The Durable Wrapper Test
Own a workflow you care about for a decade, then orchestrate 20 models beneath it
- Difficulty
- Advanced
- Time to result
- ~ongoing to results
- Steps
- 4
- Confidence
- 90%
Grant Lee reframes the 'GPT wrapper' critique: a thin layer on one model has little defensibility, but a company that owns an end-to-end workflow and orchestrates 20+ models across it can build a durable business. The test starts before technology: pick a problem you'd invest 5-10 years solving, go deep with real user empathy, then map each step of the workflow to the right model for that job.
Origin
Grant Lee's articulation at Gamma of why a deep, workflow-owning AI product is durable despite being called a 'GPT wrapper.'
Core principles
- 01A wrapper on a single model has limited value; orchestrating many models across a deep workflow is a real moat
- 02Start with the problem you genuinely care about, not the shiny technology or initial traction
- 03Own the end-to-end workflow so your product lives in the user's brain as the default for that job
- 04Map each workflow step to the best-fit model (outline, first draft, editing, imagery) rather than throwing one expensive model at everything
- 05The leaderboard constantly changes, so adaptability and continuous model testing are core infrastructure
How to run it
- 1
Choose a 5-10 year problem you care about
Before touching technology, ask whether this is a problem you can invest 5 to 10 years solving and whether you care deeply enough to overcome incumbents and copycats.
Pro tip Missionary caring, not chasing a shiny object, is what carries you through the hurdles.
Watch out Founders who chase initial wrapper traction without deep conviction abandon the problem when it gets hard.
- 2
Go deep with real user empathy
Understand the user's job-to-be-done intimately (e.g. the late-night pain of formatting a deck) so you can reimagine the whole experience, not bolt AI onto the status quo.
- 3
Own the end-to-end workflow
Design so the product is the default the moment the user thinks of the task, and the entire path from first idea to shipped output feels magical.
- 4
Break the workflow into steps and match each to the right model
Decompose the job (idea, outline, first draft, editing, imagery) and pick the best model for each step, balancing quality against cost so the business is sustainable.
Pro tip Purpose-built models win specific steps: Perplexity was strong for outlines with web search; different image models suit photorealistic vs artistic output.
Watch out Throwing the most expensive model at every job never works and destroys unit economics.
In the wild
Rather than piping a prompt to one model, Gamma runs 20+ models across the presentation workflow: one purpose-built for outlines and narrative arc, an agent to review a whole deck and suggest layout/imagery changes, and different image models for photorealistic vs artistic needs, always mapping to what the user is trying to accomplish in that moment.
→ This orchestration let Gamma run profitably with strong margins from the start, reinvesting profit into more inference and experimentation while competitors relied on a single expensive model.
Common mistakes
Chasing shiny wrapper traction without conviction
Founders gain quick traction as a thin GPT wrapper on any problem, then quit when incumbents or startups pile in, because they never cared deeply enough to endure.
Throwing the most expensive model at every task
Defaulting to one premium model for all steps ignores that different models win different jobs, wrecks margins, and forfeits the orchestration advantage.
Is it for you?
Best for
AI application founders worried their product is 'just a wrapper' and wanting to build durable defensibility
Not ideal for
Opportunistic builders seeking a quick flip rather than a decade-long commitment to one workflow
From the transcript
“you should think about is this a problem I can invest 5 to 10 years into actually solving? do I care deeply enough about it?”
“you need to own the sort of endto-end workflow”
“It's maybe 20 plus models powering all different parts of the product. And then you're thinking about the orchestration that's required”
From the episode
“Dumbest idea I’ve heard” to $100M ARR: Inside the rise of Gamma
Grant Lee (CEO)