Prompts to Pipelines v3: Five Model Generations Later
Self-contained HTML deck. Arrow keys, click, or swipe to navigate. A theme toggle lives in the top-right corner of the deck.
Overview
Between the April edition of this talk and this one, the ground moved: five model generations, one federal export-control intervention, and a summer in which the orchestration plumbing every serious team had hand-rolled started shipping natively in the harness. So this is not an update. It is a rebuild around the arc the field actually traveled: prompt engineering optimized the phrasing, context engineering optimized what the model sees, and harness engineering, 2026's discipline, optimizes the system around the model. Each era kept the previous era's tools and demoted them from strategy to hygiene.
The middle of the deck walks the two layers you still own. The standards layer: what qualifies an instruction as a standard (written from incidents, enforceable in one line, derived not remembered, context-budgeted), evals for prompts, the enforcement ladder from prose to physical rejection, and evidence discipline, where DONE is a verdict the evidence earns. The runtime layer: context as a measured budget, state that survives sessions, orchestration doctrine after Dynamic Workflows, skills as a supply chain you audit, and governance that outlives its vendors.
The proof is longitudinal: RDF's 19-week public arc, 28 releases, each answering a failure the previous one exposed. An RCE in the audit tool. CI that was green because it never ran. Three consecutive releases that fixed nothing but the framework's own honesty. Every number in the deck was re-derived from source the day it was rebuilt, because the framework's changelog once caught the author forward-copying stats, and the fix was making derived-not-remembered a CI failure.
Key Takeaways
- The discipline grew up: prompt engineering, then context engineering, then harness engineering. Each era kept the previous era's tools and demoted them from strategy to hygiene.
- The model is rented. The plumbing is shipping. The discipline is yours: it encodes what your shop means by done, and no vendor can ship you that.
- A standard is an instruction that survives contact with an agent that did not read your mind: written from incidents, enforceable in one line, derived not remembered, context-budgeted.
- The enforcement ladder: prose asks, review catches, hooks block, physical rejection. The rule that matters is the one a process enforces when nobody's reading.
- DONE is a verdict the evidence earns. Evidence blocks with re-runnable citations, a QA pass that re-runs them, and a separate calibrated evaluator, because agents overrate their own work.
- Rent the plumbing, keep the doctrine. The hand-rolled orchestration layer was deleted within weeks of the native version shipping; canonical content plus generated adapters outlives every tool it governs.