Essays
PromptQL's writing on shared context, accuracy, and building AI you can actually trust.
Defusing slop grenades and other multiplayer AI stories beyond group chat
Personal AI agents changed individual work; an AI agent in a group chat hasn't done the same for teams. Five hand-offs in team workflows that need to get better, and why the behavior shift may be the harder part.
Read→
User Identity: the access model for multiplayer AI
Multiplayer AI shouldn't mean a shared login. The agent should act as whoever is driving it — their identity, their permissions — so a shared conversation never quietly becomes shared access.
Read→
Why is it called PromptQL?
People think PromptQL is the AI you talk to. It's actually the layer underneath the AI you build — your models, your data, your context, your name. The plumbing is boring on purpose.
Read→
Designing an Auth model for Multiplayer AI
Multiplayer AI — people and agents sharing the same context — breaks the assumptions behind traditional access control. Here's how we designed an auth model that holds up.
Read→
Stop tokenmaxxing. Start contextmaxxing.
We've spent the last year realizing AI will soon do everything — provided it has enough useful context. The alpha has shifted from using AI to maintaining useful context.
Read→
2 reasons why Slack can't contain the "company brain"
Two simple reasons Slack hasn't become, and won't become, the company brain: it captures coordination rather than execution, and it's built for messaging, not prompting.
Read→
Killing Slack was the only way to make AI accurate
The three big bets we made to build AI that works accurately on internal data — and why moving beyond Slack was essential to getting there.
Read→
43% accuracy with Opus-4.6 & friends - will Text-to-SQL ever be good enough?
The Data Agent Benchmark from UC Berkeley's EPIC lab shows frontier models top out around 43% on real-world data questions — and what that means for text-to-SQL on enterprise data.
Read→
On shared context
Companies run on shared context. Becoming AI-native requires systems that can capture, maintain, and apply that context for AI. Here's the blueprint for building one.
Read→
The semantic layer is dead. Long live the wiki.
Most “AI on data” programs are just semantic-layer maximalism with a new paint job. A Wikipedia-style model is a better way to think about organizational meaning — and to make AI reason accurately.
Read→
Introducing the GenAI Assessment Framework (GAF): A 3×3 Matrix to Map Enterprise AI Needs
Introducing the GenAI Assessment Framework (GAF): a 3×3 matrix that helps AI leaders map enterprise AI needs and measure progress.
Read→
Static AI Is Already Cutting Jobs; What Happens When It Starts to Learn?
Most of the labor-market impact we're seeing comes from mediocre, static AI. What happens when these systems start to learn on the job?
Read→
Being "Confidently Wrong" is holding AI back
The failure mode that stalls “AI for data” efforts isn't psychedelic hallucination — it's confident inaccuracy: plausible answers that are wrong in subtle and costly ways.
Read→
Expose the Data Prep Tax: A Path to High-Impact AI Deployment
Every company pays a hidden “data prep tax” — weeks of setup and constant upkeep. Eliminate it to unlock the curiosity and innovation that's been dormant for years.
Read→
Find the Unanswered Questions: A Path to High-Impact AI Deployment
$100M lost to unanswered questions. Reliable AI that speaks the language of your business lets every employee ask and act instantly.
Read→
Spot the Automation Paradox: A Path to High-Impact AI Deployment
$200M lost to hidden automation gaps. AI that speaks expert logic can unlock massive value.
Read→