From prompt to prod

Materials for a short talk about Cursor and agent-style development: how to go from an idea in the chat to something you can ship, without losing your mind along the way.

Speaker: Vedran Mandić
Context: Backend engineer at Similarweb, shared at CroAI on 21 April 2026.

This repo holds the slides so you can read them at your own pace:

• from-promt-to-prod-v2.pdf — PDF export (easy to skim and search)
• from-promt-to-prod-v2.key — Keynote source (if you use Keynote)

What the talk covers

1. Cursor in context — AI-native editor on a familiar VS Code base, agents in the UI (and CLI), diffs, indexing, and team-oriented features.
2. How LLMs actually show up in the editor — tokens, system prompts, tools, history, and your messages. Why “big context” on the label is not the same as comfortable room left for your work.
3. Project setup — AGENTS.md, scoped rules, skills, commands, hooks, and MCP configuration. Rules for steady conventions, skills for repeatable workflows.
4. The agentic loop — plan before you generate, review like you would a junior’s PR, then verify with tests and your own eyes.
5. Prompting — be precise when you know the outcome; stay a bit open when you are exploring. Use paths, symbols, and docs the editor can see.
6. Parallel work — subagents, worktrees, and cloud agents when they fit the problem (without forcing complexity).
7. Extend the toolkit — MCPs, small CLIs wrapped as skills, team skill repos, and a few ecosystem ideas worth knowing.
8. Trust but verify — least privilege, branch hygiene, and keeping secrets out of plain text the model can read.

The deck is an honest personal take, not a product pitch. If something sparks an idea, try it on a small real task first.

Takeaways (short version)

• Spend real time on context: globs for rules, lean skills, fresh threads for unrelated work.
• Treat plan and review as first-class steps; execution is cheaper when the plan is right.
• Prefer scoped automation (skills, MCPs, scripts you own) over giant always-on rule packs.
• Verify everything that touches production paths, credentials, or customer data.

Check your setup against the talk (agent prompt)

This is how you can see whether your project and workflow line up with the practices from the presentation Vedran shares—context, agent loop, tooling, and safety—in a way any AI assistant can run.

• File: src/verify-your-agentic-workflows.md — copy or @-reference it in chat. The agent must (1) ask computer-wide vs project, (2) ask where you keep source code (one or more directories) for machine-wide runs so it can sample real repos—not just ~—and (3) plan the work and use subagents (or serial “tracks”) to split home/global config (Cursor, Claude, Copilot, Codex, …) from project evidence. Every section A–K includes a best-practice line from the talk plus your evidence, so a run is never an empty “not observable” list. Front-load to skip questions, e.g. computer-wide, source roots ~/dev ~/work, home scan OK.
  Note: This repo is mostly talk materials (PDF, README). Project-mode here still yields lots of “N/A to files” for D–F; the prompt says how to treat slide-only vs application repos. For a rich check, use project on a repo you ship and/or computer-wide with source roots you actually use.
• What you get back: a research plan, T1/T2 scan notes, per-dimension anchors + evidence, crosswalk, caveats, and an open closing—still no secret values in the output.

Links mentioned in the slides

• Addy Osmani — agent skills — collection and patterns for packaging reusable agent “skills” (workflows, scripts, instructions); the deck uses his treat the model like a junior quote and the full loop (plan → act → review) framing.
• Cursor “doctor” (community) — community diagnostics / health checks for a Cursor install (useful for debugging odd editor or agent behavior; not official Cursor support).
• Notes on Cursor Auto mode and model choice — empirical write-up on whether Auto mode picks a sensible model; supports the talk’s point that you should still understand model / cost / task fit instead of assuming the label is magic.
• Claude Code best practices (community collection) — curated tips and links for Claude Code (and adjacent agent workflows): prompts, context, skills, and team habits; community-maintained, not an Anthropic official doc.
• Superpowers (Obra) — a library of skills and workflows (TDD, planning, verification, etc.) designed to be loaded by agentic tools; fits the “skills + process discipline” thread in the talk.
• MemPalace — long-term memory for coding agents (MCP tools, wings / rooms / drawers, optional knowledge graph; see also the MCP integration guide).
• Skillshare (runkids) — distribute and sync a team’s skills (and related rules) via git so everyone’s agents share the same behaviors; pairs with a shared skills repo in your org.

Contact

• GitHub: github.com/vmandic — this repo: github.com/vmandic/from-prompt-to-prod
• Handles: @vekzdran, @vmandic
• Email: [email protected]

If something in the deck is unclear or dated, open an issue or reach out. Happy building.