product
Technical users running Incipit with their own model endpoint — internal/work colleagues, developers, and anyone who has cloned it (the "bring your own model" path: Ollama, LM Studio, llama.cpp, vLLM, OpenAI). They arrive with a rough, half-formed software idea and want a rigorous, paste-ready spec for a coding agent without writing it by hand. They're comfortable with a terminal and an endpoint/model settings panel; they value speed and signal over hand-holding, but a newcomer must still be able to follow the flow on first run.
The job to be done: turn a brain dump into a single structured "mega-prompt"
in minutes — calibrate stakes/form-factor, answer one batch of clarifying
questions (with smart [ASSUMPTION] defaults), draft and refine six spec
sections, optionally run a "round table" review, then copy/download the result.
A BMAD-style mega-prompt wizard. It compresses the elicitation a good spec needs (idea → calibrate → clarify → sections → final) into a fast, mostly hands-off flow, and assembles the final spec deterministically (no LLM call). It exists so that the quality of a coding-agent prompt doesn't depend on the user remembering to ask themselves every question. Success = the user pastes the output into an agent and it has what it needs, on the first try, without a back-and-forth.
Model-agnostic by design: it talks to any OpenAI-compatible endpoint and has an advanced local DiffusionGemma path. No authentication — localhost / trusted network only.
Terminal-native craft — precise, technical, hacker-tool confidence. Three words: precise, transparent, unfussy. The voice is a capable CLI, not a chat-bot: it shows the machine working (live SSE progress, a ❯ shell prompt, monospace activity blocks) rather than hiding behind vague spinners. Some energy and personality are welcome (the violet glow, the "Shoot the Moon" hands-off mode, the round-table "party"); the line is that energy must read as craft, never as marketing.
- Generic AI-SaaS look — gradient text, hero-metric blocks, endless identical icon+heading card grids, decorative glassmorphism. The default AI-startup aesthetic is the thing to actively avoid.
- Corporate / enterprise dashboard — heavy chrome, business-blue palette, data-table-everywhere density, admin-panel coldness.
- Childish / over-friendly — cartoon illustrations, mascots, rounded consumer-app cuteness, emoji as decoration. (Functional emoji as state/wayfinding markers — 🌙 moonshot, 🎉 round table, ✓ applied — are fine; decorative ones aren't.)
(Note: "loud / attention-grabbing" was not ruled out — controlled energy is on-brand, as long as it isn't marketing-landing energy on a working tool.)
- Compression over completeness. The product's whole value is respecting the user's time. Default smart assumptions, never force a blank the user doesn't care about, and keep every step to the minimum that earns its place.
- Show the machine working. Generation is slow and asynchronous; make that legible and even satisfying (terminal progress, heartbeat, the round-table transcript) instead of papering over it. Transparency is the feature.
- Tool, not toy. Assume a technical user: density, keyboard-friendliness, and confidence over tutorialization — while keeping first-run legible to a stranger who just cloned it.
- Unambiguous state. Pending / applying / applied / denied, and finding severity, must be readable at a glance. The user is making approve/deny decisions; never make them guess what state something is in.
- Bring your own everything. Stay provider-agnostic in copy and UI. Never assume a specific model, vendor, or that thinking-mode behaves one way.
Target WCAG 2.1 AA. Dark-theme by design, so body text holds ≥4.5:1 and
large text ≥3:1; keyboard navigation and visible focus on all controls;
prefers-reduced-motion honored for the always-on terminal/heartbeat motion;
live regions (role="status" / aria-live) on the SSE-driven progress and
round-table panels so non-visual users get the async updates. Severity/state is
never carried by color alone — pair it with a label or icon.