Beyond the single prompt

• A skill is a saved recipe: you teach it a task once, in plain words, and reuse it forever instead of re-explaining.
• It captures your way of doing something: the steps, the standard a good result must hit, and the "we don't do it like that" rules.
• The AI decides when to fire a skill from its description, so keep the description boring and exact, not clever.
• A skill saves the task; a Project saves your context. Use both: the Project knows your world, the skill knows the job.
• The trigger: the moment you find yourself writing the same kind of prompt a third time.

Turn this into a reusable skill I can run again later.
Name it something obvious. Write it as: when to use it, the exact steps, and the standard a good result must hit.
The task: [describe the repeatable job in plain English, including your rules and your style]
Save it, then show me the one-line command I'll use to run it next time.

The prompt above is the everyday version. When a misstep would be expensive, split the job in two: research gathers the facts, thinking makes the call.

Do deep research and give me a cited report I can trust.

QUESTION: [the specific thing you need answered]
WHY IT MATTERS: [the decision this feeds, so you can judge what's relevant]
MUST COVER: [3-5 things the report has to address, e.g. real pricing, recent reviews, market size, the catch]
SOURCE BAR: prefer recent, primary, reputable sources. Flag anything you could only find in one place as "single-source, unverified."

Deliver:
1. A direct answer up front, one paragraph.
2. The evidence, grouped by point, each claim with a citation.
3. Where sources disagreed, and which you trust more and why.
4. What you could NOT confirm.

Don't pad it. If a section has no solid evidence, say so.

Think hard about this before answering. Do NOT just agree with me.

DECISION: [the question or choice you're facing]
WHAT I'M LEANING TOWARD: [your current gut call, if any]
WHAT'S AT STAKE: [time, money, headcount, or reputation on the line]

Process:
1. Attack this from at least three independent angles (e.g. cost, risk, second-order effects, what a sceptic on the board would say). Treat them as separate viewpoints, not one.
2. Form the strongest answer.
3. Now try to DISPROVE that answer. Where is it weak? What would have to be true for it to fall apart?
4. Only if it survives, give me your recommendation, your confidence level, and the one thing most likely to break it.

If my leaning is flawed, tell me plainly. I want the truth, not reassurance.

• Describe a recurring job in plain words and Cowork runs it on a schedule: no reminding, no being there.
• One catch worth knowing: it only runs while your computer is awake and the Claude app is open. A missed run goes late, not never, so do not lean on it for anything time-critical.
• Schedule a task only after you have proven it works by hand. Automating an unreliable task just multiplies the mess.

Set this up as a scheduled task that runs on its own.

WHEN: [e.g. every Monday at 8am]
WHAT TO DO EACH RUN: [the job, step by step, in plain English]
WHAT I WANT BACK: [the exact output, e.g. a one-page summary saved to my drive]

Draft the scheduled task and show it to me to approve before you turn it on. Note: I understand it only runs when my computer is awake and the Claude app is open.

• Each sub-agent works in its own clean, separate context, so none get bogged down, then they report back and the results get combined.
• They work alone and cannot talk to each other, so this only fits work that splits into genuinely independent chunks.
• Anthropic's own test found a team of sub-agents beat a single agent by 90.2% on a research task, at roughly fifteen times the cost. Power and bill climb together. See the study.

• The step past this page: the AI writes a small plan, then fans out a large parallel job, sometimes dozens or hundreds of helpers at once, and merges everything at the end.
• /goal runs one job deep until it is right. This runs many jobs wide at the same time.
• It always asks before running. You will not trigger it by accident.
• Honestly, you will rarely need it. One person burned through half a $200 monthly plan with a single prompt at this level. If a normal chat, a skill, or /goal solves it, use that instead.

This may need a big parallel workflow. Before running anything, tell me:
1. Does this job actually break into many independent pieces that can run at the same time? If not, recommend a cheaper approach (plain chat, a skill, or /goal) instead.
2. Roughly how much will this cost in effort/tokens, and is it worth it?

If a big parallel run really is the right call:
THE JOB: [the big parallel task]
THE DELIVERABLE: [one clear, named output]
Put the helpers on the cheapest capable model, bound the scope tightly, and ask me to confirm before you run it.