project-initiator
From an idea to an evidence-backed project plan and a repository an AI agent can start building in right away.
You describe the idea, the skill asks what matters, checks today's market for every part of the project and picks technologies by your priorities — not by what AI always recommends. At the end it sets up a repository with the plan, a backlog and instructions for the agent.
- 6
- steps from idea to repository
- 3–5
- candidates for every layer
- 3
- stack variants: Top, Balanced, Budget
- 4
- agent autonomy levels — you choose
Same question, different result
Ask an ordinary chat "which stack should I use" and you get a fast but generic answer. The skill works like a consultant: it asks, checks and backs things up.
Ordinary AI answer
project-initiator
No questions, straight to a recommendation
A short interview, every question with a suggested answer — "ok" is enough
Goes straight to how to build it
First checks whether it already exists — buy, adapt or build
The same popular tools every time
3–5 candidates per layer, found by searching
Prices from memory, often outdated
Prices and limits checked during the run, with source and date
Components picked without checking they work together
Integrations, regions, lock-in and vendor count checked
Security left for "later"
Secrets, dependencies, access and spending caps handled from day one
An answer in a chat that scrolls away
A repository with the plan, decisions and a backlog
From idea to repository in six steps
Every step has a clear output — a document that stays in the project and feeds both the next step and the agent's future work.
- 01
Interview
Core questions including the purpose — for yourself, for a company or for sale — then room for your preferences and follow-ups only for what the project actually has. Tools you already use stay if they fit. Anything still unclear becomes a written assumption.
ProducesPLAN.md - 02
Market research
First it checks whether a solution already exists and, for a commercial project, who the competition is. Then for every layer — app, data, sign-in, payments, AI, hosting, security — it searches for current candidates. Every claim has a source and a date.
ProducesLANDSCAPE.mdSOURCES.md - 03
Decision
Weighted scoring by your priorities, visible in a table. Top, Balanced and Budget variants. When public data can't decide — AI model quality, say — it doesn't guess but turns the choice into a measured task.
ProducesDECISIONS.md - 04
Package and risk check
Checks the components actually work together: official integrations, known issues, data regions, transfer fees. It classifies the project by risk and writes a lite threat model.
ProducesSTACK-FIT.mdARCHITECTURE.mdTHREAT-MODEL.md - 05
Costs and hardware
Monthly costs during build, launch and growth, including where the free tier ends. What to buy, what to rent and what you don't need at all.
ProducesCOSTS.mdHARDWARE.mdRISKS.md - 06
Project bootstrap
A folder with GitHub structure: agent instructions with industry rules (privacy law, payments, app stores…) and the autonomy level you choose, security rules, CI, a backlog with the first tasks, a changelog and ADRs. How much documentation you get depends on the profile you pick.
ProducesAGENTS.mdSECURITY.mdSETUP.mdQUEUE.md
A popular tool has to win, not just come to mind first
Language models recommend the same tools regardless of the project. So the skill first writes down its first guess, and research then has to confirm or overturn it — by comparing it with alternatives against your priorities. Tools you already use don't have to compete: they just get checked against the project's needs.
01 · First guess
"I'd recommend this even without research."
02 · Compared with ≥ 2 alternatives
weighted by your priorities
Confirmed — with a documented reason
Overturned — a better pick wins
Less guessing, more evidence
Decisions you can defend
Every choice comes with sources, a comparison and one line on why not the others. Useful when a boss, an investor or a developer asks.
No surprises on the bill
Costs over three horizons and the exact points where a free tier ends and the bill jumps.
The agent knows what to do
AGENTS.md with the stack, security rules, the autonomy level you chose and the lines the agent must not cross without you. The backlog starts at T001.
In your language
The interview and documents in the language you write in. The repository can be in another — English for an international team, say.
A repository an agent can start in right away
The result isn't just an answer in a chat but a folder with a GitHub structure. Open it in Claude Code, type "start with T001" and the agent knows what to build, with what, and by which rules. You choose how much documentation you get — a minimal, balanced or complete profile — with the token cost of each explained.
Three levels of depth
Quick — one round of questions, 2–3 candidates on the core layers, other picks flagged for confirmation, ~15–20 searches
Standard — both rounds, 3–5 candidates per layer, package check, ~30–50 searches
Deep — plus every integration verified and all variants costed, 80+ searches
- my-project/
- AGENTS.md
- CLAUDE.md
- README.md
- CHANGELOG.md
- SECURITY.md
- .github/
- docs/
- blueprint/
- PLAN.md · LANDSCAPE.md · DECISIONS.md
- STACK-FIT.md · ARCHITECTURE.md · COSTS.md
- HARDWARE.md · SETUP.md · RISKS.md
- SOURCES.md (+ THREAT-MODEL, OPERATIONS)
- backlog/
- adr/
Two commands in Claude Code
The skill installs as a plugin. It needs web search — without it, it says so up front and marks every price and version as unverified.
> /plugin marketplace add hypodai/project-initiator> /plugin install project-initiator@project-initiator
Tried on real prompts
- Fitness app (Czech)A non-coder, iOS and Android, smartwatch sync, subscriptions. The whole repository and the conversation.
- AI CV feedback (English)A Next.js developer, an MVP in two weeks. The skill kept their stack and left the AI model to a measured test.
- Compared with plain AIThe same prompts with and without the skill — what it added, what it cost and where it fell short.
The examples were made at quick depth with simulated "ok" answers; prices were checked on 16 Oct 2026. The repository is public, version 0.1.0.
Then just describe your idea
Got it: a mobile workout app that syncs with smartwatches. A few questions — each has a suggested answer, "ok" is enough.
- 1. Which phones?[Default: iOS and Android]
- 2. Which watches?[Default: Apple Watch and Wear OS via HealthKit / Health Connect]
…
Sample start of the conversation
It answers and writes the documents in the language you write in. Want the repository in a different language? Just say so.
Rather go through it with a person?
The skill gives you the evidence, but the decisions are yours. If you want to talk an idea through, have an MVP built or set up agents for your team, get in touch.
Let's bring your idea to life
An MVP, a website, SaaS, reporting or an AI experiment.
It doesn't have to be a finished brief. An idea you're not sure about works too — the best things get built when both sides actually care.