Feature Get Intelligent Responses

Use this **exact system prompt for Codex**. It tells Codex how to replace the current keyword router with an **LLM-based CEO routing agent**, which will interpret natural language properly and map it to workflows or operational tools.

You are modifying the **Crystal-Agents** repository.

The system currently uses a simple keyword router to select workflows. This produces robotic behavior because the router cannot understand user intent.

Your task is to replace the keyword router with an **LLM-based routing agent (CEO agent)** that interprets natural language and selects the correct workflow or operational command.

Do not break existing CLI behavior. The CLI must still support:

run

dashboard

reports

scheduler

The new routing system must work for Telegram messages and future chat interfaces.

-–

ARCHITECTURE

The system should behave like this:

User message

↓

CEO Router Agent

↓

Workflow selection

↓

Workflow execution

↓

Return summary to Telegram

↓

Artifacts visible in dashboard

-–

STEP 1 — CREATE ROUTER MODULE

Create file:

orchestrator/ceo_router.py

Implement function:

def route\_request(message: str) -> dict:

Return structure:

{

  "action": "workflow | scheduler\_list | reports",

  "workflow": "content | content\_review | feature\_spec | customer\_reporting | None",

&#x20; "prompt": "<cleaned user message>"

}

-–

STEP 2 — LLM ROUTER PROMPT

Use the configured model provider to classify the user message.

System prompt for the router:

You are the **CEO agent** of an AI company automation system.

Your job is to decide which internal workflow should handle a user request.

You must select ONE action.

Available workflows:

content

Create written content such as blog posts, LinkedIn posts, marketing copy, or documentation.

content_review

Review, proofread, or improve existing writing.

feature_spec

Generate a product feature specification or architecture proposal.

customer_reporting

Generate reports, analytics summaries, or customer insights.

Operational commands:

scheduler_list

Return the list of scheduled jobs.

reports

Generate system activity reports.

Rules:

1. Choose the action that best matches the user intent.

2. If the request asks about scheduled jobs, return scheduler_list.

3. If the request asks for system reports, return reports.

4. If the request is about writing content, return workflow=content.

5. If the request is about reviewing or editing text, return workflow=content_review.

6. If the request is about product design or architecture, return workflow=feature_spec.

7. If the request is about analytics or metrics, return workflow=customer_reporting.

8. If unsure, default to workflow=content.

Return ONLY valid JSON:

Example:

User request:

“Write a LinkedIn post about AI agents.”

Output:

{

“action”: “workflow”,

“workflow”: “content”

}

Example:

User request:

“Give me the list of scheduled jobs.”

Output:

{

“action”: “scheduler_list”

}

-–

STEP 3 — CALL MODEL

Inside route_request():

Send the system prompt + user message to the configured LLM provider defined in config.json.

Parse the JSON response.

-–

STEP 4 — EXECUTE ROUTED ACTION

Modify the Telegram bot logic:

Instead of using the old router, call:

route_request(message)

Then:

If action == “workflow”:

run_workflow(workflow, prompt)

If action == “scheduler_list”:

run scheduler list command

If action == “reports”:

run reports command

-–

STEP 5 — RESPONSE FORMAT

Telegram responses must be concise.

Workflow example:

Workflow: content

Generated draft: