Using AI as a custom consultant to find tax savings

Executive overview

Most people ask AI generic questions and get generic answers. The fix is a structured process that extracts your specific context first, then builds a personalised consultant from it.

Run a reverse interview — AI asks you questions one at a time until it fully understands your situation. That conversation gets synthesised into a context document, which feeds a custom system prompt, which powers a project in Claude or ChatGPT.

The core insight: AI gives better advice when it knows your situation deeply — so build the context before you ask the questions.

The reverse interview

• Prompt a fast reasoning model (e.g. o4-mini) to interview you one question at a time
• Each answer informs the next question — 15–20 turns total
• Provide initial context upfront: business type, location, incorporation docs, prior tax filings, industry, offering
• End the session by asking the model to synthesise all findings into a single document
• Stop early if you feel covered — just ask for the synthesis at any point

Validating the synthesis

• Pass the full conversation and the synthesised doc to a larger model (Sonnet 4, Opus 4, or o3)
• Ask it to compare both, identify gaps, and rewrite a more complete context document
• Run through o3 as well if the task is critical — multiple passes reduce missed details

Building the system prompt

• Feed the context document to Claude or ChatGPT with a prompt requesting a system prompt for a specialist AI
• Specify today's date and the exact model being prompted — each model has different prompting best practices
• Add constraints: always research up-to-date tactics before advising; always reference the provided context
• Force the AI to be opinionated and practical — vague, agreeable responses are not useful
• Make minor manual edits if needed before deploying

Setting up the project

• Create a project in Claude or ChatGPT (not a custom GPT — projects are private)
• Paste the system prompt into the instructions field
• Upload the synthesised context document and any supporting docs (tax filings, incorporation papers) to the knowledge base
• Ask the AI what additional documents it would need — use that list to fill gaps

Querying the consultant

• Start broad: ask for general tax-saving strategies relevant to your context as of today's date
• Narrow down into the responses that are most relevant or interesting
• Validate AI-generated advice independently before acting on it
• Pass the final recommendation to a human advisor — especially for legal or tax matters