| name | project-agent-scaffold |
| description | Methodology: Project-scoped agent generation via tech-stack detection.
EN: Detects project tech stack from package manifests (package.json, Cargo.toml, go.mod, pyproject.toml, pom.xml, mix.exs, Gemfile, composer.json, Package.swift, pubspec.yaml) + secondary signals (tailwind.config.*, next.config.*, nuxt.config.*, docker-compose.yml). Classifies signals into a stack profile, proposes 1–3 project-scoped agents (project:<slug>-pro) baked from agent-template.md with stack-specific patterns. Asks approval per proposal before writing any file. Does NOT auto-create — user always confirms. Writes agent files to .claude/agents/ on approval and suggests project-agent-matrix.yaml entries for dispatch overrides.
RU: Определяет технологический стек проекта по package-манифестам + вторичным сигналам. Классифицирует в stack profile, предлагает 1–3 project-scoped агента (project:<slug>-pro) на базе agent-template.md. Спрашивает подтверждения для каждого предложения. Записывает только после согласия пользователя. НЕ создаёт автоматически.
Triggers: "scaffold project agent", "create project agent", "detect project stack", "project-specific agent", "agent for my stack", "создай agent для проекта", "agent под мой стек", "agent под мой проект", "/project-agent-scaffold"
|
| origin | forgeplan |
| disable-model-invocation | true |
| allowed-tools | Read Write Edit Bash(ls *) Bash(cat *) Bash(test *) Bash(find *) Bash(grep *) |
project-agent-scaffold — tech-stack-aware agent generation
Detects what the project is built with, proposes 1–3 tailored agents, asks
your approval for each, then writes the file. Nothing is written without
an explicit "yes".
When to use
- You just cloned a new project and want a project-scoped Claude Code agent pre-loaded with its conventions
- The canonical
agents-domain or agents-pro agents exist but need project-specific context baked in
- You want a proposal before manually writing
.claude/agents/project-*.md from scratch
When NOT to use
- The project already has project-scoped agents in
.claude/agents/ — audit with agent-advisor first
- You need a general-purpose specialist (no project context needed) — use
agent-advisor directly
- You want the agent to be forgeplan-aware (produce EVIDENCE, claim/release, etc.) — use
AGENT-AUTHORING-GUIDE.md for manual authoring; this skill generates non-forgeplan project helpers
Step 1 — Detect tech stack
Read manifests in priority order. Stop after first match per category; collect secondary signals from any depth.
Primary manifests (one match per runtime)
| File | Runtime | Key fields to read |
|---|
package.json | Node / JS / TS | dependencies, devDependencies, scripts |
Cargo.toml | Rust | [dependencies] section |
go.mod | Go | require lines |
pyproject.toml | Python | [project.dependencies] or [tool.poetry.dependencies] |
requirements.txt | Python (alt) | package names, one per line |
pom.xml | Java / Kotlin | <groupId> + <artifactId> under <dependencies> |
build.gradle / build.gradle.kts | Java / Kotlin | dependencies { ... } block |
mix.exs | Elixir | deps list |
Gemfile | Ruby | gem "..." lines |
composer.json | PHP | require and require-dev keys |
Package.swift | Swift | .package(url:...) entries |
pubspec.yaml | Dart / Flutter | dependencies: block |
Secondary signals (layer on top of primary)
| File / pattern | Signal it adds |
|---|
tailwind.config.* | Tailwind CSS |
next.config.* | Next.js (App Router if appDir: true or /app dir present) |
nuxt.config.* | Nuxt 3 |
vite.config.* | Vite bundler |
docker-compose.yml | Multi-service (informs agent scope) |
prisma/schema.prisma | Prisma ORM |
drizzle.config.* | Drizzle ORM |
supabase/ dir | Supabase backend |
convex/ dir | Convex backend |
*.tf files | Terraform IaC |
k8s/ or charts/ dir | Kubernetes |
Detection commands (run these)
test -f package.json && cat package.json
test -f Cargo.toml && cat Cargo.toml
test -f go.mod && cat go.mod
test -f pyproject.toml && cat pyproject.toml
test -f requirements.txt && cat requirements.txt
test -f mix.exs && cat mix.exs
test -f composer.json && cat composer.json
test -f pubspec.yaml && cat pubspec.yaml
find . -maxdepth 2 -name "next.config.*" -o -name "tailwind.config.*" -o -name "nuxt.config.*" 2>/dev/null
find . -maxdepth 3 -name "prisma" -type d -o -name "drizzle.config.*" 2>/dev/null
ls .claude/agents/ 2>/dev/null || echo "(no project agents yet)"
Step 2 — Classify stack profile
Combine primary manifest + secondary signals into a human-readable profile string and derive a slug.
| Example profile | Slug | Agent name |
|---|
| Next.js + Tailwind + Prisma + PostgreSQL | nextjs-tailwind-prisma | project:nextjs-tailwind-prisma-pro |
| Next.js + Tailwind + Drizzle + Supabase | nextjs-drizzle-supabase | project:nextjs-drizzle-supabase-pro |
| Nuxt 3 + Tailwind + Prisma | nuxt-tailwind-prisma | project:nuxt-tailwind-prisma-pro |
| React + Vite + Express + PostgreSQL | react-express-pg | project:react-express-pg-pro |
| NestJS + Prisma + PostgreSQL | nestjs-prisma | project:nestjs-prisma-pro |
| Rust + Tokio + Axum | rust-axum | project:rust-axum-pro |
| Rust + Axum + Diesel + PostgreSQL | rust-axum-diesel | project:rust-axum-diesel-pro |
| Go + Gin + GORM | go-gin-gorm | project:go-gin-gorm-pro |
| Go + Echo + gRPC | go-echo-grpc | project:go-echo-grpc-pro |
| Python + FastAPI + SQLAlchemy | python-fastapi-sqlalchemy | project:python-fastapi-sqlalchemy-pro |
| Python + FastAPI + ML (torch/transformers) | python-fastapi-ml | project:python-fastapi-ml-pro |
| Django + PostgreSQL | django-pg | project:django-pg-pro |
| Elixir + Phoenix | elixir-phoenix | project:elixir-phoenix-pro |
| PHP + Laravel | php-laravel | project:php-laravel-pro |
| PHP + Symfony | php-symfony | project:php-symfony-pro |
| Dart + Flutter | dart-flutter | project:dart-flutter-pro |
| Monorepo (mixed runtimes) | monorepo-<primary> | up to 3 agents, one per sub-project |
Slug rule: 2–4 components, kebab-case, most discriminating first, omit obvious (js, ts, web).
