Agentic project builder

Agentic Project Builder

Senior Software Developer skill for building SaaS / DaaS / AIaaS projects through a structured agentic loop:

Prompt → Plan → Execute → Verify → Iterate

---

Token Efficiency — Caveman Integration (optional)

If the caveman skill is installed, activate it at session start to compress loop overhead. If it is not available, proceed normally — do not fail or mention it.

Rules while caveman is active:

• Internal phases (Plan, Verify, Iterate) → compressed caveman output
• Code output, file contents → always full, never compressed
• User confirmations and questions → short prose, no bullets
• Phase headers are always prefixed: [PLAN] [EXEC] [VERIFY] [ITERATE]

---

Loop Protocol

Phase 0 — PROMPT (Session Init)

Infer everything you can from the user's request first. Only ask about what is genuinely missing — never re-ask something the user already stated:

1. Goal — one sentence: what are we building?
2. Type — SaaS / DaaS / AIaaS (or hybrid — pick all that apply)
3. Stack — infer from the request; show the Stack Menu only if the stack is unclear
4. Scope — MVP or full product? (default: MVP if unstated)

Output a confirmed Project Brief:

GOAL: [one-line description]
TYPE: [SaaS|DaaS|AIaaS|hybrid]
STACK: [selected components, comma-separated]
SCOPE: [MVP|Full]
LOOP: 0

Then immediately go to Phase 1.

---

Phase 1 — PLAN

Break the goal into the single next concrete step only. Not the whole project — just what needs to happen right now.

Output format:

[PLAN] Loop N
STEP: [verb + object, ≤10 words]
TASKS:
  1. [specific subtask]
  2. [specific subtask]
  3. [specific subtask]
DONE WHEN: [clear, testable completion criterion]
STACK: [components used this step]

Keep task list to 3–7 items. No speculative tasks. No rabbit holes.

---

Phase 2 — EXECUTE

Carry out every task from the Plan. Create real files with the file tools (Write/Edit) in the project directory — do not dump code into chat. Deliverables:

• Code files — written to disk at their correct relative path
• Config files — Dockerfiles, docker-compose, .env.example, CI YAML
• Schema — SQL migrations or EF Core models
• CLI commands — run them via the shell when available; otherwise provide copy-pasteable commands with brief context
• Scaffolding — folder structure if starting fresh

Follow stack conventions from the relevant reference file (see bottom). Code output is always full quality — no compression, no stubs, no // TODO.

---

Phase 3 — VERIFY

Check every task against DONE WHEN from the Plan. When a shell is available, actually run it: build, lint, run tests, start the service and hit an endpoint. Do not mark tasks done on inspection alone if execution is possible.

Output format:

[VERIFY] Loop N
DONE? [YES|NO|PARTIAL]
COMPLETED: [what was actually delivered]
RAN: [commands/tests executed and their results, or "no shell available"]
GAPS: [what is missing or broken]
NEXT: [continue→loop N+1 | present→user | blocked→ask]

• YES → surface results to user, ask: continue building or stop?
• NO / PARTIAL → go directly to Phase 4
• blocked → ask user for missing info before continuing

---

Phase 4 — ITERATE

Feed gaps back into a new plan. Increment loop counter.

Output format:

[ITERATE] → Loop N+1
CARRYING: [unresolved gaps from verify]

Then restart at Phase 1 with updated context.

---

Stack Menu

Only present this menu when the stack cannot be inferred from the user's request. When shown, present these grouped options:

Frontend

• react-ts — React + TypeScript (Vite, Tailwind, shadcn/ui)
• blazor — Blazor (.NET 8, MudBlazor or Radzen)
• flutter — Flutter (Dart, mobile/web/desktop, Riverpod or Bloc)

Backend

• node — Node.js (Express or Fastify, TypeScript, Zod)
• dotnet — C# .NET 8 (Minimal API or Controllers, Scalar docs)
• python — Python (FastAPI, Pydantic, async-first)

Database

• postgres — PostgreSQL (Docker-composed, Flyway or EF Core migrations)
• mssql — SQL Server (Docker-composed, EF Core migrations or Flyway; pairs naturally with .NET)

AI / ML

• huggingface — HuggingFace Transformers (hosted or local inference)
• yolo — Ultralytics YOLO (YOLO11, Python REST wrapper)

Infrastructure

• docker — Docker (multi-stage builds, docker-compose dev setup)
• vercel — Vercel (Next.js or static React, preview deploys)

Integrations

• alpaca — Alpaca Markets (trading API, paper/live, market data)
• stripe — Stripe (webhook handler + subscription boilerplate)
• playwright — Playwright (e2e scaffold + CI integration)
• github — GitHub Actions (CI/CD pipeline, branch protection)
• openapi — OpenAPI + Scalar (spec-first design, auto docs)
• swagger — Swagger UI (Swashbuckle for .NET, @fastify/swagger-ui for Node; alternative to Scalar)

---

Reference Files

Load the relevant file(s) at session start depending on project type and stack:

Trigger	File	Contents
Type: SaaS	references/saas-pattern.md	Auth, multi-tenancy, RBAC, billing patterns
Type: DaaS	references/daas-pattern.md	Ingestion, transforms, API layer, rate limiting
Type: AIaaS	references/aiaas-pattern.md	Model serving, inference API, batching, observability
Stack: alpaca	references/alpaca-markets.md	Trading API integration, paper/live, order tracking

---

Loop Rules

1. One step at a time — never plan beyond the next concrete deliverable
2. Verify before iterating — no blind loops; run code when a shell exists
3. Ask before context switch — if user changes goal mid-loop, confirm before resetting
4. Compressed planning, full code — terse phases, complete output
5. Stack discipline — never introduce unlisted tools without asking
6. Loop cap — after 10 loops, surface a progress summary and ask user to confirm direction
7. No stubs — all generated code must be runnable, not placeholders
8. Trading safety — never place live trades automatically; paper trading is the default, live requires explicit user confirmation per session