ImpactMe.AI · NECA Accounting Workshop · October 27, 2026

Thinking Like AI: Cheat Sheet

Prompts are the fish. Thinking like AI is the fishing. Specific prompts go stale in about a month. This page is the part that lasts.

What an LLM actually is (plain language)

ChatGPT, Claude, Gemini, and Copilot are all built on large language models (LLMs): software trained on enormous amounts of text to predict what words should come next. When you ask a question, it doesn't look up a stored answer. It writes a response, word by word, based on patterns it learned.

That's why it's excellent at language (drafting, explaining, summarizing, rewording) and why it can be shaky at things that need exact precision (arithmetic across many rows, specific dates, citing a real rule). It isn't a calculator or a database. It's closer to a very well-read assistant who works fast, never gets tired, and sometimes confidently misremembers.

The big shift: Excel is deterministic: same inputs, same answer, every time. AI is probabilistic: same question, slightly different answer each time. Knowing that difference is half the skill.

The sorting framework: where does this task belong?

Key idea: An AI being capable of a task does not mean you have a working, trustworthy process. Before you use AI for something, sort it.

	What it means	Accounting examples
Good at	Language and thinking work where "pretty good" is useful and you can judge quality at a glance	Drafting a retainage follow-up email. Turning rough notes into an SOP. Explaining WIP to a new PM. Brainstorming causes of a variance. Writing an Excel formula from a description. Suggesting questions for a job review.
Needs a human check	AI does the first pass; a knowledgeable person reviews before it's used	Variance commentary for the owner. A journal entry memo. A cleaned-up data export. A pay app review checklist. Spotting likely matches in a reconciliation.
Never alone	Anything with a hard, checkable answer or real consequences. Humans decide; deterministic tools (Excel, your accounting system, a calculator) do the math	Final totals and tie-outs. Retainage and billing amounts. Lien and notice deadlines. Tax and compliance positions. Payroll and union fringe calculations. Anything that gets signed, filed, or paid.

Rule of thumb: If there's one right answer you can check, check it with a tool that gives the same answer twice.

The 5-part prompt pattern

Role: "You are a construction accountant..."
Context: Who you are, what the data is, what you already know.
Task and format: Exactly what you want back, and its shape (table, bullets, email, word count).
Constraints: What to avoid, what not to change, length, tone.
Check yourself: "List your assumptions. Flag anything you're unsure of. Show your math step by step."

You won't need all five every time. But when the answer disappoints, a missing part is usually why.

How to iterate: the first answer is a draft

Push back. "That's too formal." "You missed the change orders." "Use our numbers, not examples."
Add what it didn't know. Most weak answers come from missing context, not a weak tool.
Ask it to critique itself. "Act as a skeptical reviewer of your last answer. What's most likely wrong?"
Ask for options. "Give me three versions: gentle, neutral, direct."
Start fresh when it's stuck. A new chat with a better first prompt often beats ten corrections.

Why answers vary (and why that matters)

Ask the same question twice and you'll often get two different answers. That's by design: the tool picks among likely next words with some built-in variety, which helps with writing but means it's not repeatable like a formula.

Two traps to avoid:

"It gave different answers, so it's broken." Not necessarily. Variation is normal for writing tasks. For tasks with one right answer, variation is your signal to verify with Excel.
"It gave the same answer twice, so it's right." No. Repeatable does not mean correct. A tool can be consistently wrong.

Confidence, fluency, and correctness are three different things

Fluency: It reads well. AI is almost always fluent.
Confidence: It sounds sure. AI often sounds sure even when it's guessing.
Correctness: It's actually right. Only checking tells you this.

A polished, confident paragraph is not evidence of a correct one. In accounting, the most dangerous output is the one that looks finished.

The verify habit

Make these automatic for anything with numbers, dates, thresholds, or compliance:

Tie out: Does the total match the source report, the GL, or the bank?
Recompute one row: Pick one line and do the math yourself. If one row is wrong, assume others are.
Check against source: Every number in a draft should trace back to a document you trust.
Count rows: After any cleanup, same number of rows and same grand total?
Confirm rules at the source: Contract terms, lien deadlines, and tax rules come from the contract, the statute, or your advisor, not the AI.

The reviewer needs to know enough to catch the mistake. If you couldn't spot an error in the output, you're not ready to rely on the tool for that task yet. That's fine. Use it to learn first.

Adam Murphy · ImpactMe.AI · AI Made Useful & Human · adam@impactme.ai · impactme.ai