EStG or UStG, i don't trust the answer.
same experience for French tax. prompting in French with specific reference to the CGI article number gets much better results than an English prompt.
one technique that works well: i ask the agent to produce a "note de synthèse" (summary memo) format — forces structured output with citations, applicable articles, and a conclusion. much easier to review than freeform text.
Sign in as a verified accountant to reply.
Sign in