Prompt Forge

Every time a non-developer asks "how should I actually use AI at work?" the same explanation gets repeated and hands stay still. Books are heavy; lectures are one-way and rarely match the asker's own scenario.

Internal use for a grad-school cohort (SKKU EMBA). Inside the same class, tool familiarity varies a lot, and "when do I pick Claude vs Gemini vs GPT?" keeps coming back as a question.

Non-developers who want to bring AI into their work but don't know where to start, and the mentor who'd like to sit beside them and shape the prompts together.

Lock three shared principles in first - "AI drafts, humans verify, no sensitive data" - then layer per-tool guides, a five-step roadmap, scenario courses, and a hands-on IDE on top, and "non-developers bringing AI into their work" stops being aspirational.

• AI usage guide - 12 scenarios, all built on the same three principles ("AI is the drafter, humans verify, no sensitive data")
• AI combo cheatsheet - "when you want X, pair these tools" written as 17 collapsible combo cards
• Claude 5-step roadmap - from install to first automation, laid out as a stepper one step per page
• Gemini usage guide - sits alongside Claude as a complement, with one comparison table for "when Gemini, when Claude"
• Courses - EMBA peers' on-site classes packaged as online courses for async catch-up
• Prompt sandbox - model selector, system-prompt presets (e.g. "polish a Korean workplace email tone"), and immediate chat-based practice
• Phase split (content track and practice track in parallel) - grading and answer sharing called out explicitly as the next milestone

• The three shared principles are pinned at the top of every page - the tool changes, the posture doesn't
• Instead of separate guides per tool, the comparison is what gets surfaced - the learning goal is "the right tool for the task," not tool tribalism
• Sandbox starts from workplace presets ("polish an email tone") rather than a blank textbox - so the first session has somewhere to land
• Login page surfaces "first request can take up to 30 seconds" - the cold-start wakeup reads as expected behavior, not a broken site

Internal use. Paired with six on-site EMBA peer classes; user numbers not disclosed. Phase 1 (auth, content, design, billing infra) is shipped; Phase 2 (practice IDE, grading, answer sharing) is in progress - the IDE is done, grading and sharing are next.

With the content track (guides, roadmap, courses) shipped, the platform is moving from "valuable just by reading" to "valuable by trying." The next round is grading + answer sharing, turning "solo learning" into "learning while comparing."

Used inside the cohort. External expansion is on hold - the space is narrow and crowded.

A learning tool isn't a write-once piece - every new content card has to apply the same principles for trust to hold. Surfacing "AI drafts / humans verify / no sensitive data" on every page enforces that regression check by design, so it doesn't have to live in human hands.

• FastAPI
• Python
• SQLAlchemy 2.x (async)
• Alembic
• PostgreSQL
• React
• Vite
• TypeScript
• TanStack Query
• Tailwind v4
• shadcn/ui
• Playwright
• GitHub Actions