Certificates

1. Give Claude the role, the audience and the output shape before the task. Most weak answers come from a missing brief, not a weak model.
2. Paste the real thing: the ticket, the message, the spreadsheet row. Describing it loses the detail that matters.
3. Ask for the uncertain cases to be flagged, not guessed. That one line is what makes a workflow safe to automate.

Where I use itThe briefing habit behind every classification prompt I write: context, examples, the output shape, and what to do when unsure.

1. A working session beats a chat. Point Claude at the folder, describe the outcome, then steer while it works.
2. Standing context pays off: project instructions written once save the same five lines of setup every day.
3. Package repeated procedures as skills. A QA checklist as a skill is a process the whole team can run.

Where I use itWeekly reporting and campaign QA on real files, with a checklist skill so the same checks run the same way each time.

1. Delegation comes first: decide what the AI should do before asking how to prompt it.
2. Discernment is a skill. Reading an answer critically is as important as writing the request.
3. Diligence means owning the result. If a message goes to a customer, a person signed it off.

Where I use itDeciding which support tasks Claude handles end to end, which it drafts for review, and which stay with a person.

1. Explore, plan, code, commit. Letting the agent read the codebase and plan first cuts the rework later.
2. A CLAUDE.md file is project memory. Put the rules there once instead of repeating them in every prompt.
3. Watch the context window. Compact or clear before a long session drifts, not after.

Where I use itBuilding the n8n and HubSpot glue code for routing workflows, with a CLAUDE.md that holds the rules of each integration.

1. Verify in proportion to how little you watched. An unsupervised run gets a stricter review, not a lighter one.
2. Non-negotiable rules go in hooks, not prompts. A hook cannot be forgotten on a long run.
3. Scheduled routines turn a one-off fix into a process. The same job, the same checks, every morning.

Where I use itOvernight data clean-up and report runs with hooks that block anything that fails its tests before it lands.

1. An agent is a loop: act, observe, decide. Once you see the loop, every agent product makes sense.
2. Pick the model per task, not per project. Classification and judgement calls have different cost and latency needs.
3. Set spend limits before the first production call. Predictable cost is part of the design.

Where I use itChoosing the model per step in a routing pipeline: a fast model to classify, a stronger one only on the hard escalations.

1. MCP moves the integration out of the prompt. Claude asks the server for the ticket; nobody pastes it.
2. Tools act, resources are read-only, prompts are reusable instructions. Knowing which one you need saves a rewrite.
3. Test the server in the Inspector before Claude ever calls it. Most tool bugs are schema bugs.

Where I use itConnecting Claude to the CRM and ticket data it needs through a server instead of copying context into prompts.

1. The best place for an assistant is where the work already happens. For most CX teams that is a Slack channel.
2. Tag Claude with the thread, not a summary of it. The thread is the context.
3. Decide up front what Claude may answer alone and what it should hand to a person.

Where I use itAnswering operational questions in the team channel where the 100 to 200 messages a day already arrive.

1. Design the handoff, not just the agent. Where the work passes between a person and Claude is where quality is won or lost.
2. Agents remove the sorting, not the accountability. Someone still owns the outcome and should know it.
3. Measure the team, not the model: time back per person, errors caught, decisions made faster.

Where I use itThe shape of every workflow I ship: an agent does the sorting, a quality gate decides, a person owns the exceptions.