Back to search
agent-creator
MarketplaceWorks with Claude CodeGitHub
Build a team of Claude Code agents for any project, with model routing per agent.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Checked today
Badge for your README
Markdown
[](/listings/juan16pcxix--agent-creator)HTML
<a href="/listings/juan16pcxix--agent-creator"><img src="/api/listings/juan16pcxix--agent-creator/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
**As a plugin** (recommended, gets updates):
/plugin marketplace add Juan16pcXIX/agent-creator /plugin install agent-creator@agent-creator
**By hand**: copy `plugins/agent-creator/skills/agent-creator` into `~/.claude/skills/` (all projects) or `.claude/skills/` (one project).
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 14 lines
- {
- "name": "agent-creator",
- "description": "Build a team of Claude Code agents for any project, with model routing per agent.",
- "owner": {
- "name": "Juan16pcXIX"
- },
- "plugins": [
- {
- "name": "agent-creator",
- "source": "./plugins/agent-creator",
- "description": "Turns any project, brand, person or goal into a tree of Claude Code agents, each on the cheapest model that does its job well."
- }
- ]
- }
README.md · README · 113 lines
- # agent-creator
- **Give it a project. Get back a team of Claude Code agents, each on the cheapest model that can do its job well.**
- [Leer en español](README.es.md)
- agent-creator is a skill for [Claude Code](https://code.claude.com). Point it at a project folder, a conversation, or a short description ("a streetwear brand", "I make music", "our logistics paperwork"), and it:
- 1. **Reads the root**: what the project is, what work keeps repeating, and what you already have (your agents, skills and connectors such as Canva or Notion).
- 2. **Scans the field briefly**: how this kind of project is usually run and what is missing.
- 3. **Interviews you**: how expert the agents should be, your budget, where to install them, what language they work in.
- 4. **Draws a tree**: departments (branches) and single-purpose subagents (leaves), with a verifier wherever mistakes are expensive.
- 5. **Routes every agent to a model** with a 10-point rubric: Haiku for mechanical work, Sonnet for most production work, Opus for strategy and high-stakes calls.
- 6. **Lets you pick the team**, department by department: tick the agents you want, keep or upgrade the ones you already have, and add any agent it did not think of ("one that builds the carousels in Canva").
- 7. **Researches each department** at the depth you chose and writes it into knowledge files the agents read.
- 8. **Writes the agent files and an orchestrator skill** (`/your-project`) that routes requests to the right agent, and **validates the result** with a bundled script.
- ```
- norte (/norte)
- ├── marketing ── norte-marketing-lead [sonnet · medium]
- │ ├── strategist [opus · high] 2-4 week content and campaign plan
- │ ├── trend-scout [haiku · medium] this week's trends, with links
- │ ├── carousel-writer [sonnet · medium] one carousel from a brief
- │ └── fact-checker [sonnet · medium] checks claims, prices, dates
- └── finance (no lead: only two leaves)
- ├── bookkeeper [haiku · low] categorizes transaction exports
- └── cashflow-forecaster [opus · medium] 13-week forecast, risky weeks
- ```
- *A real output for a fictional clothing brand. All files are in [examples/clothing-brand](examples/clothing-brand/.claude).*
- ## Why
- Building a team of agents by hand takes hours, and most people end up with one of two bad teams: a few giant agents that do everything badly, or every agent on the most expensive model "to be safe". A Haiku call costs about a twentieth of a Sonnet call. If the mechanical half of your work runs on Haiku and only the strategy runs on Opus, the team does the same work for a fraction of the tokens.
- ## Install
- **As a plugin** (recommended, gets updates):
- ```
- /plugin marketplace add Juan16pcXIX/agent-creator
- /plugin install agent-creator@agent-creator
- ```
- **By hand**: copy `plugins/agent-creator/skills/agent-creator` into `~/.claude/skills/` (all projects) or `.claude/skills/` (one project).
- ## Use
- In Claude Code, inside your project:
- ```
- /agent-creator
- ```
- or just ask: "create a team of agents for this project", "split my brand into departments with agents". You can also describe a root that has no folder yet: "build me agents for my music career: I write, record vocals at home and release a single every 6 weeks."
- When it finishes you get `/your-project`, an orchestrator skill that knows the whole team, plus the agent files in `.claude/agents/`. Run `/agent-creator` again later to add, merge or remove agents: it reads the saved `team.json` and works as a diff.
- ## How models are chosen
- Each agent is scored 0 to 2 on five things: reasoning depth, ambiguity, cost of an error, horizon, and judgment. The total (0 to 10) maps to a model:
- | Score | Model | For |
- |---|---|---|
- | 0-4 | Haiku | Files, formatting, lookups, data entry, checklists |
- | 5-7 | Sonnet | Writing from a brief, coding, research, fact-checking |
- | 8-10 | Opus | Multi-week plans, strategy, money and high-stakes decisions |
- Then it adjusts: your budget moves everything up or down a tier, agents that run many times per request drop a tier, verifiers are never weaker than what they check, and anything an audience reads is at least Sonnet. Fable is used only if you choose "max quality". Full rubric: [model-routing.md](plugins/agent-creator/skills/agent-creator/references/model-routing.md).
- ## How the tree is cut
- - A **leaf** does one recurring job with one output and a checkable "done when". One-off tasks don't get agents.
- - Every department where mistakes cost money, reputation or truth gets a **verifier** that did not produce the work.
- - A **lead** agent exists only when a department has 3+ agents that need planning or sequencing. Otherwise the orchestrator calls them directly, which is cheaper.
- - 3 to 7 departments, 2 to 6 agents each.
- Details and archetypes (business, creator, musician, software, logistics, research, personal goals): [references/](plugins/agent-creator/skills/agent-creator/references).
- ## What's in the repo
- ```
- .claude-plugin/marketplace.json marketplace entry
- plugins/agent-creator/
- .claude-plugin/plugin.json
- skills/agent-creator/
- SKILL.md the 8 phases
- references/ decomposition, model routing, interview, research, archetypes
- templates/ leaf agent, lead agent, orchestrator skill, team map, team.json schema
- scripts/validate_team.py checks a generated team (stdlib Python)
- examples/
- clothing-brand/ a full generated team that passes validation
- musician/ a team map, in Spanish
- ```
- Check a generated team yourself:
- ```
- python plugins/agent-creator/skills/agent-creator/scripts/validate_team.py examples/clothing-brand/.claude --team norte
- ```
- ## Roadmap
- - [ ] More examples (software startup, logistics office, YouTube channel)
- - [ ] Cost estimate per typical request, measured on real runs
- - [ ] Evals: a fixed set of project descriptions and a rubric for the trees they should produce
- - [ ] Ready-made team packs per industry
- Ideas and pull requests are welcome. Open an issue with the kind of project you tried it on and the tree you got.
- ## License
- [MIT](LICENSE)
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.