agent-creator

Agent Profile Creator

Create agent profiles stored as config.json in ~/.aider-desk/agents/{name}/ (global) or {project}/.aider-desk/agents/{name}/ (project-level).

Read all reference files before proposing a profile to ensure accuracy:

• references/agent-profile-schema.md — Complete property reference
• references/subagent-guide.md — Subagent config, system prompts, context memory, filtering
• references/tool-approval-guide.md — All tool groups, correct tool keys, approval strategies
• references/profile-examples.md — Complete working examples

Q&A Process

Gather complete requirements before proposing a profile. Ask focused questions, propose internally, and present one final summary for approval.

Step 1: Purpose & Behavior

Ask: "Describe the agent's purpose and what it should do."

Based on the response, internally propose:

• Name (derived from purpose)
• Max iterations (default 0 = unlimited; use a finite value like 250 for bounded tasks, or lower for focused tasks like review/analysis: 150)
• Tool groups (based on purpose — see Step 3)
• Custom instructions (free-text guidance appended to the system prompt)
• System prompt — does the agent need a fully custom system prompt that replaces the default built-in one? If the user describes specific behavioral constraints, role, or identity, propose a systemPrompt.
• Subagent config — should this profile be usable as a subagent? If yes, gather subagent-specific system prompt, invocation mode, color, description, and context memory mode.

Step 2: Provider & Model

Ask: "Which provider/model? (format: provider/model, e.g., anthropic/claude-sonnet-5)"

• Split by first slash to get provider and model
• Use as-is (correct obvious typos only)
• If the user does not specify, default to anthropic / claude-sonnet-5 (matching DEFAULT_AGENT_PROFILE / DEFAULT_PROVIDER_MODELS in packages/common/src/agent.ts). You may ask the user for their preferred provider/model, but if none is given, use these defaults.
• Optionally ask about temperature and maxTokens if the purpose implies precision (low temp) or creativity (high temp)

Step 3: Tool Groups & Approvals

Based on the purpose, propose tool groups to enable. Present each group with a brief rationale:

For each enabled group, propose tool approval defaults (always, ask, never). Only include tools in toolApprovals that deviate from the default "ask".

If useSubagents is enabled, ask: "Should this agent use all available subagents or only specific ones?" — set enabledSubagentIds if filtering is needed (omit for all).

Step 4: Subagent Configuration

Ask: "Should this profile be usable as a subagent by other agents?"

If yes, gather:

• Subagent system prompt — specialized prompt for subagent role (distinct from the main agent systemPrompt)
• Invocation mode — automatic (needs a clear description) or on-demand
• Context memory — off (default), last-message, or full-context
• Color — visual identifier
• Description — required for automatic invocation; should contain domain keywords

Step 5: Advanced Settings (Optional)

Only ask if user mentioned any of these:

• Temperature (0.1-1.0)
• Max tokens
• Min time between tool calls (rate limiting, in ms)
• MCP servers to enable (enabledServers)
• Rule files (markdown files placed in ~/.aider-desk/agents/{name}/rules/)
• Auto-compaction overrides (autoCompactThresholdPercentage, autoCompactThresholdTokens, autoCompactionType)
• Bash tool settings (allowedPattern / deniedPattern regex)

If not mentioned, skip this step entirely.

Step 6: Location

Ask: "Global profile (all projects) or project-specific? (default: global)"

Step 7: Review & Confirm

Present one complete summary with all proposed properties:

• Name, provider/model, temperature/maxTokens (if set)
• Tool groups enabled with rationale
• Tool approvals (which are "never", which are "always"; rest default to "ask")
• System prompt (if any)
• Custom instructions (if any)
• Subagent config (if enabled)
• Subagent filtering (enabledSubagentIds, if applicable)
• Rule files (if any)
• Location (global/project)

Ask: "Here's your agent profile. Should I create it?"

Do not ask for confirmations on individual items. Only one final approval.

