Skip to content

Prototyping fast: a hello world in one day

تیز پروٹوٹائپ: ایک دن میں 'ہیلو ورلڈ'

35 min read

Three ways to see it

  1. A one-day prototype has three ingredients. A clear input, a clear output, and a single happy path between them. Skip authentication, skip logging, skip styling, skip every administrative concern that does not change whether the core idea works. For the FBR tax-return summariser: paste the text of a single return into a Streamlit textbox, call Claude with the system prompt from lesson four, display the JSON output as a table. That is the prototype. It runs on the engineer's laptop. It is not production. It tells you in one hour whether the model can do the job at all.

  2. The single most powerful test is the user test. Take the working prototype to one real user, not the project sponsor. Sit beside them. Watch them try. Do not explain. Note every place they hesitate, every place they ask a question, every place the output disappoints them. One hour with one user replaces a week of internal debate. The Pakistani version of this lesson is that the user must be the actual analyst, not their boss. The boss tells you what the analyst should want; the analyst tells you what the analyst actually wants. The two are never identical.

  3. Three rules that protect a prototype. First, no real data on day one. Synthetic returns that look like real ones are fine. Real returns wait until the DPIA is signed and the redactor is in place. Second, no real authentication on day one. Run on localhost. Authentication is a six-week project that has nothing to do with whether the model works. Third, no production hosting on day one. Hosting is the moment when 'a thing on my laptop' becomes 'a thing the bank is responsible for', and that transition deserves its own checklist later, not now.

Quick check

Quick check: what makes modern AI different from a rule-based program?

The why-tree

Why-tree level one: why ship in a day rather than a month? Because the unknowns compound. A month-long design before any code rests on assumptions that become harder to challenge as more is built. A day-long prototype tests the assumptions cheaply, before anyone is emotionally invested.

Try this with Claude

AI-edge prompt to try with Claude or ChatGPT: 'Generate a 60-line Streamlit app in Python that takes a pasted tax return, calls Claude Sonnet with my system prompt, and displays the JSON result as a table. Add five synthetic Pakistani tax return texts as built-in examples. No authentication. No external storage. Comments in Urdu.' Run it. Fix every hallucinated import.

Sources

Sources and further reading. Streamlit docs (docs.streamlit.io). Gradio quickstart (gradio.app/quickstart). Anthropic Python SDK (github.com/anthropics/anthropic-sdk-python). Eric Ries, 'The Lean Startup' (lean.startup.com) for the discipline of minimum viable product. Nielsen Norman Group, 'How Many Test Users in a Usability Study?' (nngroup.com/articles/how-many-test-users).