A Claude skill that turns a vague software idea into a developer-ready specification — in one conversation.
Runs a structured 25-question adaptive interview with a non-technical client, silently calibrates to their technical literacy, and produces five production-ready outputs:
- Technical Specification — markdown doc with screen inventory + user flow map
- Implementation Prompt — paste-ready LLM code-gen prompt
- React/HTML UI Mockup — working visual skeleton of the primary user flow
- Tool Recommendations — personalised Claude ecosystem setup
- Project Style — custom Claude style tuned to the project's communication needs
Domains: Web apps / SaaS · Mobile · Automation & bots · Data pipelines & dashboards
SKILL.md # Main skill — execution protocol, card rendering, UX rules
references/
domain-probes.md # Domain-specific deep-dive questions (e-commerce, SaaS, etc.)
stack-defaults.md # Tech stack inference rules by domain + tier
security-probes.md # Compliance flags, risk pattern library (HIPAA, PCI, SOC 2)
ui-patterns.md # Design reference mappings (Notion, Linear, Stripe, etc.)
spec-template.md # Output 1 — Technical Specification template
impl-prompt-template.md # Output 2 — Implementation Prompt template
mockup-rules.md # Output 3 — React/HTML Mockup rendering rules
tool-recommendations.md # Output 4 — Claude tool recommendation logic
style-template.md # Output 5 — Project Style generation template
cache-strategy.md # Developer reference — API caching for programmatic use
Install as a skill via your Claude Desktop config. The skill auto-triggers on phrases like:
- "I want to build something"
- "help me scope this"
- "my client needs a build plan"
- "I have an app idea"
Paste the contents of SKILL.md as a system prompt. Load reference files on demand as the interview progresses (the skill's Reference Files table tells you when to load each one).
| Principle | In practice |
|---|---|
| Infer, don't interrogate | Never ask "what stack?" — earn the recommendation |
| Speak their language | 4 vocabulary tiers, silently applied |
| One idea per question | Multi-part questions = multi-part confusion |
| Gate the output | Nothing generates until scope is locked |
| Summaries build trust | Every round ends with a client-correctable reflection |
Oleg Strutsovski — cyrillical00