Step 8: Create

On confirmation:

1. Choose a unique, readable id (e.g., code-reviewer, security-auditor). No UUID is required — readable names matching the examples are fine, as long as it is unique.
2. Build the complete config.json — verify structure against references/profile-examples.md
3. Do not include ruleFiles in config.json — rule files are dynamically discovered from the rules/ subdirectory
4. Do not include isSubagent — it's a runtime flag
5. Create the directory and write config.json
6. If rule files were requested, create them as .md files in a rules/ subdirectory next to config.json

Tool Approval Strategy

Keys: {group}---{tool} (three dashes as separator)

Default is "ask". Only explicitly set:

• "never": Tools completely irrelevant to the agent's purpose
• "always": Safe, essential, read-only operations

Do not include tools that should remain at default ("ask").

Only include tool keys that exist in references/tool-approval-guide.md.

System Prompt Strategy

Two distinct system prompt fields exist:

1. Top-level systemPrompt — overrides the default built-in system prompt when this profile runs as the main agent. Rules files and customInstructions are still appended. Use when the agent needs a completely different identity or behavioral framework.

2. subagent.systemPrompt — used when this profile runs as a subagent. Passed directly to the subagent's LLM. If empty, falls back to the top-level systemPrompt.

When to use which:

• Agent works only as main agent → set top-level systemPrompt
• Agent works only as subagent → set subagent.systemPrompt
• Agent works in both roles → set both; they can differ for context-appropriate behavior
• Agent doesn't need custom prompts → leave both empty; default system prompt is used

Subagent Configuration

Every agent can be a subagent. Set subagent.enabled: true to allow delegation.

Context memory modes:

• "off" — fresh each time (default for most agents)
• "last-message" — receives last result from previous runs
• "full-context" — full conversation history (for complex iterative analysis)

Subagent filtering: enabledSubagentIds?: string[] on the parent profile controls which subagents it can use. Omit to allow all. Set to an array of profile id values to restrict. Note: enabledSubagentIds holds profile id values (the same id field set on each profile), not directory names — they only match when a directory name happens to equal the profile id.

See references/subagent-guide.md for detailed guidance.

Minimal config.json Structure

{
  "id": "my-agent",
  "name": "my-agent",
  "provider": "anthropic",
  "model": "claude-sonnet-5",
  "maxIterations": 0,
  "minTimeBetweenToolCalls": 0,
  "enabledServers": [],
  "toolApprovals": {
    "power---file_read": "always",
    "power---bash": "ask"
  },
  "toolSettings": {},
  "includeContextFiles": false,
  "includeRepoMap": false,
  "usePowerTools": true,
  "useAiderTools": false,
  "useTodoTools": true,
  "useSubagents": false,
  "useTaskTools": false,
  "useMemoryTools": true,
  "useSkillsTools": true,
  "useExtensionTools": true,
  "disabledExtensionTools": [],
  "customInstructions": "",
  "subagent": {
    "enabled": true,
    "contextMemory": "off",
    "systemPrompt": "",
    "invocationMode": "on-demand",
    "color": "#3368a8",
    "description": ""
  }
}

Note on defaults shown above: id is any unique readable string (no UUID needed). When the user does not specify provider/model, default to anthropic / claude-sonnet-5. maxIterations: 0 means unlimited. includeContextFiles and includeRepoMap both default to false — the values above are defaults; set them to true only when you want context files / repo map included. The subagent color default is #3368a8.

Validation Checklist

Before writing config.json:

• Unique name
• Unique, readable id (no UUID required)
• Provider/model present (use as-is; default to anthropic/claude-sonnet-5 if unspecified)
• Tool keys use {group}---{tool} format with three dashes
• Tool approval values are "always", "ask", or "never"
• Only tools from references/tool-approval-guide.md are in toolApprovals
• systemPrompt set only if custom behavior is needed
• subagent.enabled reflects whether this profile should be delegable
• If subagent.invocationMode is "automatic", description is non-empty
• If useSubagents is true and filtering needed, enabledSubagentIds is set
• useExtensionTools and disabledExtensionTools included (default: true / [])
• ruleFiles NOT in config.json (dynamically discovered)
• isSubagent NOT in config.json (runtime flag)
• All boolean tool group flags present
• Directory location matches user preference (global/project)

Resources

• references/agent-profile-schema.md — Complete schema with all properties and types
• references/subagent-guide.md — Subagent config, context memory, filtering, system prompts
• references/tool-approval-guide.md — All tool groups and correct tool keys
• references/profile-examples.md — Complete working examples
• assets/templates/config.json.template — Template
• assets/examples/sample-profile.json — Example profile