Internal cheat sheet · steal this
printed on our office wall ↓

Which Claude do I use?
And how do I make it know my business?

The one-pager we give every client. Four models, one effort dial, three ways to load your business in. Save it, share it, argue with it.

01 · The four models
Claude

Haiku 4.5

"The sprinter"
speed
cost

Near-instant and the cheapest. Use it when the task is mechanical and volume is high.

Claude

Sonnet 5

"The workhorse"
speed
cost

The everyday default. Best smarts-per-dollar for 80% of real work.

Claude

Opus 4.8

"The heavy lifter"
speed (has Fast Mode)
cost

Deep work on big material. Reads entire codebases and contract stacks (1M-token context).

Claude

Fable 5

"The frontier"
speed
cost

The newest flagship family. Hardest reasoning, longest agent runs, best taste.

House rule: start on Sonnet. Drop to Haiku the moment the task feels mechanical. Escalate to Opus or Fable when being wrong is expensive — a lost client, a broken deploy, a mispriced deal.
02 · The effort dial (most people never touch it)

Every model also has a thinking / effort setting: how long it reasons before answering. Low = instant answer. High = it works the problem first. This changes results more than most model swaps.

EffortWhenFeels like
lowLookups, rewrites, formatting, anything you'll check yourselfA fast colleague answering from memory
medium/highDebugging, planning, maths, anything multi-stepSomeone taking a walk to think it through first
maxThe genuinely hard stuff — let it grindA specialist locking the door for an afternoon
Cheat: a mid model on high effort often beats a big model on low effort for puzzles and debugging. But for taste — writing, strategy, judgment — the bigger model wins even on low. Puzzle → effort. Taste → model.

Where it lives:  claude.ai → model picker ▾ → Extended thinking Claude Code → type /model → pick model + effort

03 · Teach it your business (this is the unlock)
1

Make a Project, load your gold

In claude.ai: sidebar → Projects → new. Upload your website copy, your 2–3 best proposals, 5 real customer emails. Then set instructions: "Write like these examples. This is our tone, our offer, our prices." Every chat in the project now knows your business.

Look for:

📁 Projects+ Add to project knowledge⚙ Set custom instructions
2

Connect your real tools

Settings → Connectors. Plug in Gmail, Drive, Calendar (and any MCP tool your systems expose). Now it answers from your actual emails and files: "Summarise every open quote from this month's email."

Look for:

🔌 Connectors Connected: Gmail🔍 Search & tools (in the chat box)
3

Show, don't describe

The 3-example rule: paste two or three gold-standard outputs before you ask for a new one. Models copy patterns far better than they follow adjectives. "Professional but friendly" loses to three real examples, every time.

Here are 3 quotes we actually sent and won. Match their structure, tone and length. Now write one for: [new client details]
04 · When to use what — pin this bit
The taskModelEffortPlus
Label / sort / extract from 1,000 thingsHaiku 4.5lowClear format example
Reply to a customer emailSonnet 5low3 past replies as examples
Write a proposal or quoteOpus 4.8mediumProject with your winners loaded
Debug code / fix a spreadsheet formulaSonnet 5 → Opus 4.8highPaste the error, not a description
Read a 100-page contract / whole codebaseOpus 4.8medium1M context — give it everything
Pricing, strategy, "what should we do?"Fable 5maxReal numbers in, always
An agent that works for hours unattendedFable 5 / Opus 4.8highConnectors + tight instructions
Meeting → summary → action listHaiku 4.5lowYour action-list template
Tech Horizon Labs · we build AI systems for Australian businesses · techhorizonlabs.com
Is AI recommending your business? Free 60-sec check → areyoufoundbyai.com