INSIDER

The AI Model Playbook

INSIDER The AI Platform for the Diamond and Jewelry Industry
The companion guide

The AI Model Playbook

You saw the tree. Here is the depth behind it. Four models, what each one is really for, and the exact prompts to run from your own desk, in the language we actually use.

Most of us reach for the same model every time and wonder why the output feels flat. The model is a tool. Picking the right one for the task is the first professional habit worth building.

01

Haiku 4.5

Lightweight · cost-saver

Reach for it when speed and volume matter more than depth. Haiku is fast and cheap. It is built for the small, repeatable jobs that fill a trading day: the reformatting and tidying that eats an hour before the real work starts.

Good to know  It is the lowest-cost model in the lineup, $1 and $5 per million tokens. Its context window is the smallest of the four, which is plenty for single messages and short listings. Tip: paste a whole supplier list in one request rather than a few rows at a time, and you cut both the time and the cost.
Best for
  • Reformatting supplier lists into one layout
  • Translating a client message across languages
  • Fast unit and currency conversions
  • Tidying a rough note into a readable sentence
Not for
  • Strategy, pricing, or negotiation prep
  • Anything where a wrong nuance costs you
  • Long documents that need real reasoning
“Clean this supplier list into a table with these columns: [stone ID, shape, weight, color, clarity]. One row per stone. Keep every value exactly as written. Do not correct, round, or guess anything. Where a field is missing, write MISSING.”
02

Sonnet 5

The everyday workhorse

Your default for real work that does not need the heavy model. Sonnet handles the bulk of a working week: drafting, replying, structuring. It is also the model to connect to your tools, so it can read your own files instead of guessing.

Good to know  It runs well below Opus cost and handles long documents comfortably. Tip: for most of your week this is the right model. Only move up to Opus when a task genuinely needs deeper reasoning, not just because it is available.
Best for
  • Memos, client replies, follow-up emails
  • First drafts of a collection or campaign story
  • Spec sheets and product descriptions at volume
  • Cleaning and reformatting exported data
Then connect it
  • WhatsApp chat exports
  • Google Drive and WordPress
  • Your CRM and inventory sheets
  • 50+ more, so it works from your real data
Not for
Bulk reformatting of long lists, where Haiku does the same job for less. Also not for a decision with real money behind it, where Opus earns its cost.
“You are a [jewelry brand copywriter]. Write a product description for this piece: [paste the spec sheet]. Tone: understated, no superlatives. Do not invent gemological data. Leave any unknown as [brackets].”
03

Opus 5

The deep-work model · near-frontier

The one you bring into the room for decisions that carry weight. Opus 5 moved up a level. It now comes close to the top model on most work, at a fraction of the cost, which makes it the right default for almost every serious task. Do not ask it to draft immediately. Ask it to question you first, then build with you.

Good to know  Opus 5 reaches near-frontier intelligence at exactly half the price of Fable, $5 and $25 per million tokens against Fable’s $10 and $50. It holds several long documents in a single pass. Tip: drop several tenders in together and ask what repeats across them, rather than reading each one alone.
Best for
  • Pricing strategy for a line or a stone
  • Comparing results across several tenders
  • Planning a sourcing or inventory decision
  • Structuring a full marketing campaign
How to work it
  • Give it the goal and your constraints
  • Tell it not to draft until it has asked you
  • Refine in steps, not in one shot
  • Keep effort on high, always
Not for
Quick one-line jobs and routine drafting. Running those here is paying for depth you are not using. Send them to Sonnet, or to Haiku if it is volume.
“You advise a diamantaire. My goal: [decide what to do with slow-moving stock]. My constraints: [margin, stock, timeline]. Do not draft yet. Ask me clarifying questions first, then we refine the approach step by step.”
“Here are results from [four tenders]. Compare them: where my bids landed against the winning levels, by [category and size range], and what repeats across all four. Use only the figures I gave you. Do not estimate or fill gaps, mark anything missing as [missing]. Then tell me what you would watch in the next tender.”
04

Fable 5

The most capable model

Save it for the small share of work that truly earns it. Fable is still the most capable model, and it costs more. Now that Opus 5 handles most deep work, that share is smaller than it was. Reach for Fable on the hardest end-to-end tasks, or when Opus stalls. Let it run a full workflow rather than a single step.

Good to know  It is the most capable model Anthropic sells, at exactly twice the cost of Opus 5. It also carries strict safety limits, so a small share of sensitive requests are handled by Opus instead. Tip: keep Fable for the one genuinely hard problem a week that pays for the extra cost, and run everything else on Opus.
Best for
  • Researching a new market end to end
  • Modeling a full product launch
  • Reasoning through a genuinely hard trade-off
  • Work that spans many connected steps
Keep in mind
  • It costs more tokens, so use it with intent
  • Opus 5 now covers most of what you need
  • Escalate to Fable, do not start there
Not for
Anything Opus already handles well, which is most deep work. Reach for Fable only when Opus stalls, otherwise you are paying double for the same answer.
“Here is my goal: [enter the Asian retail market with a fancy color line]. Here are my constraints: [budget, timeline, existing relationships]. Think through the trade-offs, propose 2 to 3 approaches, and recommend one with your full reasoning.”

Five habits that hold across every model

  1. Never let it invent specs. No carat, color, clarity, or certificate it was not given. Unknowns stay in [brackets].
  2. Give it a role and a constraint. “You are a jeweler advising a retail buyer” beats a blank request every time.
  3. Ask it to question you first on anything that matters. The clarifying questions are where the value is.
  4. Connect your own data so it works from your files, not from guesses about the trade.
  5. Match the model to the task. Speed for the small jobs, depth for the decisions. That habit alone changes your output.

AI is moving fast. Stay in the loop.

Two places to stay connected, both free, both built for our trade.

WhatsApp Stay informed. Quiet updates. No noise. Join the WhatsApp group → Facebook Join the conversation. Real questions. Real use cases. Just opened. Join the Facebook group →
The model is the tool. The expertise is yours.
by INSIDER · insaiders.com
The quiet group