Introduction
There's a moment in every software organisation where the unit of work flips: from "build the thing" to "build the thing that builds the thing." AI accelerates that flip. When agents can produce interactive content at the speed of prompts, the leverage isn't in writing more games or pages or modules — it's in writing the meta-system that produces them, including the meta-meta-system that improves the meta-system.
Why this matters
- Volume targets that were impossible (200 interactive games in a year) become possible — but only via meta-tooling.
- Quality scales sub-linearly with team size; framework leverage breaks the curve.
- Each layer of meta-tooling compounds the next, so early investment pays back exponentially.
- Hand-crafting at AI-output volume is impossible; you stop shipping.
Core concepts
Layers of leverage
Layer 0: build a thing. Layer 1: build a tool that builds things. Layer 2: build a system that builds tools. Each layer is harder, slower to start, but produces step-change throughput.
Specifications as the new code
When agents generate the implementation, the lasting artefact is the spec: schemas, evals, content rules. Invest in specs as your durable IP.
Templates and constraints
Open-ended generation is unreliable. Templates with bounded variation are how you get scale and consistency at the same time.
Pipelines for content
CI/CD, but for non-code artefacts: generate, validate, A/B test, deploy. Same disciplines, new outputs.
Practical patterns
Spec-first generation
Define the schema for the artefact (game, page, lesson). Generate against the schema; validate before shipping.
Eval-driven authoring
Each generated artefact is judged against rubrics (engagement, correctness, accessibility) before publication.
Component libraries for agents
A library of vetted building blocks the agent must use. Limits creativity in good ways.
Two-phase pipeline
Phase 1: agent drafts; Phase 2: human approves/edits. Rebalance the ratio as quality improves.
Pitfalls to avoid
- Building Layer 1 before you understand Layer 0; the meta-tool encodes the wrong abstractions.
- Letting framework complexity outpace the team's ability to maintain it.
- Generating volume without quality gates; you flood your platform with mediocre output.
- No way to iterate on the meta-system; you're stuck with v1 forever.
Key takeaways
- 1Move up the leverage stack deliberately; don't skip layers.
- 2Specs and evals are durable; generated artefacts are disposable.
- 3Constrain the generative surface; it's the only way to keep quality.
- 4Treat the meta-system as a product, with users (your authors) and a roadmap.
Go deeper · external resources
Curated reading list to take you from primer to practitioner. All links are external and free to read.