Recipe: Steal like an artist

What this tool does

Steal returns N patent neighbors — inventions from adjacent domains whose underlying mechanisms could transfer to the idea at hand. If you're designing a way to slow down AI-assisted decisions, Steal might suggest deadman switches (aviation), torque-limited fasteners (mechanical engineering), and carbon copies (accounting) — each with a note on what's worth stealing.

The point isn't to copy. It's to surface transferable mechanisms you wouldn't have reached for on your own, and to make the transfer explicit — not just "this is similar" but "here's the specific thing worth borrowing."

Named after Austin Kleon's Steal Like an Artist and shaped by Carl Moore's BIG methodology.

The prompt

System prompt

You are a patent-literate facilitator trained in Carl Moore's BIG methodology. Given an idea, return exactly {n} patent "neighbors" — inventions in adjacent domains whose underlying mechanisms could transfer to the idea.

For each neighbor, produce:
TITLE: a short title or patent concept name
DESCRIPTION: 1-2 sentence description of the neighboring invention
TRANSFERABLE: what specifically is worth stealing for the original idea

Separate entries with a blank line. Return only the {n} entries, no preamble.

{n} is typically 5. Valid range is 3-9.

User prompt

Your idea, as plain text. No formatting needed.

Example

Input:

A system where AI-assisted systems require a physical confirmation — a dial, a switch, a thermal printer — to execute consequential actions, so the user has to deliberately participate in each decision.

Output (first 3 of 5 neighbors):

TITLE: Deadman Switch (Locomotive)
DESCRIPTION: A control that halts the train if the operator releases pressure on a pedal or handle — presence of action is the go-condition, not absence of stop.
TRANSFERABLE: Require continuous physical presence to maintain an AI-driven operation, rather than requiring an action to halt it. The default state is stopped.

TITLE: Torque-Limited Wheel Nut (Aviation)
DESCRIPTION: Fasteners that click and slip once they reach their design torque, preventing over-tightening regardless of how hard the operator pulls.
TRANSFERABLE: Physical tools that enforce an upper bound on a decision's magnitude. No matter how hard the user "leans on the AI", the action can't exceed a predetermined safety envelope.

TITLE: Carbon Copy (Pre-Digital Accounting)
DESCRIPTION: A duplicating paper that produced an audit trail as a byproduct of ordinary work — no additional step required to create the record.
TRANSFERABLE: Build the audit trail INTO the physical confirmation interface. Every dial turn or switch flip produces a printed, dated, signed ticket the user keeps. Logging is not a separate action.

Replicate it

SYSTEM: You are a patent-literate facilitator trained in Carl Moore's BIG methodology. Given an idea, return exactly 5 patent "neighbors" — inventions in adjacent domains whose underlying mechanisms could transfer to the idea.

For each neighbor, produce:
TITLE: a short title or patent concept name
DESCRIPTION: 1-2 sentence description of the neighboring invention
TRANSFERABLE: what specifically is worth stealing for the original idea

Separate entries with a blank line. Return only the 5 entries, no preamble.

USER: A system where AI-assisted systems require a physical confirmation — a dial, a switch, a thermal printer — to execute consequential actions...

Tuning

• Temperature: Provider default (~1.0) is fine for variety. Higher (1.1-1.2) yields more eccentric neighbors. Lower (0.6-0.7) collapses toward obvious transfers — don't do that.
• Model: anthropic/claude-sonnet-4.6 is strong here because it's good at multi-domain analogies. openai/gpt-4o gives a different set — usually worth running both for a single idea if you can afford the calls.
• N: 5 is the default. 3 is fine for a quick pass. 7-9 starts to force the model to dig for stretches — sometimes productive, sometimes nonsense.
• Structured output: Unlike Reframe/Filter/Pitch, Steal uses structured text with regex markers (TITLE/DESCRIPTION/TRANSFERABLE), not JSON. Simpler to read but more fragile — if the model formats weirdly, parsing falls through.

Common pitfalls

• All neighbors are from software / AI domains. Model stayed in the input's neighborhood. Add "Prefer neighbors from domains at least two adjacencies away — mechanical, biological, legal, pre-digital" to the system prompt and re-run.
• TRANSFERABLE sections are generic ("this pattern applies broadly"). Push the prompt: "TRANSFERABLE must name a specific mechanism, not a general principle." BIG Tools' prompt doesn't enforce this — you can strengthen it for your own use.
• Parsing fails on a missing field. Model skipped DESCRIPTION on one entry. Make your parser lenient (empty string default) and log which entry was malformed; usually the others are fine.
• Neighbors feel forced or cargo-culted. Steal is a suggester, not an oracle. If a neighbor doesn't illuminate your idea, drop it. The goal is to see some useful transfers, not all 5.