AI Development · 2026-07-05 · Michael Ditter
Stop Prompting. Start Briefing.
The model was never the bottleneck — the brief was. How I design what an AI sees before I ask, from Sentinel's multi-agent pipeline to Fortune 500 rollouts.
I blamed the model for months. Drafts came back fluent, confident, and useless — right facts, wrong audience, wrong register, no point of view. Then I audited my own inputs and found the defect. I had been telling the AI what to produce and nothing about who it was supposed to be while producing it. The failure was never in the answer. It was in the brief.
The Tricks Era Is Over
Two years ago, getting good output meant collecting incantations — phrase it this way, bolt on "think step by step," stack the magic words and hope. That game is finished. What replaced it is context engineering: deliberately designing everything the model sees before your request ever lands. Not the phrasing of the ask. The environment around it.
Here's the reframe that changed my work. Treat the model like a search engine and you get search-engine results — generic, defensible, flat. Treat it like a brilliant hire on day one, with zero institutional knowledge and infinite patience, and the output changes class. Nobody hands a new hire a task with no background and grades them on the result. You brief them first. Same discipline here. Same payoff.
The Four Moves I Make in Every Brief
I've shipped more than 100 prototypes — voice agents, climate systems, retrieval pipelines, a narrative game. Four moves show up in every one that worked.
1. Assign the seat before the task
The first sentence of my brief names whose chair the model sits in. "You are a CFO three days out from a board meeting, deciding whether this vendor renewal survives the budget cut" does more work than any instruction that follows it — one sentence collapses the answer space from everything to the handful of responses that specific person would actually write. Skip the role and you get prose optimized for nobody in particular. That is exactly what it will read like.
2. Write constraints a machine could grade
"Write a summary" always yields a summary. That's the trap — output appears, so it feels like progress. I write constraints I can check in seconds: 150 words, zero jargon, one risk named explicitly, close with the action I should take next. Pass or fail, no debate. Testable constraints do double duty — they make good output repeatable and bad output obvious. Vague instructions hide failures inside competent-sounding prose, and hidden failures are the expensive kind.
3. Show two examples instead of describing ten requirements
Examples are the highest-return element of a brief and the one most people skip. Two real samples of what good looks like, dropped straight into the prompt, carry structure, tone, and the weird edge cases all at once — no requirements list gets close. Description is lossy. Demonstration isn't. When I stopped writing longer instructions and started curating better examples, my revision cycles collapsed.
4. Save the brief — it's an asset, not exhaust
A prompt that works once is a lucky roll. A prompt that works fifty times is infrastructure. Every recurring deliverable in my operation — status brief, contract summary, launch draft — has a saved, versioned prompt behind it, with a purpose statement and pass/fail criteria attached. When I improve one, I change a single variable and re-run, the way you'd run any experiment. And I edit ruthlessly: any line that doesn't move the output gets deleted. What survives is the smallest brief that still produces consistent results.
Where This Stopped Being Theory for Me
Sentinel made me prove it. Sentinel is the climate-risk system I ran live at the SCSP AI+ Expo in Washington, D.C. — a multi-agent Claude pipeline wired into real-time NOAA, NWS, CMS, and NPI data. One agent parses storm feeds. Another maps exposure against healthcare facilities. A third writes the alert a human actually reads. No person sits between those agents to clarify intent — each one runs cold, off a written brief: the role it plays, the constraints its output must pass, examples of correct output, the exact format the next agent expects.
Every misfire in that pipeline traced to the same root cause. Never phrasing. Always a missing constraint or a bad example. In a multi-agent system, context engineering stops being a technique and becomes the architecture — the briefs are the load-bearing walls.
The enterprise version is the same lesson at lower speed. In the Fortune 500 rollouts I've worked, the teams stuck at "AI writes our first drafts, badly" are prompting. The teams compounding — output measurably better month over month — are briefing, and saving the briefs where the next person can find them.
The Model Forgets. Plan for It.
Context windows are enormous now — 1,000,000 tokens on the frontier — and none of it survives the session. Close the tab and your collaborator is a stranger again by morning. The fix is standing memory: project-level buckets that hold brand voice, audience definitions, recurring constraints, house terminology. Load them once. Draw on them in every session after.
Supermemory, a vendor in the memory space, benchmarks the payoff at roughly 8 hours a week for professionals who maintain standing memory versus starting each session from zero. I discount vendor math on principle — but the direction matches what I see in my own weeks. Setup costs an afternoon. Then it amortizes across every session you will ever run.
The Bar: Someone Else Can Run It Cold
Two capacities decide who actually gets faster with these tools. Agency — you design what the model sees; you own the brief. Taste — you can tell whether what came back is right. Context engineering is the muscle behind the first, and it quietly trains the second, because writing testable constraints forces you to define what right means before you ever see an answer.
My definition of done is simple. A colleague picks up the prompt, runs it with zero explanation from me, and the output still passes spec. Until it clears that bar, it's a note to self. After it clears, only the inputs change — the instructions hold.
The model was never the bottleneck. The brief was. Fix the brief.
Adapted from THE UPLOAD — my living AI guide for working professionals. The full playbook, with copy-ready prompts and a narrated audio edition, lives there.