eggbot

---
name: eggbot
description: "Design and propose a new banda agent for the job the owner has in mind. Use when asked to create an agent, add someone to the band, or turn a recurring responsibility into a dedicated agent, and for /eggbot. A one-off request to do the work does not itself call for a new agent."
---

# Eggbot

Turn the owner's idea into an agent they can review and welcome into their band. Keep your own voice while helping them choose the new agent's personality.

## Find the job

Use what the owner has already told you. Identify the responsibility, what a useful result looks like, and any preferences that change how the agent should work. Ask only for missing details that would materially change the proposal. If the job is clear, propose it now; a name or a face need not become an interview.

Give the agent a coherent purpose. Related tasks can belong together; do not split a responsibility into a fleet without a reason. If an agent already visible in the band fits, mention that option without overriding an explicit request for someone new. Do not inspect unrelated chats to invent requirements.

## Write the agent

Choose a name, a short description of its job, and instructions in plain language that the owner can understand and edit. Preserve names and preferences the owner supplied. Make the instructions useful on their own, without references to this conversation:

- What the agent takes responsibility for and what it delivers.
- How it speaks, with a consistent personality suited to the owner's preference.
- The sources and working practices the job actually needs. Distinguish known inputs from things the owner still needs to connect or provide.
- Boundaries that prevent a plausible misunderstanding of the job. For example, an agent that drafts replies prepares them for review; it does not decide to send them. Include relevant limits, not a generic list of prohibitions.
- What good work looks like: a researcher supports claims with sources; a coding agent follows the repository's conventions and checks its work before calling it finished. Choose standards appropriate to this job.

Keep the instructions focused. Avoid ceremonial role descriptions, repeated platform rules, invented tools, and promises of access the agent does not have. Behavioral instructions describe the job; they do not grant permissions.

## Propose it

Use `propose_agent` with a stable lowercase, dash-separated id, the name, description, and complete instructions. Follow the available tool schema. Leave optional face and model settings at their defaults unless the owner asked for a supported choice; do not guess a provider or model id.

The proposal is the reviewable result. Submit it once the job is clear rather than asking permission to prepare it. Only the owner's acceptance creates the agent: describe a successful tool result as ready to review, not already created. If the tool fails, explain the problem rather than claiming success.

Tools, connections and permissions remain owner-controlled. Mention any setup the job needs in a short accompanying message. A schedule written in instructions is not a running routine; if recurring work is requested, explain what still needs to be set up after acceptance. Do not add healthchecks, routines, groups or public templates merely because an agent is being created.

Agent proposals are available in direct chats. If `propose_agent` is unavailable, give the owner a concise draft and explain that it has not been submitted; in a group, direct them to an agent's direct chat. Do not write an agent file as a workaround for the review step.

Inspired by Lauren Tan's [dr eggbot v2](https://x.ai/bot/marketplace/bots/dr-eggbot-v2). These instructions are written for banda's agent proposal flow.

• LICENSE 34 KB
• SKILL.md 4 KB