This is how I move from using coding agents as a product thinking partner to building a "working proof of concept" (i.e. more than a shallow prototype, less than production code) with an AI coding agent.
TLDR: These are prompts I use to create a spec, plan, and agents file, and then I just have to say "let's execute."
If you're already using an AI coding agent (like Cursor, Claude Code, or similar) as a product thinking partner, here's a video walkthrough of how to turn your thoughts into functional prototypes. The steps are also written out below.
The UI library in the video is Wired Elements and the font is Gloria Hallelujah.
Create a fresh new directory for your prototype, and open it in your AI coding agent (or change into it).
Tell your AI coding agent:
Let's clone
https://github.com/talsraviv/from-thinking-to-coding
into a new top-level directory in this project called "from thinking to coding/"
Download the zip file from GitHub and drag the from thinking to coding directory into your existing AI copilot directory. Your AI agent will then be able to reference these templates using @from thinking to coding/.
Once integrated, the AI can see and use these templates when you reference them with @ mentions in your conversations.
Draft a document covering the "why" and high-level "what" - the product opportunity, user needs, success criteria, and what you're building.
I love having my AI coding agent guide me conversationally, pulling the context out of me. I'll create a new file called opportunity assessment.md and copy my personal favorite headings in there:
# Objective
# Target customer
# Success
# What I believe
# What I need to research
# Solution directions
# Risks to validate + how to validate them cheaply/quickly
Then I'll ask the AI coding agent to interview me:
I want your help in filling out `@opportunity assessment.md` I want you to converse with me and ask me questions one by one to help fill this out. Let's keep those sections as is and fit into them.
Think about what's missing, interview me, and pull insights and ideas out of me, and brainstorm with me.
Let's converse just enough to make a great lightweight doc from our collaboration. Wait on writing the doc until our conversation is done.
To write the doc, keep it succinct and readable (no filler) but make sure to capture all the gold. Use our original conversation words for the vivid/evocative stuff.
If you really wanna wing it, just dictate what you're thinking about building and why into your AI coding agent chat box, e.g.
I want to build a children's science museum version of ChatGPT demonstrating visibly how extended thinking works.
Pro tip: If you want to build a learning POC, you might start with a deep research on the topic you want to prototype and include it at the start of the conversation.
Let's @1-create-a-spec/ for @opportunity assessment.md
The AI will generate a detailed technical specification as a new section inside the original document.
Continue in the same thread:
Now let's @2-create-a-plan/
The AI will add a phased implementation plan to your document.
Continue in the same thread:
Let's @3-create-agent-instructions/
The AI will create or update an AGENTS.md file in your code project directory with development guidelines.
Use the spec, plan, and AGENTS.md as guides while building. Reference them as needed during development.
Continue in the same thread:
Let's execute the plan!
- I got tons of inspiration from Jesse Vincent, especially this blog post on using coding agents in general this blog post on self-sufficiency and this repository.
- I absolutely loved humanlayer's "Writing a good Claude.md" post and incorporated insights into this repo.
- It was very cool to see these insights cohere with Anthropic's experience applying the Claude Agent SDK