5 signs you actually know how to use AI

You don't keep a document of 500 saved prompts

Current models follow plain, specific instructions well. What changes the output isn't a magic phrase, it's context: what you're making, who it's for, what good looks like and what to avoid. People who get great results ask for what they want in normal language, then correct it in the next message.

So instead of a library of prompts, keep one habit: make Claude ask you for the missing context before it starts.

I need [what you want, for example a follow-up email to a client who went quiet after a proposal].

Before you write anything, ask me up to 5 questions about the context you're missing: who it's for, what a great result looks like, the tone, and anything I want you to avoid. Ask them all at once.

Once I answer, write the first version, then tell me the one assumption you're least sure about so I can correct it.

You run it by hand before you automate it

AI can only automate a process that already works. If the steps are fuzzy in your head, an automation just produces wrong results faster and with more confidence. Doing it by hand once is how you find the steps you'd forgotten were steps.

1. ✋ Do the task yourself once, with Claude beside you. Work through it in a normal chat, one step at a time, and say out loud what you're checking at each point.
2. 📋 Turn that run into written steps. Paste the prompt below at the end of the chat so Claude writes up exactly what you did, including the judgment calls.
3. 🔁 Run the written steps two more times. Use them on real inputs. Wherever you had to step in and fix something, update the steps.
4. ⚡ Only now automate it. Once three runs come out right without you stepping in, it's ready to become a skill, a scheduled task or an agent.

Look back over everything we just did in this chat. Write it up as a step-by-step process someone else could follow to get the same result.

For each step include: what goes in, what comes out, and any judgment call I made (for example "I skipped leads with no company name"). Mark any step where the right answer depended on my taste rather than a rule.

Then list the steps that are safe to automate and the ones that should stay with a human, and explain why for each.

You know when not to use it

Some tasks take longer to check than to do. In a 2025 randomised trial by METR, experienced open-source developers took 19% longer to finish tasks when they used AI tools, even though they believed afterwards that AI had sped them up by about 20%. The study is here. Feeling faster and being faster aren't the same thing.

Your setup is boring

The best setups are folders and text files. One folder per project, one instructions file saying who you are and what good looks like, and notes in plain text. Elaborate systems built over a weekend tend to get abandoned. A boring one gets used every day.

1. 📁 One folder or Project per area of work. In the Claude app, create a Project for each area and put your notes and reference files in it. In Claude Code, use a normal folder on your computer.
2. 📄 One instructions file. In a Project, that's the project instructions. In Claude Code, it's a file called CLAUDE.md in the folder, which Claude reads automatically at the start of every session. How CLAUDE.md works.
3. ✏️ Update it when Claude gets something wrong twice. That's the whole maintenance plan. The file improves every week without you sitting down to "build a system".

Interview me to write the instructions file for this project. Ask one question at a time, and stop at 8 questions.

Cover: what this project is, who the work is for, what a great result looks like with one example, the tone I want, the mistakes you should never make, and the tools or files you should use.

Then write it as a short plain-text file under 300 words, with simple headings, no fluff. Keep only rules that would change your output.

You know your subagent formations

A subagent is a separate Claude with its own clean context, given one job and reporting back. Once you can run more than one, the skill is choosing how they work together. Anthropic's guide to building effective agents names these patterns, and their multi-agent research system (a lead agent directing subagents) beat a single Claude Opus 4 agent by 90.2% on their internal research evaluation.

Where to run them

1. 💻 Install Claude Code. Subagents run in Claude Code, which is included in the Pro and Max plans. Plans and pricing.
2. 📂 Open it in your project folder. Run claude in the folder with your files and CLAUDE.md.
3. 🧩 Just ask for the formation. Claude Code can launch subagents on its own when you describe the formation in your prompt, like the ones below. To save a worker you reuse often, type /agents and create it there. Subagents docs.

Here is the job: [describe the task and what the finished result should be].

Before doing anything, tell me which formation fits best: a single agent, orchestrator and workers, consensus voting, or a debate with a referee. Explain in two sentences why, and name the formation you rejected and why.

Then estimate roughly how much more work it is than a single agent, and wait for me to say go.

Act as the orchestrator for this job: [task].

1. Split it into 3 to 5 parts that don't depend on each other, and show me the split first.
2. Launch one subagent per part, in parallel. Give each a clear brief: its goal, what to return, what to ignore, and the format for its answer.
3. When they report back, check each result against its brief, send back anything incomplete, then combine everything into one final result.

List any contradictions between workers instead of quietly picking one.

I need a reliable answer to this: [question or item to classify or check].

Launch 3 subagents in parallel. Give each the same instructions and the same material, but don't let them see each other's work. Each returns an answer, a confidence level from 1 to 5, and the evidence it used.

Report the majority answer. If they don't all agree, show me where they split and the strongest evidence on each side, so I can make the final call.

I'm deciding: [the decision, for example whether to rebuild our onboarding emails or fix the existing ones].

Launch two subagents. One argues for option A, the other for option B. Each gets the same context: [paste the context]. Each writes its strongest case, then reads the other's case and writes a rebuttal.

Then launch a third subagent as the referee. It reads both cases and both rebuttals, and decides which argument is stronger and why. It also names the one fact that would flip the decision.

So, how many have you got?

Pick your lowest sign and fix only that one this week. One boring habit, used every day, beats five you set up and forget.

Proof

Building effective agents (Anthropic)
How Anthropic built its multi-agent research system
Improving factuality and reasoning through multiagent debate
METR study on AI and experienced developers
Claude Code subagents docs