Docs / Getting started / Sending messages

Sending messages

CloudyBot runs in threads in the dashboard. Each message you send is scoped to that thread — the model sees prior turns, tool results, and files you reference. Treat every message as a mini spec: goal, inputs, constraints, and output format. Vague chat works for brainstorming; execution needs specifics.

New here? Run one end-to-end job first: your first task. Context rules: threads & memory.

Mini-spec checklist

• Goal — what you need and why. “I need a competitor pricing table so I can update our sales deck.”
• Inputs — URLs, paths under ~/files/, or named integration objects (Notion database, sheet tab, GitHub repo). Upload attachments in the thread or reference files already in your workspace.
• Constraints — tone, length, read-only vs write, timezone, “do not send email,” “no login required,” or “stop after 5 results.”
• Format — markdown table, numbered list, JSON, Slack-ready paragraph, or “save as output.csv in my workspace.”

If any line is missing, the agent may guess — often correctly, sometimes not. One missing field is cheaper to fix in a follow-up than restarting in a new thread.

Vague vs executable

• Weak: “Help with our blog SEO.”
  Stronger: “Crawl https://oursite.com/blog — list every post URL, title tag, and meta description. Flag titles over 60 characters. Markdown table.”
• Weak: “Check Notion.”
  Stronger: “Read database ‘Launch tasks’ — count items where Status is Blocked. Bullets: task name + owner. No edits.” (Requires Notion connected — integrations.)
• Weak: “Fix this file.”
  Stronger: “Read draft.md in my workspace. Fix spelling, add an H1 from the first line, save as draft-clean.md.”

Attachments and file references

Upload small files in the chat UI when the task depends on local content (notes, CSV, screenshot). For follow-up turns, refer by filename — the agent resolves paths under your account workspace. Large binary exports are better produced by the agent (file_* tools) than pasted inline. Skill docs: file editing, file management.

When to ask for a plan first

On large or risky work — multi-page browser flows, writes to Notion/Sheets, outbound email or WhatsApp — ask for an outline or numbered plan, then reply with “proceed with step 1 only” or “approved — send”. That pattern saves credits and prevents surprise sends. Free accounts have 0 outbound WhatsApp sends; email and other caps are in limits FAQ.

Follow-ups and corrections

Stay in the same thread when iterating on one job — the agent already has tool output and filenames. Use short deltas: “Add a fourth column for annual price.” or “Redo step 2 — use the EU pricing page instead.” Start a new thread when the topic changes; unrelated context dilutes answers (see threads & memory).

Specialists and skills

Messages to a Specialist should name the duty, not the whole product. Example: to Scout, “Run today’s competitor homepage check for {domains} — same format as last time.” Skills must be enabled on that Specialist and credentials connected under Settings → Connect. The skills index lists what each tool family can do.

When messages fail

Common fixes: create a thread first, connect the integration, narrow the scope, or split browser work across turns. Troubleshooting · Status · Contact.