Cortex Config

Cortex Config Skill

Create a Cortex YAML config from a short interactive conversation. Asks the user a few questions, designs the agent/team structure, shows a plan, then writes the file.

Step 1 — Choose mode

Ask the user which mode they want. Present all three with one-line descriptions:

1. Workflow (DAG) — structured execution with team dependencies. Teams run in parallel tiers, upstream results feed downstream prompts. Best for multi-step projects with clear ordering.
2. Mesh — autonomous agents with optional messaging. Each agent works independently, can see a roster of peers, and messages others only when needed. Best for parallel workstreams.
3. Gossip — decentralized knowledge sharing. Agents explore a topic from different angles, findings are exchanged periodically via gossip protocol. Best for research and brainstorming.

If $ARGUMENTS contains a mode or enough context to infer one, skip this question.

Step 2 — Gather project details

Ask questions in a single round (use AskUserQuestion). Tailor questions to the chosen mode.

All modes

1. Project name — what is this called?
2. What are agents working on? — describe the objective in 2-3 sentences. The user MUST provide a real description — without it you cannot design meaningful agents. If the answer is too vague, ask follow-ups.
3. Cluster context — any shared background all agents should know? (optional)

Mode-specific questions

Workflow (DAG): 4. How many teams? — suggest a number with names based on the description (e.g., "3 teams: backend, frontend, integration"). Include "auto" and custom options. 5. Model — sonnet (recommended), opus, or haiku 6. Permission mode — acceptEdits (recommended) or bypassPermissions

Mesh: 4. How many agents? — suggest specific agents with names and roles based on the description (e.g., "3 agents: market-sizing, competitor-analysis, product-strategy") 5. Model — sonnet (recommended), opus, or haiku 6. Permission mode — acceptEdits (recommended) or bypassPermissions

Gossip: 4. How many agents? — suggest specific agents with names and topics based on the description 5. Rounds — how many gossip exchange rounds? (suggest a default based on scope, typically 3-5) 6. Model — sonnet (recommended), opus, or haiku

Critical rule

Do NOT proceed to Step 3 until you have a concrete understanding of what the agents should accomplish. If the user gives a vague answer, ask follow-up questions. You need enough detail to write specific prompts — not generic boilerplate.

Step 3 — Design the config

From the user's answers, design the full config.

Workflow (DAG) design

• Name each team after its domain (e.g., backend, frontend, data-pipeline)
• Each team gets a lead with a descriptive role
• Add members with distinct roles and focus areas (2-5 per team)
• Write a rich context block per team
• Determine DAG dependencies — no circular deps allowed
• Design 4-7 tasks per team with summary, details, deliverables, and verify

Mesh design

• Name each agent after its domain (e.g., market-sizing, competitor-analysis)
• Give each agent a specific role (one-line description of what they are)
• Write a detailed prompt for each agent — what they should research/build/analyze, what deliverables to produce, and how to structure their output
• Set mesh settings (heartbeat/suspect/dead timeouts — defaults are usually fine)

Gossip design

• Name each agent after its exploration angle (e.g., competitor-analyst, market-sizer)
• Give each agent a topic and detailed prompt
• Set gossip settings (rounds, topology, exchange interval)
• Add seed knowledge if relevant

Step 4 — Present the plan

Show a formatted summary for user review. Format depends on mode:

Workflow plan

## Cortex Config: <name>
Mode: workflow (DAG)

### Defaults
model: <model> | max_turns: <N> | timeout: <N>min | permission: <mode>

### Execution Tiers

**Tier 0** (parallel):
  - <team>: <lead role> + <N members>

**Tier 1** (depends on: <tier 0 teams>):
  - <team>: <lead role> + <N members>

### Output
  <filename>.yaml

Mesh plan

## Cortex Config: <name>
Mode: mesh

### Defaults
model: <model> | max_turns: <N> | timeout: <N>min | permission: <mode>

### Agents
  - <name>: <role>
  - <name>: <role>

### Mesh Settings
heartbeat: <N>s | suspect timeout: <N>s | dead timeout: <N>s

### Output
  <filename>.yaml

Gossip plan

## Cortex Config: <name>
Mode: gossip

### Defaults
model: <model> | max_turns: <N> | timeout: <N>min

### Agents
  - <name>: <topic>
  - <name>: <topic>

### Gossip Settings
rounds: <N> | topology: <type> | interval: <N>s

### Output
  <filename>.yaml

Ask: "Does this look right? I'll write the config."

Step 5 — Write the YAML

Write the file to the output path from $ARGUMENTS, or default to examples/<project-name>.yaml inside the cortex directory.

Workflow schema

name: "<project name>"

defaults:
  model: <model>
  max_turns: <number>
  permission_mode: <mode>
  timeout_minutes: <number>

teams:
  - name: <team-name>
    lead:
      role: "<role description>"
    members:
      - role: "<member role>"
        focus: "<what this member owns>"
    context: |
      <rich domain context>
    depends_on:
      - <team-name>
    tasks:
      - summary: "<short description>"
        details: "<specific requirements>"
        deliverables:
          - "<file path>"
        verify: "<shell command>"

Mesh schema

name: "<project name>"
mode: mesh

cluster_context: |
  <shared context for all agents>

defaults:
  model: <model>
  max_turns: <number>
  permission_mode: <mode>
  timeout_minutes: <number>

mesh:
  heartbeat_interval_seconds: 30
  suspect_timeout_seconds: 90
  dead_timeout_seconds: 180

agents:
  - name: <agent-name>
    role: "<one-line role description>"
    prompt: |
      <detailed assignment — what to do, what deliverables to produce,
       how to structure output. Be specific.>

Gossip schema

name: "<project name>"
mode: gossip

cluster_context: |
  <shared context for all agents>

defaults:
  model: <model>
  max_turns: <number>
  permission_mode: <mode>
  timeout_minutes: <number>

gossip:
  rounds: <number>
  topology: <full_mesh|ring|random>
  exchange_interval_seconds: <number>

agents:
  - name: <agent-name>
    topic: "<knowledge topic>"
    prompt: |
      <detailed exploration instructions>

seed_knowledge:
  - topic: "<topic>"
    content: "<starting knowledge>"

Quality checklist before writing

• Project name is non-empty
• Every agent/team has a name
• No duplicate names
• Agent prompts are specific and actionable (not generic filler like "do research")
• Workflow: all depends_on references exist, no circular deps, every team has tasks
• Gossip: rounds >= 1, valid topology
• Mesh: positive timeout values

Step 6 — Print next steps

After writing the file, print mode-appropriate next steps:

Workflow

<filename> written!

Next steps:
  1. Review:  cat <filename>
  2. Dry run: mix cortex.run <filename> --dry-run
  3. Run:     mix cortex.run <filename>

Mesh

<filename> written!

Next steps:
  1. Review:  cat <filename>
  2. Dry run: mix cortex.run <filename> --dry-run
  3. Run:     mix cortex.run <filename>

Gossip

<filename> written!

Next steps:
  1. Review:  cat <filename>
  2. Dry run: mix cortex.run <filename> --dry-run
  3. Run:     mix cortex.run <filename>

Source: itsHabib/cortex — distributed by TomeVault.