How to Prompt Your Agents

Brandstaq agents are autonomous, but they still need clear direction. Talk to them like a manager talks to a smart new hire — clear intent, specific outcome, no ceremony. This guide covers the patterns that route the first time, the verbs the system uses to auto-pick skills, and the most common mistakes.

The 3-line rule

Every good prompt has three parts. Drop any of them and you'll get a worse result.

1. What you want done — the verb. "Post", "research", "draft", "send", "deploy".
2. The object — what specifically. "this image", "the Q2 launch", "Buffer's pricing", "my LinkedIn account".
3. The constraint — when, where, how, for whom. "before Tuesday", "on Twitter", "matching our brand voice", "for the sales team".

Bad:  "make a post"
Good: "Draft a Twitter post about our pricing change, in our voice,
      for review before publishing"

The good version names a skill (draft_post), names the platform (Twitter), names the constraint (review approval mode), and gives the agent a topic. It will route to the right skill on the first try.

Verbs that auto-pin a skill

Brandstaq's chat agent watches your message for imperative verbs that map to specific skills. When it sees one, it pins that skill before the LLM can decide otherwise — so you skip the "let me think about which skill to use" round-trip.

The full mapping lives in the Skills Catalog. Browse by category to see what verbs surface what skill.

When to be vague vs specific

Vague prompts work when:

• The intent is obvious from one verb ("research our top competitors" → research)
• You want the agent to decide ("plan today's posts" — agent picks topic, platform, voice)
• The brand context already covers the constraint (voice, audience, key messages)

Specific prompts work when:

• You have a tight constraint the agent can't infer ("schedule for Tuesday 9am ET, not now")
• You want a particular skill, not the agent's choice ("use generate_chart, not generate_pdf")
• The brand context is broad and the agent might pick the wrong angle ("for our Pro tier audience specifically")

A reasonable rule: if you've explained the same constraint twice in a session, save it as a brand learning so you don't have to say it again.

"Always reply to negative reviews within 4 hours, not days"
   ↓
agent calls add_brand_learning, then auto-applies it forever

Common mistakes

1. Asking for the work AND the skill

Bad:  "Use the generate_pdf skill to make a one-pager about our launch"
Good: "Make a one-pager about our launch"

You don't need to name the skill. The agent picks it. Naming it wastes tokens and sometimes overrides a better choice.

2. Stacking unrelated requests in one message

Bad:  "Post on Twitter, then check engagement, then deploy a sales agent,
       then book a meeting with John"
Good: (one request, then the next, then the next)

The agent CAN compose flows ("research X then write a thread") but stacked unrelated requests degrade routing quality. Send them one at a time unless they're genuinely a single workflow.

3. Leaving the constraint blank

Bad:  "Plan some content"
Good: "Plan 3 LinkedIn posts for next week, focused on our Pro tier"

"Some content" is a polite way of saying "I don't know what I want." The agent will pick something, and it might be fine, but you'll iterate more than if you'd named the constraint up front.

4. Talking about the agent in third person

Bad:  "Tell the agent to post on Instagram"
Good: "Post this to Instagram"

You're talking to the agent. It will route the request itself.

5. Confirming work that was already done

Bad:  "Did you post that?" (after the agent already replied "Posted at 9:14am ET")
Good: (trust the agent's reply)

Brandstaq agents don't lie about completion — they tell you exactly what they did with timestamps and links. If you re-ask, the agent will confirm but it's wasted tokens.

Power patterns

Compose flows in one prompt

"Research what people are saying about our new feature on Twitter,
then draft 3 reply-to-feedback threads in our voice for review"

The agent chains research → draft_post_angles → review queue.

Handoff to a deployed agent

"@Pelumi take a look at last week's engagement and plan accordingly"

The @AgentName mention routes the request to that specific agent's session, not the chat agent. The agent uses its own configured schedule, voice, and goals.

Set defaults so you don't repeat yourself

"Set the default approval mode to review_first for all my agents"
"Set the social manager's voice to Nova"
"Cap this agent at 1000 tokens per day"

These calls update brand and agent config. Subsequent work picks them up automatically — you don't have to mention "review first" or "use Nova" in every message.

Next steps

• Browse the full Skills Catalog to see every verb the system understands.
• Read Voice & Tone to teach the agent how your brand sounds.
• Read Approval Modes to control when work goes live without you.