Base agent mapping
| Stack | Closest marketplace base | If no marketplace base |
|---|
| Next.js | agents-domain:nextjs-developer | — |
| Nuxt / Vue | agents-domain:frontend-developer | — |
| React + Vite | agents-domain:frontend-developer | — |
| NestJS | agents-domain:nodejs-developer | — |
| Rust | none in marketplace | Import from VoltAgent/awesome-claude-code-subagents |
| Go | none in marketplace | Import from VoltAgent/awesome-claude-code-subagents |
| Python / FastAPI | agents-pro:ml-developer (ML-heavy) or generic | — |
| Django / Flask | none specific | Generic project agent |
| Java / Spring | none specific | Generic project agent |
| Elixir / Phoenix | none specific | Generic project agent |
| PHP / Laravel | none specific | Generic project agent |
| Dart / Flutter | none specific | Generic project agent |
Step 3 — Generate proposals (one per candidate)
For each candidate (max 3 total), render a proposal block:
PROPOSAL: project:<slug>-pro
Base: <marketplace base or "none — generic template">
Stack: <human profile string>
Project-specific patterns to bake in:
- <pattern 1 from stack, e.g. "Next.js App Router conventions (server components, use client boundary)">
- <pattern 2, e.g. "Prisma migration workflow (npx prisma migrate dev)">
- <pattern 3, e.g. "Tailwind v4 config (CSS-first, no tailwind.config.js)">
Target file: .claude/agents/project-<slug>-pro.md
matrix.yaml: project:<slug>-pro → <phase> override (optional)
Approve this proposal? [yes / modify / skip]
<<NEED_USER_INPUT: approve / modify / skip for project:<slug>-pro>>
Present all proposals before writing any file.
Step 4 — On user response
| Response | Action |
|---|
yes / approve | Write agent file (Step 5). Record in final report. |
modify | Ask what to change (name, scope, base, patterns). Re-render proposal. Loop. |
skip / no | Note as skipped. Continue to next proposal. |
Step 5 — Write agent file (on approval only)
Use plugins/fpl-skills/templates/agent-template.md as the skeleton. Replace every TODO with:
| TODO placeholder | Replacement |
|---|
TODO-kebab-name | project-<slug>-pro |
TODO-role | project-specific <stack> developer |
TODO-one-line-scope | short scope sentence (e.g. "You handle all Next.js App Router + Prisma work in this repo.") |
EN: TODO in description | project-specific description with stack patterns |
RU: TODO in description | Russian translation |
| Triggers | stack-specific natural-language triggers |
model: | sonnet (default for project agents — adjust to opus if reasoning-heavy) |
color: | pick a stable hex (see palette below) |
tools: | Read, Grep, Glob, Write, Edit, Bash — NO forgeplan MCP (project agents are Profile C-coder) |
| HARD RULES section | project-specific invariants (e.g. "Never run migrations without user confirmation") |
Color palette (pick one per agent, avoid duplication with existing agents):
| Stack | Suggested hex |
|---|
| Next.js / React | #0070F3 |
| Rust | #CE422B |
| Go | #00ADD8 |
| Python | #3572A5 |
| Elixir | #6E4A7E |
| PHP | #777BB4 |
| Flutter | #54C5F8 |
| Java / Kotlin | #F89820 |
| Generic | #546E7A |
Write to: .claude/agents/project-<slug>-pro.md
After writing, suggest the project-agent-matrix.yaml entry:
overrides:
- phase: build
stack_signal: "<primary manifest filename>"
agent: "project:<slug>-pro"
Step 6 — Final report
After all proposals are resolved:
project-agent-scaffold complete
written: <N> agent(s)
- .claude/agents/project-<slug-1>-pro.md
- .claude/agents/project-<slug-2>-pro.md [if any]
skipped: <M> proposal(s) — <reason if given>
next: Run `forgeplan list` to verify project-agent-matrix.yaml entries.
Run `/agent-advisor` to confirm dispatch mapping for new agents.
Anti-patterns
- Do not write files without explicit approval. Always present the full proposal block and wait for
<<NEED_USER_INPUT:>> to resolve. Auto-writing bypasses the user's security boundary.
- Do not propose more than 3 agents. Cap at 1 per major runtime. If a monorepo has 5 sub-projects, pick the top 3 by LOC or user priority.
- Do not depend on
agents-domain availability. The base mapping is advisory. If the user's installation lacks agents-domain, generate from the generic template anyway.
- Do not add forgeplan MCP tools to proposed agents. Project-scoped agents are Profile C-coder (source files only). If the project needs forgeplan-aware project agents, direct to
AGENT-AUTHORING-GUIDE.md.
- Do not invent stack signals not present in the files. If
package.json has no Prisma entry and no prisma/ dir, do not assume Prisma.
Related skills
