Models and modes
The trained predictor itself. Different models trade depth, speed, and cost.
Fastest, cheapest Claude in the family. Best for volume work where a miss is cheap.
The daily driver. Strong default for drafting, analysis, and everyday tool use.
Deep-work model. Use when a shallow miss would survive into IC or a customer-facing deliverable.
Frontier model. Reach for it only after Opus has clearly failed on the same problem.
How thorough a response should be. Separate from which model you picked.
Drafting space before the final answer. Lets Claude reason privately, then show or use that trail.
The word-pieces Claude reads and writes. You pay and fill the context window in these.
Everything the model can see right now. Finite, and yours to manage.
Surfaces
The default conversation surface. You stay in the loop between turns.
Workspaces that hold instructions, files, and chats for one stream of work.
Polished deliverables Claude produces as documents, apps, or other shareable outputs.
A mode/tooling path for gathering sources and returning cited synthesis.
Desktop-centered agentic work: plan, act across files and tools, return finished deliverables.
The software workshop surface. Same agentic loop, pointed at repos and code tools.
Claude inside another app, such as Chrome, Excel, or Outlook, so you skip copy-paste shuttles.
Stack vocabulary
Tools that ship with Claude: search, fetch, code execution, file creation. No setup.
Model Context Protocol. Open standard for how tools plug into AI apps. A socket, not a product.
A prebuilt plug into a named service, running under your permissions.
A folder of instructions and examples Claude loads when relevant. Procedure, not access.
Anything Claude can call that is not pure prediction: search, code, your systems.
One concrete request Claude makes to a tool, with arguments and a returned result.
Agents and loops
A model in a loop: plan, act with tools, observe the result, revise, repeat until done.
The four beats: Plan → Act → Observe → Revise. Stops when the goal is met, not when the prose ends.
A worker spun up for a scoped slice of work, often with different tools or instructions.
The finish line you write before the loop starts. Without it, agents stall or stop early.
Fluency and judgment
Deciding what Claude owns versus what stays human.
Briefing Claude clearly enough that the work is constrained and checkable.
Inspecting outputs for quality, risk, and fit before they leave your desk.
The ongoing habit of verifying, iterating, and owning the result.
Choosing what enters the window: files, instructions, history, and tools. Prompting is only one slice.
Safety and routing
Are you allowed to share this with an outside system?
Who is affected if the answer or action is wrong?
What reliable evidence can confirm the result?
Where can you stop, reverse, or recover if things go wrong?
Malicious instructions hidden in content Claude reads, aimed at steering tools or outputs.
Give Claude the smallest access that still completes the job. Writes need approval by default.
The connector or tool path is broken. Prompting harder will not fix it.
The tools worked, but Claude made a bad call. Needs better briefing, skills, or human review.