Scaling up — deployment, ROI & cost control
Adopting Claude on your own is easy; getting a measurable gain across a team takes a method. Here is a 4-week deployment framework, followed by the habits that keep you from wasting your credits.
The 4-week method
Map it out. List everything that recurs more than once a week. Separate thinking (creating, deciding) from execution (organizing, extracting, filling in). Spot the tasks where quality depends on human focus.
Automate 3 tasks — no more. Meeting notes, CRM-driven email follow-ups, document summaries. A lasting system starts with 3 workflows that are 100% reliable, not 10 shaky ones.
Create Skills. Turn each recurring task into a pre-configured command (e.g. /followup, /notes). Your workflow, your voice, reusable across the whole team.
Measure and iterate. Calculate the time saved per person per week on your data, identify the next task to automate, and document the system so it can be handed off.
Controlling your costs — the right anti-waste habits
Technical note: prompt caching also cuts costs on repeated prompts.
📄 PDF → Markdown
Converting a heavy PDF to .md before analysis dramatically lightens the context, for the same result.
🧠 Plan in Chat, execute in Cowork
Brainstorming costs little in Chat; save Cowork for handling files.
🔎 Search / Connectors on demand
Turn them on only when the task requires it — an unnecessary search weighs down every exchange.
✏️ Edit rather than resend
Edit your message instead of piling on follow-ups: you start again from a clean context.
🎯 One task per message
Three clear requests beat one confused message — for you as much as for Claude.
⚡ The right model
Sonnet 5 every day; Opus 5 for heavy work; Fable 5.1 for the summit (see the single rule, module 02).
🗂️ Reuse Projects
Load your files once into a Project; no need to re-paste them into every conversation.
🔁 New thread regularly
Over a very long conversation, Claude rereads everything: open a new thread per topic.
📝 Summarize before starting over
Session dragging on? Ask "summarize our exchange in key points (decisions, figures, to-dos)," then start fresh in a new thread with that summary. Claude no longer rereads 40 messages, just 10 lines. Near the limit, Claude does it automatically — but doing it yourself keeps it from losing a figure or an important detail.
🧹 Clean up before pasting
Don't paste an entire web page (menus, ads, cookies) or an email thread with signatures: keep the useful passage. On raw content, this is the most spectacular saving.
🖼️ Lighten your images
A very high-resolution screenshot or photo costs a lot, often for nothing. If the text stays legible, shrink it before sending.
✂️ Ask for brief
What Claude writes counts too. "Answer directly, no preamble" or "in 5 bullets" avoids needless pages when you just want the essentials.
📋 Concise Project instructions
A Project's instructions are reread with every message. Keep them short and sharp: a rambling instruction is paid for continuously and Claude follows it less well.
Optimizing your tokens — the playbook for recent models
A token is Claude's unit of account: a piece of a word (an English word ≈ 1.3 tokens). Everything Claude reads — your messages, your files, the conversation history — and everything it writes is counted in tokens. Understanding how that meter runs means doubling the value of your plan without spending a euro more.
How your budget is counted
Two counters run in parallel: a limit per 5-hour session and a weekly limit. Track them in real time in Settings → Usage — two progress bars, no surprises.
Every message rereads the whole conversation. On the 40th exchange of the same thread, you pay to reread the previous 39. That's the reason for the "new thread per topic" habit seen above.
Everything counts: message length, attachments (one PDF page ≈ 1,500 to 3,000 tokens!), tools turned on (Search, connectors), the model chosen, the effort level, and creating artifacts.
Projects work in your favor: their content is cached — a document loaded into a Project doesn't "reweigh" with every message, unlike the same document re-pasted into each conversation.
The levers specific to Opus 5 and Fable 5.1
⚙️ Set the effort level
Fable 5.1 always thinks ("adaptive" reasoning): the effort slider is your main lever. Low effort for the routine, high for the strategic — it's the number-one factor in consumption.
💰 A premium model — dose it
Fable 5.1 costs twice as much as Opus 5 per token. Save it for the summit; Opus 5 for heavy work, Sonnet 5 for most everyday needs.
🧠 1 million tokens of context
Enough to load a whole case file at once. But a massive context is paid for with every exchange: load what's useful, not the full archive "just in case."
🛟 Refusals aren't billed
If a safety guardrail declines a request (fewer than 5% of sessions), nothing is counted and the automatic fallback to Opus takes over.
Time a recurring task this week, then redo it with Claude and compare.
Expected result : A before / after reference figure, on your own data.