Skip to content

jeremyary/agent-scaffold

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

45 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Agent Scaffold

A template repository that provides a complete Claude Code multi-agent system for software development -- from product discovery through deployment. It ships pre-configured with 17 specialist agents, 16 convention rules, 8 skills, 2 git hooks, and a shared permissions baseline so your team can start building with AI-assisted workflows immediately.

This README describes the template itself. Once you've set up your project using /setup or START_HERE.md, replace this file with your project's own README.


How It Works

The scaffold turns Claude Code into an AI development team. Each agent has a defined role, model tier, tool permissions, and behavioral rules. The main Claude Code session acts as an orchestrator, routing your requests to the right specialist based on a keyword-matching decision matrix.

You describe what you need
        |
   Main session (orchestrator)
        |
   Routes to specialist agent(s)
        |
   Agent reads project rules + relevant code
        |
   Agent produces output (plan, code, review, etc.)

Agents don't operate in a vacuum. Convention rules constrain how they write code, structure errors, handle secrets, and format commits. Skills provide reusable workflow templates. Hooks automate compliance tasks at commit time.

Agents

17 specialist agents, each defined in .claude/agents/. Agents are invoked with @agent-name in your prompt.

Planning and Strategy

Agent Model Description
Product Manager Opus Product discovery, PRDs, roadmaps, feature prioritization, persona definition
Architect Opus System design, ADRs, technology trade-offs, high-level structure
Tech Lead Opus Feature-level technical designs, interface contracts, implementation approach
Requirements Analyst Sonnet Requirements gathering, user stories, acceptance criteria

Implementation

Agent Model Description
Backend Developer Sonnet Server-side code, API handlers, business logic, middleware
Frontend Developer Sonnet UI components, client-side state, accessibility, responsive design
Database Engineer Sonnet Schema design, migrations, query optimization, data integrity
API Designer Sonnet OpenAPI specs, API contracts, REST/GraphQL interface design

Quality and Security

Agent Model Description
Code Reviewer Opus Code quality analysis, standards adherence, best practices (read-only)
Test Engineer Sonnet Test strategy, test authoring, coverage analysis
Security Engineer Sonnet Vulnerability analysis, threat modeling, security audit (read-only)

Operations and Support

Agent Model Description
Project Manager Sonnet Epic/story breakdown, work estimation, Jira/Linear/GitHub exports
DevOps Engineer Sonnet CI/CD pipelines, Docker/K8s configs, infrastructure as code
SRE Engineer Sonnet SLO/SLI definitions, runbooks, alerting, incident response
Performance Engineer Sonnet Profiling, bottleneck identification, optimization
Debug Specialist Sonnet Structured root cause analysis, systematic bug diagnosis
Technical Writer Sonnet Documentation, changelogs, API docs

The default model assignment uses an expanded hybrid strategy: Opus for planning and review agents (where errors cascade), Sonnet for implementation agents (where well-defined contracts make Opus unnecessary). This is configurable per-agent.

Nine agents have cross-session memory enabled, so they learn your project's patterns, decisions, and preferences over time.

Convention Rules

15 rule files in .claude/rules/. Rules are loaded automatically -- 11 globally via CLAUDE.md imports, and 4 scoped to specific file paths (API, UI, database code, and Python files).

Rule Scope What It Governs
code-style.md Global TypeScript + Python naming, formatting, imports, component patterns
python-style.md **/*.py Python-only style alternative (for projects without a frontend)
git-workflow.md Global Branch naming, conventional commits, AI assistance trailers
testing.md Global Coverage targets, test naming, AAA pattern, isolation rules
security.md Global Secrets handling, input validation, auth, transport, dependencies
error-handling.md Global RFC 7807 error responses, HTTP status codes, error class hierarchy
observability.md Global Structured logging, log levels, correlation IDs, health checks, metrics
api-conventions.md Global REST resource design, pagination, filtering, versioning, rate limiting
ai-compliance.md Global Human-in-the-loop, AI marking, sensitive data prohibition, licensing
agent-workflow.md Global Task chunking limits (3-5 files, 5-7 steps), context engineering
review-governance.md Global Plan-review-first, anti-rubber-stamping, PR size limits, two-agent review
architecture.md Global Monorepo structure, package managers, inter-package dependencies
api-development.md packages/api/** Backend framework patterns (FastAPI by default)
ui-development.md packages/ui/** Frontend framework patterns (React/TanStack by default)
database-development.md packages/db/** ORM patterns, migration conventions (SQLAlchemy/Alembic by default)

Rules are meant to be customized. The scaffold ships with opinionated defaults (TypeScript + Python, monorepo, REST APIs) that you adjust to match your stack during setup.

Skills (Slash Commands)

8 skills in .claude/skills/. User-invocable skills are used as slash commands in Claude Code.

