a builder's codex
codex · operators · Tomasz Tunguz · ins_skill-distillation-frontier-teacher

Frontier models can write procedural SKILL.md files that cheaper local models execute, decoupling reasoning cost from inference cost

By Tomasz Tunguz · General Partner, Theory Ventures · 2026-05-29 · essay · Skill Distillation

Tier B · TL;DR
Frontier models can write procedural SKILL.md files that cheaper local models execute, decoupling reasoning cost from inference cost

Claim

Frontier AI models can author and refine atomic procedural SKILL.md files that smaller, cheaper local models then execute verbatim, creating a cost architecture where expensive frontier reasoning happens once and cheap local inference runs repeatedly.

Mechanism

The skill distillation pattern has three layers: a local markdown knowledge base of workflow files; atomic SKILL.md files authored and refined by an advanced model (the "teacher" step); and an agent loop running Plan, Tool Call, Observe, Refine cycles with API integrations. The small model does not need to reason from first principles. It only needs to follow the procedural steps already worked out by the frontier model. This decouples the cost of generating institutional knowledge from the cost of applying it. As models change, the skill files stay stable; the team swaps the execution model without rewriting the procedures.

Conditions

Holds when: the task is procedural enough to document as repeatable steps; the frontier model's procedure files have been tested against real-world cases; and the execution environment has the required API integrations.

Fails when: tasks require genuine situational judgment that cannot be reduced to documented steps; the frontier model writes procedures at too high a level of abstraction for the smaller model to execute; or the environment changes faster than the skill files are updated.

Evidence

"The frontier model becomes a teacher. The library becomes the company's institutional knowledge. The student becomes whichever model happens to be cheapest this quarter."

"Skill distillation retrieves procedures. The smaller model doesn't have to know how to evaluate a company. It just has to know how to follow the steps."

The example setup uses approximately 80 workflow files, 17 Rust API integrations, and runs execution on Qwen 35B or Gemma 26B.

Signals

Counter-evidence

The approach assumes tasks can be fully specified as documented steps. Tasks that benefit from genuine multi-step reasoning, creative problem-solving, or nuanced judgment degrade when executed by smaller models following documented procedures alone. The pattern fails quietly: the small model follows the steps and produces a plausible-looking output that is subtly wrong in ways the steps cannot catch.

Cross-references

Open the interactive view → View original source → Markdown source →