The AI-native founder
Four weeks. Real users. A loop that runs every day. Use AI to turn a clear goal into a small working product, then improve it with evidence.
Reviewed for this workshop · Official sources · Check live plans for current access and prices.
01 · The way you work
You own the user, the goal and the decisions. Agents help you research, build and test. A workflow repeats the useful routine.
Talk to five people you can reach. Watch how they solve it today. Ask AI to organise the evidence and challenge your assumptions.
Keep: one user, one recurring need, real examples.
Give the agent the goal, context, constraints and a clear finish line. Attach examples. Ask for a short plan and the next smallest useful step.
Keep: a brief the team can reuse.
Make one real input become one useful result. Ask the agent to run it, test it and show what changed. Try it yourself, then watch a user use it.
Keep: a working link and evidence of actual use.
Save 5–10 examples: normal requests, missing input, duplicates and past failures. Use fixed rules for facts you can test, plus a separate review for judgment.
Keep: pass/fail results, not “the AI says it works”.
Choose its trigger, allowed actions, retry limit and stop condition. Log every run. Keep human approval for sensitive actions and exceptions.
Keep: a daily run, a cost cap and a stop button.
Review repeat use, failures, cost per useful result and your minutes. Change one thing. Compare old and new on the same saved cases before shipping.
Keep: one measured improvement each week.
Replace the brackets. Keep the brief, examples and decisions in your project files so the next session can pick up the work.
GOAL Help [five named testers] get [one useful result] every [day/week]. CONTEXT They currently [do this]. Here are real examples, approved sources and what we learned: [attach]. CONSTRAINTS Use the current project and the smallest useful change. Daily budget: [amount]. Keep secrets out of code. Agree before the first live send, spend or data change. Then work within those limits; ask before exceeding them. Flag missing facts. DONE WHEN One input becomes a checked result. These examples pass: [attach]. Show the working result, checks, changed files, remaining risks and how to stop it. First: propose a short plan. Then build and verify the first slice.
02 · Choose your tools
One assistant. One coding agent. One place for your evidence. Add a cloud runtime when the job is ready to repeat.
Think & research
Use the assistant you already have. Explore options, understand unfamiliar work and draft interview questions. Give it your evidence; ask for sources and uncertainty. On eligible paid plans, Claude Desktop can also work on files in a folder you allow; this was called Cowork.
Cost (checked 13 Oct 2026): Free tiers, with optional paid plans: ChatGPT, Claude, Gemini.
Leave with: a decision, a reason and a next test.
Build & verify
Let an agent work in your project: edit files, run commands and tests, and fix failures. Give it a small deliverable and an explicit way to verify it.
Access: Codex plans and current limits. Free access is limited; available features depend on plan and rollout. API usage is billed separately.
Access: Claude Code plans. Check current plan access and usage limits before paying.
Leave with: a working slice, a readable change and test evidence.
Connect & repeat
Connect a form, schedule, model call, check and delivery. n8n provides a visual workflow; an agent can also write a small program that runs on a server.
Cost (checked 13 Oct 2026): n8n Community Edition is free to self-host; hosting and connected APIs may cost extra.
Leave with: a trigger, a run log, a retry limit and an owner.
Make useful media
Start with your assistant for simple visuals. Add Higgsfield when video is part of the product. Specify audience, format, reference and what a good result must show.
Cost: Check current plans and credits. Verify the terms for your intended commercial use and budget for the actual generations you need.
Leave with: one useful asset. Check accuracy, consent and rights before publishing.
A shared document or sheet holds interviews, source links, test cases and costs. GitHub keeps versions of code and briefs. Save decisions there; do not rely on one long chat.
Start free where possible. Pay for one bottleneck. Budget separately for subscriptions, API calls and hosting; check how your chosen route is billed. Challenge support →
Gemini Notebook (formerly NotebookLM) grounds answers in selected sources and provides citations. Good for interview notes and course material. Open the cited passage and check the claim.
Grok Bot offers a cloud computer for delegated work. Review its work and permissions; give it a bounded task.
Access: See the current Grok Bot access requirements; plan access and regional features can change.
Tagging @bot on X is not available for EU-based accounts, including Latvia.
GitHub Pages fits static class demos and project documentation. Choose an appropriate app host for a customer-facing service or backend.
Cost (checked 13 Oct 2026): Pages and Actions with standard hosted runners are free for public repositories.
n8n Cloud is managed hosting with a trial, then paid plans. Community Edition is free to self-host; the server and upkeep are yours. Actions schedules can be delayed or dropped. Choose a runtime that meets your users' timing needs.
03 · A loop in practice
Test the result by hand first. Then make this small routine repeat. The tools serve the job.
A hosted schedule starts the run.
Read employer pages the team selected.
Match roles to the student's stated interests.
Reject expired roles, duplicates and unsupported claims.
Send once; keep failed deliveries visible.
Record clicks, feedback, cost and your minutes.
Approve the first editions. Save normal and broken examples. Put secrets in the host's secret settings; give each connection only the access it needs.
Enforce a daily spend cap and limits on results and retries in the workflow. Stop or ask you when checks fail. Put the stop control and failure log where the team can find them.
Invite the next five users with a useful sample. Deliver a checked shortlist. Each week, test one change against the same cases and user feedback.
Keep your first loop simple. Add parallel agents or more tools when one measured bottleneck justifies them. Build the control panel →
04 · Keep up without losing the day
A launch post is a lead to investigate. Your own result decides whether a new tool earns a place in the stack.
Use a private X List: the tools you use, two builders and people in your users' world. Read that instead of an endless feed.
Find the official release or docs. Check the date, availability, price and limits. Grok can help find sources; read them yourself.
Use the same input as your current tool. Compare quality, minutes and cost. Queue a bounded experiment if it needs longer.
Write one line: claim → evidence → result → decision. Adopt it only if it improves your product. Then return to building.
Create a private List and pin it using your existing account. Pick a few accounts, not every one below.
Add local founders, customers and people doing the job you want to improve. Product demand is a signal too.
Ask Grok: “Find the original announcement for this claim. Link the official docs. What is available to me, and what would I need to test?” Keep private work out of public replies.
Do one thing now
Use the next hour to put one useful result in front of a real person.
The working method above adapts the official guidance below to this four-week challenge. Product capabilities, availability and prices change; follow the live provider pages before paying. Subscription access, API charges and hosting are separate. Check your account and the provider before spending.