Problem
Team creation during squad init is expensive and inconsistent:
- High token cost — The LLM improvises all role definitions from scratch every time (charter content, expertise, boundaries, voice)
- Slow setup — Multiple LLM round-trips to generate boilerplate charters that end up thin and generic
- Inconsistent quality — Charter depth varies wildly depending on the LLM's creativity in that moment
- No institutional knowledge — There's no curated baseline for what makes a good "Backend Developer" charter vs a "Frontend Developer" charter
- Limited scope — Squad is used for all kinds of teams (marketing, sales, game dev, etc.), but the current
personalityForRole() in cast.ts only knows ~9 engineering-flavored patterns
Proposed Solution
Add 20 built-in base roles that serve as starting points during team casting. Each base role includes deep, curated charter content (expertise, boundaries, voice, ownership, routing patterns) that provides ~90% of the charter out of the box. A lightweight LLM refinement pass adapts each base role to the specific project context (stack, frameworks, domain) — adding the remaining 10% at ~1/5th the current token cost.
The 20 Base Roles
12 Software Development roles (primary use case):
lead, frontend, backend, fullstack, reviewer, tester, devops, security, data, docs, ai, designer
8 Category Generalist roles (covering broader team types):
marketing, sales, product, project-mgmt, support, game-dev, media, compliance
Init Flow: Start from Base, Refine for Context
- User describes their project
- Deterministic keyword matching picks 4-5 base roles from catalog (no LLM needed)
- Present matches with one-liner descriptions; user confirms/swaps/adds
- Lightweight LLM pass refines each base role for project context (e.g., base
backend → "Node.js/Express Backend Dev with Stripe knowledge")
- Character name casting uses existing universe system
- Fallback to full LLM generation if no base role fits
Deliverables
packages/squad-sdk/src/roles/ — types, catalog (20 role objects), public APIuseRole('backend', { name: 'kane' }) — SDK builder for squad.config.tssquad roles — CLI command to list/search base roles- Updated
cast.ts — generateCharter() and personalityForRole() use catalog
- Updated
coordinator.ts — Init mode uses catalog-based matching + refinement
- Updated Init Mode prompts in
squad.agent.md templates
Attribution
Base role content is inspired by and adapted from msitarzewski/agency-agents (MIT License) by AgentLand Contributors. Proper attribution will be included in source comments, generated charter headers, and documentation.
Problem
Team creation during
squad initis expensive and inconsistent:personalityForRole()incast.tsonly knows ~9 engineering-flavored patternsProposed Solution
Add 20 built-in base roles that serve as starting points during team casting. Each base role includes deep, curated charter content (expertise, boundaries, voice, ownership, routing patterns) that provides ~90% of the charter out of the box. A lightweight LLM refinement pass adapts each base role to the specific project context (stack, frameworks, domain) — adding the remaining 10% at ~1/5th the current token cost.
The 20 Base Roles
12 Software Development roles (primary use case):
lead, frontend, backend, fullstack, reviewer, tester, devops, security, data, docs, ai, designer
8 Category Generalist roles (covering broader team types):
marketing, sales, product, project-mgmt, support, game-dev, media, compliance
Init Flow: Start from Base, Refine for Context
backend→ "Node.js/Express Backend Dev with Stripe knowledge")Deliverables
packages/squad-sdk/src/roles/— types, catalog (20 role objects), public APIuseRole('backend', { name: 'kane' })— SDK builder forsquad.config.tssquad roles— CLI command to list/search base rolescast.ts—generateCharter()andpersonalityForRole()use catalogcoordinator.ts— Init mode uses catalog-based matching + refinementsquad.agent.mdtemplatesAttribution
Base role content is inspired by and adapted from msitarzewski/agency-agents (MIT License) by AgentLand Contributors. Proper attribution will be included in source comments, generated charter headers, and documentation.