Topic
The architectural principle that using AI inference for deterministic, computational tasks is the most common and most avoidable performance and cost problem in AI workflows.
Target Reader
Knowledge entrepreneurs, analysts, and coaches who are using AI heavily but noticing that some tasks feel slow, expensive, or unreliable. They’re not programmers. They think Python is not for them. They’ve never considered that the tool they already have (Claude, Gemini, ChatGPT) can write the code that replaces its own inference for the right class of task.
The Fear / Frustration / Want / Aspiration
Frustration: AI feels slow or hits limits when processing large files or datasets. Fear: AI costs are adding up and they’re not sure if the ROI is there. Want: faster, more reliable results without needing a developer or a higher subscription tier. Aspiration: the analytical capabilities of a data team, on a solo budget.
Before State
They’re using Claude to filter spreadsheets, aggregate data, and process files with hundreds of rows — tasks that have a single correct answer. It’s slow, sometimes fails, and always costs inference credits. They believe the only upgrade path is a better model or a higher-tier subscription.
After State
They understand the inference/computation distinction at an intuitive level. They know how to describe a transformation task in plain English and get working Python back. They’ve run at least one script on a dataset. They’re thinking about which other repetitive tasks they’re overpaying inference to handle.
Narrative Arc
Scott Delinger, working on a nonprofit budget, processed 90 million research records in one night with Python he’d never written before — code Gemini wrote in a conversation. Lou’s Gears pipeline went from 6 hours to 15 minutes with the same architectural move. Both arrived at the same principle from different directions: AI is most valuable when it’s making decisions. It’s expensive overkill when it’s doing arithmetic at scale. The unlock is knowing which is which — and knowing that AI will write the code for you.
Core Argument
The most common AI efficiency leak isn’t model choice or prompt quality — it’s using inference to do the work that code handles faster, cheaper, and more reliably, because most people don’t know they can ask AI to write the code instead.
Key Evidence / Examples
- Scott: 90,900,000 entries → 128,000-line dataset, 10 hours 16 minutes, 500MB RAM peak, going back to 1993. “This would not have been possible without the collaboration with Gemini on writing Python.”
- Lou: Gears pipeline 6–7 hours → 15 minutes. “Claude handles only routing and evaluation decisions, while Python handles the computational heavy lifting.”
- The embedded script pattern: scripts inside skill definitions, spawned by AI at the right moment, zero API cost for the computation itself
- Insight - Process Architecture Transmits Judgment More Reliably Than Individual Prompts — the meta-principle about keeping AI in the right lane
- Insight - CLI-First Micro-Apps — Why the Most Durable Personal Tools Skip the UI — the durable-tools angle on the same idea
Proposed Structure (5–7 beats)
- The problem most people can’t see — when you use AI for deterministic tasks, you’re paying inference prices for work a calculator would do faster
- Scott’s story — 90 million records, a nonprofit budget, Python he didn’t know, one conversation with Gemini, one overnight run
- Lou’s pipeline — from 6 hours to 15 minutes; the architectural shift that changed the math
- The decision rule — “Does this task have a deterministic correct answer?” If yes, it belongs in code. If it requires judgment, that’s where inference earns its price.
- You don’t need to learn Python — you need to describe the task clearly; that’s a different skill, and it’s learnable
- The five-step pattern — describe → get code → test on sample → run at scale → inference for edge cases only
- What becomes possible — the analytical capabilities of a data team, available at solo-practitioner cost, limited only by description quality
Related Insights
- Insight - Code Is for Computation, Inference Is for Judgment
- Insight - Process Architecture Transmits Judgment More Reliably Than Individual Prompts
- Insight - CLI-First Micro-Apps — Why the Most Durable Personal Tools Skip the UI
- Insight - Use AI to Compress the Iteration Cycle, Not Replace the Thinking
Editorial Notes
The two concrete examples (Scott’s 90M records, Lou’s 6h→15min) are load-bearing. Don’t summarize them — use them in full. The reader needs to feel the scale difference viscerally to take the principle seriously. Keep the tone accessible: the goal is to convince someone who thinks coding is not for them that they can apply this today, in their next AI conversation.
Next Step
- Approved for drafting
- Needs revision
- Deprioritised