Seed, Sprout, and Harvest describe AI that runs inside your finished app, not how much AI helped you build. No model call? You can still enter as an AI-assisted build and skip the ladder.
Level 1
Seed
One prompt produces one useful response
A direct model call answers, summarises, rewrites, or generates.
Answers or creates
Level 2
Sprout
It grounds a response or calls a tool
Approved documents, data, or an API provide real-world context.
Uses evidence or a tool
Level 3
Harvest
It runs a bounded multi-step process
A code-controlled workflow or model-directed agent reaches a result.
Completes a process
Add a modality (optional)
Voice, images, and video are input or output choices you can add at any level. They do not raise the level or score by themselves, so use one only when it improves the experience.
Voice — Accept spoken input or return a spoken response.
Images — Understand or create an image, screenshot, or diagram.
Video — Analyse a clip, find key moments, or create a short video.
See it grow
Start with a Seed you can demo, then add features as you find time — the same idea grows as the AI does more.
Learn a language with a little help.
Seed
Generates example sentences from a word.
Sprout
Quizzes you on your own vocab list.
Harvest
Plans lessons, quizzes you, and tracks your progress.
Every level
Meet the same quality bar
Complexity is optional. Evidence, safe behaviour, and a working outcome are not.
1
Prove the outcome
Demo one representative task from input to a result you can verify.
2
Evaluate beyond the happy path
Try at least three cases, including one edge or failure case, and share what happened.
3
Make failure visible
Show uncertainty, missing evidence, and tool errors instead of guessing or silently continuing.
4
Keep people in control
Use public, dummy, or approved data and require approval before consequential actions.
Starter challenges
Pick one — a scoped idea beats a big vague one. Each works for any project type.
Seed
Rewrite Helper
Problem: A message, summary, or note takes too long to get right.
Build: An assistant that summarises, rewrites, translates, or reformats text.
Done when
It completes one clear job — summarise, rewrite, translate, or reformat — on an example a judge can run live.
A first-time user can supply input and understand the result with no extra instructions.
A messy, empty, or off-topic input is handled gracefully instead of returning a broken or confusing result.
Seed
Draft It
Problem: Starting from a blank page is the hardest part.
Build: Generate a first draft — an email, a plan, a snippet — from a short brief.
Done when
A one-line brief produces a usable first draft with the requested purpose, audience, and format.
Changing one detail in the brief visibly changes the draft as expected.
The output is offered as an editable starting point to refine, not a fixed final answer.
Sprout
Company Concierge
Problem: Customers ask the same product and support questions over and over.
Build: A branded support chatbot with a defined persona that answers from one business’s own knowledge — products, policies, and FAQs.
Done when
It holds a consistent persona and tone, and answers common questions from the supplied business knowledge.
A judge can trace each answer back to the business facts it was given — no invented prices, specs, or promises.
For anything out of scope or unknown, it admits it and points to a human or contact channel instead of guessing.
Sprout
Live Lookup
Problem: A useful answer needs current or approved data the model does not have.
Build: Give the AI a tool — an API, a search, a database — it can call to fetch what it needs.
Done when
It calls an approved API, search, or database and shows the value plus when it was retrieved.
A failed, slow, or empty lookup shows a clear fallback message — never an invented answer.
Harvest
Do It For Me
Problem: A repetitive, multi-step task eats time that could go elsewhere.
Build: An agent or workflow that completes the task end to end.
Done when
One representative task runs end to end to the intended result, with the steps it took visible.
The run has a completion condition and a hard step, time, or retry limit, so it always stops.
A failed step is reported and either stops or follows a defined fallback.
Any action that sends, changes, deletes, spends, or publishes waits for explicit human approval.
Harvest
The Team Play
Problem: One component cannot reliably finish a task that needs distinct skills, tools, or boundaries.
Build: A workflow or small set of specialised agents that hand off or combine work toward one result.
Done when
Each agent or step has a distinct responsibility with a defined input and output.
The demo shows at least one handoff and how a failed handoff is handled.
You can explain why one agent or a simpler workflow would be less reliable.
The full run has a completion condition and a hard step, time, or retry limit.