Skip to main content
Instructions are a key lever on agent quality. Skills give an agent knowledge and tools give it capability, but instructions can help decide whether the output or actions are useful.

​
Focus on outcomes, not steps

Describe what a good result looks like. Let the agent figure out how to get there. Agents are stronger at executing toward a defined outcome than at following a rigid procedure, and step-by-step instructions tend to break the moment the input shape changes. Weak—step-by-step
Open the Support channel. Read every data point from the last seven days. Group them into themes. For each theme, write a paragraph. Add a quote. Put it all in a doc.
Strong—outcome-focused
Summarize new data in the Support channel from the past week. Group findings by theme, lead with the most urgent issue, and include one direct quote per theme. If there are no findings, say so and do not create them. Output as a doc titled “Support digest—[week of].”
The second version tells the agent what “done” looks like. It can adapt when the data is thin, when a theme dominates, or when nothing urgent has come in.

​
Anatomy of a strong instruction

Every effective instruction answers four questions:
  • What is the agent producing? A doc, a channel comment, a tag applied, a message sent.
  • What data should it pull from? A specific channel, project, folder, or time window.
  • How should the output be structured? Sections, ordering, length, format.
  • What decisions does the agent get to make? Which items to prioritize, what to filter out, what to do when it can’t complete it’s task.
If any of these is unclear, the agent will guess.

​
Common patterns

Weekly digest
Every Monday, summarize new data added to the [Channel name] channel over the past week. Group findings by theme, lead with the most urgent issue, and include one direct quote per theme. Output as a doc in the [Folder name] folder.
Tagger
For each new data added to project X, apply tags from the taxonomy in the attached skill doc. Only apply tags that clearly match—if a data point doesn’t fit any tag, leave it untagged rather than forcing a match.
Summarizer
When triggered, read the attached transcript and produce a one-page summary with three sections: what was discussed, key decisions, and open questions. Include timestamps for anything the reader might want to revisit.
Competitor watch
Every Friday, search the web for public announcements from [competitors] over the past week. Summarize each into two sentences—what shipped and why it matters to us. Output as a doc and email me.

​
How personas actually change output

A persona shapes tone and phrasing. It doesn’t change what the agent does or what it can access. Think of it as adjusting the voice, not the job. Same instructions, different personas Instruction: Summarize the week’s support tickets grouped by theme.
  • Persona: “Concise analyst” → short sentences, bullet points, no editorializing
  • Persona: “Friendly researcher” → warmer framing, more context around each theme, gentler language on urgent issues
  • Persona: “Skeptical PM” → leads with the pattern that suggests a product problem, questions assumptions in the data
Personas are most useful when the output has a human audience. For agents that only tag data or route items, a persona adds nothing—skip it.

​
Common mistakes

Too vague. “Summarize the channel” leaves format, cadence, ordering, and destination undefined. Every run will look different. Too rigid. Listing every step forces the agent down a brittle path. When the data doesn’t match the shape you assumed, the whole run derails. Mixing multiple jobs. An agent that tags data, writes summaries, and sends notifications will do all three worse than three focused agents would do individually. Assuming context. The agent only knows what you’ve written into instructions and what’s attached as skills. If a term is internal shorthand, define it or attach a glossary.

​
Test before you schedule

Run any new agent on demand two or three times before setting a recurring trigger. Read the output, adjust the instructions, and run again. 20 minutes of iteration up front prevents a month of scheduled runs producing the wrong thing.