invent-by-hand
What it does
Takes a fellow's stated idea and forces an original product concept out of it — one a frontier model would not hand to every competitor who typed the obvious prompt. It peels the idea to the single invariant need underneath, generates a divergent set of concepts instead of the first safe one, ties each to a specific row of the fellow's proprietary corpus, and then runs the test that gates the whole skill: the generic-prompt test. Write the one-line prompt a competitor with no proprietary data would type into a model. If the model's default answer lands on your concept, the concept is a commodity — the machine already gives it to everyone — and it dies. What survives is unreachable by that prompt for a nameable reason: your trapped corpus, a tacit judgment the model cannot see, or a non-obvious inversion. The output is an invented concept plus the rationale for why a generic prompt could not produce it, ready to hand to wedge-five-questions.
The Icarus reframe
In the age of the machine, the danger is not that you cannot generate ideas; it is that you generate the same idea as everyone else. Ask a model "what should we build in factory operations?" and it returns the average of everything ever built there — a copilot, a dashboard, a chatbot — the one concept no one can defend, because the model will hand it to your competitor on the same afternoon. So this skill inverts the usual brainstorm: it invents by hand first, grounded in the fellow's own corpus, and then uses the machine only as an adversary. The rule is one line — if a generic prompt could produce it, it isn't yours yet — and it is enforced literally: every candidate is run against the prompt a competitor would type, and only the concepts that prompt cannot reach, for a reason you can name, are allowed to survive. The concept that survives is still a hypothesis; the test proves no competitor can prompt their way to it, not that a customer wants it.
When to use / When NOT
Use when a fellow has a stated idea (or a validated core need and a proprietary corpus) and needs the actual thing to build — the concept — and wants it to be non-obvious rather than the default the machine gives everyone. Trigger phrases: "what should we actually build", "invent the solution", "make it non-obvious", "give me a concept a competitor couldn't just prompt for", "what's the product idea here".
Do not use when:
| Request | Belongs to |
|---|---|
| "Is this the smallest thing to build first / will it get adopted / is my wedge sharp" | wedge-five-questions (sibling, section 06). It carves the invented concept into the first adopted slice and tests it 5/5. Run this to invent the concept, that to test the first build. |
| "Where's our proprietary data / what data can't a competitor get" | yoda-data-sourcing (section 03). It finds the corpus this skill stands on. Run it first — a concept with no band-P corpus to anchor to is invention on sand, and this skill routes back there. |
| "Design a cheap experiment to validate this idea / build a pretotype" | concierge-probe / probe-matrix (section 04). That is where a chosen concept meets behaviour cheaply. This skill produces the concept worth probing; it does not design the probe. |