← AI 101

Intermediate training

Choosing AI

When Claude is the right tool, when it isn't, and when something else — a spreadsheet, a colleague, the LOS, the rate sheet — will get you a better answer faster. The judgment that closes Level 2.

17 minutes Builds on Module 2.3 Includes tool-fit tester

Start here

What you'll be able to do after this lesson

Classify a task quickly

Sort any task you're handed into Claude-strong, Claude-weak, or Claude-wrong in about thirty seconds, using a reusable mental algorithm.

Name the better tool

When Claude is the wrong answer, know what the right answer is — the LOS, a spreadsheet, the rate sheet, a colleague, the policy doc.

Avoid common misuse

Recognize the three or four patterns at GMFS where employees keep reaching for Claude when something else would land faster and cleaner.

Three tests

The decision algorithm in one paragraph

The language test

Is the work mostly text — drafting, summarizing, explaining, comparing, structuring? If language is the bottleneck, Claude is in the running. If language is incidental and the real work is something else, the right tool probably is too.

The accuracy test

Does the answer have to be exactly right, in a way you can't verify yourself in a minute? Math, regulated decisions, live data. Claude is fluent, not always correct — so anything that has to be exact needs the tool that's actually exact.

The data test

Does the task need live or proprietary data Claude doesn't have? Today's rate sheet, your LOS pipeline, your borrower file. Without the data, Claude is reasoning about an empty room.

Trainer note: Claude is a strong generalist — not a calculator, not a database, not a regulator, not a colleague. It's better than every other tool at the things it's better at, and worse at everything else. The skill is knowing the difference before you start.

Wrong-tool surfaces

Three places employees most often reach for the wrong tool

Math & finance

DTI, LTV, exact rate math, amortization, fee comparisons. Use a spreadsheet or the LOS. Claude can format and explain the output once the math is done elsewhere.

Use a sheet

Live data

Current rates, today's pipeline, real-time inventory. Use the source system. Claude has no live feed into your tools and no current knowledge of anything moving today.

Use the source

High-stakes compliance

Exceptions, adverse-action reasoning, fair-lending decisions. Use the human approver with documented authority. Claude can prepare the file; the human decides.

Use the human

Tool choice

Five rules for picking the right one

Start with "what would good look like"

Describe the deliverable in one sentence. The right tool usually becomes obvious in that sentence — or it becomes obvious that the task is ambiguous, which is its own answer.

Ask whether language is the bottleneck

If the slow part is reading, writing, summarizing, or explaining — Claude is in the running. If the slow part is data entry, math, or a system click, look elsewhere.

Check whether the answer needs to be exact

Exact arithmetic, exact citation, exact regulatory wording — those need the exact tool. Approximate language work is where Claude shines.

Check whether the inputs are reachable

Does Claude actually have access to the source data, file, or document? If you can't attach or paste the input, the task isn't ready for Claude — find the data first.

Ask a teammate when it isn't obvious

Thirty seconds of someone else's judgment beats twenty minutes of trying. "Is this a Claude task?" is a healthy question to ask out loud.

Prompt upgrade

Wrong-tool prompt

Build me this week's rate sheet for 30-year conventional, FHA, VA, and USDA at FICO tiers 660 through 800.

Right-tool prompt

Work-ready prompt

Attached is this week's rate sheet (CSV exported from the rate engine). Reformat it as a clean table for client-facing use, group by program, and write a one-paragraph plain-English explanation for FICO 720 borrowers. Do not compute or interpolate any rates — use exactly what's in the file.

When Claude is the right tool

Four patterns where it clearly wins

Draft-then-edit

Anywhere the blank page is the cost. Emails, memos, status updates, talking points. Claude produces the draft; you bring the judgment.

Summarize-then-decide

Long documents you have to make a call on. Attach the source, ask the specific question, get the version of the document you actually need.

Extract-and-structure

Messy text in, structured output out. Action items from a thread, entities from a document, a table from an unstructured paragraph.

Explain-then-teach

Translating complex material for a specific audience. Underwriting reasoning for an LO, a guideline change for a realtor, a policy for a new hire.

Wrong-tool red flags

Five signs you reached for the wrong thing

The task needs exact arithmetic — a spreadsheet does this in one cell The task needs live or current data Claude has no access to The task is a regulated decision a human has to own The task is fundamentally about a human relationship — a call, not a draft The task could be done faster with one keyboard shortcut or a saved query

Employee rule: The point of this curriculum isn't to use Claude on everything. It's to use Claude on the things it makes better, and to recognize — quickly — the things it doesn't. The wrong-tool reach is the most common mistake on this list.

Sort your week

Six exercises to calibrate your judgment

The goal isn't to memorize categories. It's to develop a fast reflex for "which tool" — and to feel the cost of getting it wrong before it costs anything real.

Open the tool-fit tester in this lesson. Click each of the five sample tasks and read the three-test scorecard. Pay attention to why each test passes or fails.
Classify ten recent tasks from your last week into Claude-strong, Claude-weak, or Claude-wrong. Be honest — half the value is noticing which ones you got wrong.
List three tools you reach for every day that Claude cannot replace. Write down what each does that Claude can't.
Find one task you've been over-using Claude on. Pick a new tool for it next week and time the difference.
Find one task you've been under-using Claude on. Pick a real one this week and run it through a proper brief.
Share both findings with a teammate. The fastest way to build this judgment as a team is to compare what each person classified differently.

Completion standard

You've finished this module when you can run the three tests in your head in under thirty seconds for any task — and when "is this a Claude task?" is a question you ask before opening the chat, not after.