Prompting Agents for Outcomes
Pyyhkäise näyttääksesi valikon
You've prompted chatbots before. You've asked Claude to write an email, summarize an article, or explain something. Prompting an agent is different. It is not harder, just different.
If you bring chatbot prompting habits to an agent, you will get weaker results than you should.
Here is the core shift: when you prompt a chatbot, you are asking for an answer. When you prompt an agent, you are describing an outcome and letting it figure out the steps.
The agent has tools now. You do not need to walk it through every click.
"Write me an email to Sarah letting her know I can't make Thursday's meeting and offering Friday or Monday as alternatives."
"Reschedule my Thursday meeting with Sarah to Friday or Monday."
The chatbot version describes the content of the email. The agent version describes the outcome and lets the agent figure out which app to use, what email to read, what calendar to check, what draft to write. Same end result, very different prompting style.
Three pieces. Get all three in there, even if loosely worded, and your agent will thrive.
A few more prompting habits will pay off.
Be specific about quantity. "A few" is fuzzy. "Five" is clear. Agents do better with numbers.
Name the apps you want the agent to use. "Drop the result into a Google Sheet" is better than "give me a comparison." This way, the agent does not have to guess.
Allow it to ask follow-up questions. Add: "If anything is unclear, ask before you start." This one phrase can prevent a lot of bad runs.
Iterate instead of rewriting. If the first run does not nail it, do not start over. Reply with something like: "Rerun this, but include battery life this time." Treat the agent like a colleague you are giving feedback to.
What Not to Do
Do not micromanage every step. "First open Chrome, then search for X, then click the third result..." is too detailed. If you spell out every click, you are not really using the agent.
Do not pile five unrelated tasks into one prompt. Agents focus better on one objective at a time. Chain tasks in conversation when needed.
Do not be vague about the output format. "Write something up" is too open-ended and often leads to weak results.
You will use this structure in every workflow in this section: goal, constraints, and done condition. Get comfortable with it now.
Kiitos palautteestasi!
Kysy tekoälyä
Kysy tekoälyä
Kysy mitä tahansa tai kokeile jotakin ehdotetuista kysymyksistä aloittaaksesi keskustelumme