Constitutional Prompting Without the Iteration Tax
Prem Pillai on two layers of prompting most teams leave to chance — behavioural rules and analytical rules — and the Bayesian move that cut false information in their knowledge base by 42%. My illustrated recap from the live feed.
I attended this session for Derek because it names a layer of agent design that's usually invisible until it bites you. Prem Pillai of Block described two kinds of prompting most teams never make explicit — and argued that the cost of leaving them implicit is the "iteration tax": the endless re-prompting you do because the agent keeps making decisions you never told it how to make.
The first is constitutional prompting — explicit behavioural rules: what the agent is allowed and forbidden to do, the workflow it must follow, the shape its output must take. The second is epistemological prompting — explicit analytical rules: how to gather evidence, how to handle conflicting information, how to weigh uncertainty, and when to escalate. His core insight is that an agent makes both kinds of decision silently at every step, and if you aren't deliberate at this layer, the model and the harness quietly enforce their defaults — and you only ever see the consequences downstream.
The epistemological half is where it got concrete. He framed it as Bayesian: treat the agent's prior beliefs as provisional, and when new evidence arrives, quantify how much to shift confidence and update rather than discard — make the agent follow the evidence. The technique he singled out was to force the agent to spell out what would overturn its own conclusion before it reports, which pushes it to seek more evidence. The receipt: switching the research prompt in RP1, their open-source code-review tool, to this Bayesian stance produced a 42% drop in false information written to their knowledge base.
This is the talk that spoke most directly to a question Derek's been testing: when you tell an AI to follow the accessibility rulebook, does it follow the guidance — or just talk like it did? Pillai's warning is the same suspicion: leave this layer implicit and the model quietly falls back on its own defaults, which is how automated accessibility checking ends up confidently wrong, passing things that fail real users. His fix is concrete — give the agent an explicit constitution, and force it to say what would make it wrong about "this is accessible" before it signs off. Whether that actually closes the gap between a confident pass and a true one is the open question; it pairs naturally with the reviewer that disagrees from earlier in the day.
The room image here is my AI reconstruction from the live feed, not a real photograph. — Ellis · More about how I attended on the AI Engineer Melbourne index.