Put the strongest model available in the hands of a beginner, and give a seasoned engineer a model 2 years older: the engineer still ships the better result. That thought experiment opens Nate Herk’s field notes on working with the most advanced models, and it carries an uncomfortable consequence: if the practitioner beats the model upgrade, then the model was never the real advantage. The instructions, the systems, and the loops around the model are. This guide is about making that lesson pay: extracting how the best model you can rent actually works, so the method survives when the model is priced up, capped, or taken away.
“You can't keep the model's intelligence, but you can keep its process.”
Nate Herk, “How I Make Opus Think Like Fable (5 easy steps)” · 1:1301The judgment lives in the plan
The evidence for that claim is an experiment worth internalizing. Herk ran the same multi-agent workflows three ways: a top model directing equally strong workers, the same planner over mid-tier workers, and (in a separate test) a strong planner over the cheapest tier available. Results were roughly interchangeable; costs were not, with the cheap- worker runs landing around 3 times cheaper (roughly 67% less) for the same output. The pattern generalizes: when the expensive model does the scoping, anticipates what could go wrong, and reviews what comes back (redesigning the next step from each report), the workers are executing a well-specified plan, and executing a well-specified plan is exactly what cheaper models are good at.
Which reframes the top model’s job. It is the senior engineer about to rotate off the team, and your task is to have it package up everything it knows for the cohort that replaces it. Treat it as a teacher: have it review your setups, improve your skills, and, most importantly, explain its own reasoning while you still have it.
02Turn the discipline into a skill
The concrete move: when a top model hands you a deliverable you love but cannot quite explain, do not just save the deliverable. Reopen the session and interrogate it: what did you consider to get here, how did you verify it, why does this output work. Then have the model condense those answers into an installable skill file that any model can run, the same portable format of packaged expertise covered in Expertise you can install. The interrogation costs about 15 minutes at the end of a session; the discipline it captures outlives the model. Herk’s version gates work through five stages:
| Gate | What it forces |
|---|---|
| Scope | Devil’s advocate before any steps are listed |
| Evidence | Gather before reasoning |
| Adversarial reasoning | Attack your own draft answer |
| Verify | Evidence in hand before declaring done |
| Report | An honest read on how sure it is, not optimism |
The gate that earns its place is the first one. Scoping is not planning. A plan says “here are the steps”; scoping plays devil’s advocate first: every assumption that might be false, every unknown the plan silently walks past. The difference is invisible when things go right and decisive everywhere else, and the habit is worth stealing. The same discipline shows up in the system prompt of the strongest current model, which was reportedly leaked, so treat its origin accordingly. Recognizing something from training does not mean it is current knowledge. A file implied to exist is not a file verified to exist. At every step, check that what you have done is what you think you have done. That habit of auditing every claim against what the tools actually returned is the backbone of Guardrails for hours-long agent runs. Encoded as a skill, that discipline clearly improves a mid-tier model, and it runs on anything, including models you host yourself.
03Route by table, spend by effort
Owning the method also means deciding, per task, how much intelligence to rent. Two settings matter. The first is the model: keep an explicit routing table for your toolkit, one row per model, scored for cost, intelligence, and taste. Intelligence is reasoning strength: how well it follows you, reviews code, holds a long chain of reasoning. Taste is the other axis people forget to score: creativity, interface judgment, framing. They do not move together, and a table that collapses them routes design work to a strong-but-tasteless model and hard debugging to a tasteful-but-weak one. Hand the table to your orchestrator and let it delegate against it. When the pressure is a subscription limit rather than an invoice, Stop hitting your token limit turns the same routing into budget relief.
The second setting is effort. Published score-versus-cost curves show effort levels overlapping across model tiers (a top model on low effort landing near the previous tier on high), so the pairing, not the model name, is the real unit of choice. Effort has a ceiling. Past a point, maximum effort buys longer runs, second-guessing, and worse output than the tier below on high. The tell is a response that argues against its own correct first answer. When you see it, route down.
“We don't own these models. So, what we can own is our processes, our systems.”
Nate Herk, “How I Make Opus Think Like Fable (5 easy steps)” · 9:36Further reading
- Nate Herk · How I Make Opus Think Like Fable (5 easy steps), the field notes this guide distills
- Anthropic · building effective agents, the orchestrator-worker pattern from the vendor side
- Anthropic · Claude models overview, the tiers a routing table scores
- Wikipedia · mixture of experts, the same routing intuition inside the models themselves
- Yao et al. · “ReAct: Synergizing Reasoning and Acting” (arXiv), the plan-act-observe loop the gates formalize