Command Description
/setup Interactive project setup wizard -- walks through every configuration step
/review Combined code quality + security review on the current branch
/status Project health check -- lint, type check, tests, dependency audit
/adr Create an Architecture Decision Record interactively

Additional skills are referenced internally by agents (not invoked directly):

Skill Used By
project-conventions All agents -- technology stack details, directory layout, environment config
workflow-patterns Orchestrator -- sequencing templates for SDD, bug fixes, refactoring, etc.
pm-exports Project Manager -- export templates for Jira, Linear, GitHub Projects
sre-templates SRE Engineer -- SLO definitions, runbooks, alerting rules, capacity plans

Hooks

2 git hooks in .claude/hooks/.

Hook Purpose
prepare-commit-msg Appends Assisted-by: Claude Code trailer to AI-assisted commits automatically
sensitive-data-check.sh Scans for credentials, PII, internal hostnames, and API keys before tool execution

The commit hook activates via symlink or git config core.hooksPath. The sensitive data check is opt-in -- enable it in settings.local.json (personal) or settings.json (team-wide).

Permissions

.claude/settings.json provides a shared permissions baseline committed to git:

  • Pre-approved: Git read commands, test runners, linters, type checkers, formatters, Docker build/compose, Terraform plan/validate, file operations (mkdir, cp, mv, touch)
  • Blocked: rm -rf, force push, git reset --hard, chmod 777, pipe-to-bash, reading .env files
  • Intentionally excluded: Package install commands (npm install, pip install, uv add) -- agents must ask before adding dependencies

A .claude/settings.local.json.template provides a starting point for personal overrides (WebFetch domains, org-specific tools). Copy it to settings.local.json (which is gitignored) and customize.

Orchestration

The main Claude Code session routes requests using a decision matrix in .claude/CLAUDE.md. For complex, multi-step work, it sequences agents using patterns defined in the workflow-patterns skill:

  • Sequential chain -- tasks in strict order, each feeding the next
  • Parallel fan-out -- independent tasks run concurrently, then synchronize
  • Review gate -- implementation followed by mandatory review before proceeding
  • Iterative loop -- profile, fix, verify, repeat until targets are met

The recommended lifecycle for non-trivial features is Spec-Driven Development (SDD): product plan, architecture, requirements, technical design, task breakdown, implementation, review -- with review gates between phases. Quick tasks (bug fixes, single-file changes) skip the ceremony and go directly to the relevant specialist.

Getting Started

Option 1 -- Interactive setup (recommended):

  1. Use this template to create a new repository (or copy the directory)
  2. Open the project in Claude Code
  3. Run /setup -- the wizard asks questions and configures everything

Option 2 -- Manual setup:

  1. Use this template to create a new repository (or copy the directory)
  2. Follow START_HERE.md -- 11 steps covering project identity, tech stack, style rules, permissions, agents, domain rules, hooks, and secrets

Both paths produce the same result. The wizard is faster; the manual guide gives you more context about what each setting does and why.

Key Design Decisions

Why 17 agents instead of one general-purpose agent? Specialization reduces hallucination. A backend developer agent loaded with API conventions produces better code than a generalist that has to context-switch between UI patterns and database migrations. Each agent loads only the rules relevant to its domain.

Why Opus for planning, Sonnet for implementation? Errors in planning cascade through everything downstream. A flawed architecture decision costs far more to fix than a flawed function implementation. Opus handles the high-leverage decisions; Sonnet handles execution within well-defined contracts.

Why so many rules? Rules are cheap context (short files, always relevant) that prevent the most common AI code generation failures: inconsistent style, missing error handling, rubber-stamped reviews, over-scoped tasks, leaked secrets. They encode team knowledge that would otherwise be re-explained every session.

Why task chunking limits? At 95% per-step reliability, a 20-step autonomous chain succeeds only ~36% of the time. Keeping tasks to 5-7 steps and 3-5 files resets the error chain and keeps compound failure probability manageable.

Why a repo template instead of a Claude Code plugin? Plugins cannot ship arbitrary files like git hooks, settings templates, or path-scoped rules -- they only support a subset of what the scaffold provides. A repo template gives you the full file tree on day one with no bootstrap step required.

Customization

Everything in this scaffold is meant to be modified. During setup you will:

  • Set your project's maturity level (PoC, MVP, Production) which governs how much rigor agents apply
  • Choose your tech stack and adjust rules to match
  • Remove agents you don't need (no frontend? delete frontend-developer.md)
  • Add domain-specific rules for your industry (healthcare, fintech, etc.)
  • Configure which model tier each agent uses based on your budget
  • Set up permissions for your specific toolchain

See START_HERE.md for the full walkthrough, or run /setup for the interactive version.

Documentation

Document Purpose
START_HERE.md Complete setup guide (11 steps)
CLAUDE.md Project configuration read by all agents
.claude/CLAUDE.md Agent routing matrix and orchestration patterns
docs/ai-native-team-playbook.md Team practices for AI-assisted development
docs/ai-compliance-checklist.md Developer quick-reference for AI compliance

About

provides agent scaffolding for an environment resembling enterprise software development - from proof-of-concept to production-ready

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages