Product Update

Practice With a Clock, Video With a Label, and a Loop You Can See

The Simulate products, generated video I was on the fence about until I wasn't, the two pages that show whether your preparedness loop is actually running, and a free, public set of our skills for whatever AI model you already use. What shipped this month, and why each piece is built the way it is.

Two weeks ago I wrote several thousand words about procurement being the hard part. While I was writing them, the team shipped more than I could keep track of. We put up a new preppr.ai so the website finally describes the product we have. This post is about what's behind it: the Simulate products, generated video, the two pages that show whether your preparedness loop is actually running, and the skills we've started giving away.

Usual disclosure. I run the company. Read accordingly.

Simulate: practice with a clock on it

In the last post I said a tabletop tests a plan but doesn't test time to decision, that Europe is well ahead of us on actual simulation, and that the United States is still write a document and talk at each other. I also said the pull toward live response is universal, and that AI hasn't earned response yet. The simulation work is my answer to both of those at once. If people want reps under time pressure, give them reps under time pressure, in a synthetic incident, with the labor done by machines and the decisions kept by people.

There are four ways to do it now, three of them under a Simulate entry in the sidebar and one inside the exercise designer. Every one of them ends in Improve.

Simulate Ops is a multiplayer practice Operation Center. People take seats on a room map. AI crews hold the seats nobody filled. A Threat team plays the hazard and the adversary. Preppr Verified physics is the truth underneath, an incident clock runs (it says hour five of twenty-four, never tick), and everything goes on one ledger. Pick an operating model, Incident Command, business continuity, or one of two FEMA EOC charts, set the dials, and start. The opening situation lands as a full-screen alert and ends on a button that says take your desk. From there your desk sorts the work into decide now, escalated to you, and for awareness. Callers ring your seat, and if the host allows it they ring your actual phone. Your section's AI crews send you one staff memo with a recommendation, the risk and who agreed, instead of a pile of escalations, and you can send it back with a question. You can call a huddle, and the clock keeps running while you talk. The host sees the pressure on the room and what's coming up, and can watch any AI seat work: its desk as it sees it, its reasoning step by step, and a button to take the seat over. We rebuilt nine of these screens last week from what a live run showed people needed.

The Operation Center board for Riverside Flood, run 1, at hour 8 of 24: one bar per incident hour grouped by phase, what the scenario did on the left, and what the room is doing on the right, filtered to the four items that need you.

The board at hour eight. The scenario on the left, the room on the right, and the four things waiting on you.

The Resources Support Section Chief's desk: an incoming call saying all access roads are now impassable, tiles for Decide now, Escalated to you and For awareness, three escalations, and the reading pane showing the ask and why now.

Your desk. A caller from the field on the line, three escalations, and the reading pane for the one you picked.

Running beside every room is a hollow run, a second simulation nobody acts on, so the debrief can set the world you got against the world you would have gotten by doing nothing. Findings come out in three lanes that are never added together: the people, the AI crews, and the plan. People's findings and the plan's can be filed to Improve. The crews' findings go to a benchmark and never become a gap in your record, because a machine's mistake is not your corrective action. There's an Experiments layer that runs the same incident with people and with crews under identical, locked and hashed conditions and measures the difference seat by seat and hour by hour, with outside reviewers and a kill switch, and you can register your own agent to play a seat. Any finished exercise can be run in an Operation Center with one click: its modules become operational periods, its injects go to the seats its stress map names, its discussion questions become decision points. Synthetic incidents only. Simulate Ops launches at IAEM26, and you can take a seat and try it there.

Watching the Mass Care Coordinator, an AI seat: the strip of every seat in the room, what the seat sees on its desk, and its reasoning step by step, with buttons to tell it something, pause it, or take over the seat.

Watch a seat. What an AI crew sees, what it decided and why, and a button to take the seat yourself.

Rehearse Decisions with SimCell is the just-in-time rehearsal I described two weeks ago. Describe a real situation, or start from a scenario starter, or let Preppr draft one from the news and your knowledge base. Name who responds, up to twelve seats, kept exactly as you wrote them. Give it an inject outline, typed or lifted from an exercise's MSEL. Check the plans before the run and it lists what they don't settle. Then it runs each decision you're weighing against a hold baseline and shows you the diff: where the plan parted from what happened, what the organization could not carry out when a plan asked the world for something it couldn't do, which tactics only worked on some paths. You can park every adversary move for review and accept, edit or withdraw it. You can continue from any path with the world state and the seats' memory carried forward. You can sweep one setting across its values and find the tipping point, with a verification re-run before the report is allowed to call it a cliff. It never recommends a decision. That sentence is in the help text and I intend to keep it there. What a rehearsal finds is filed to Improve as synthetic evidence that can corroborate a pattern but never create one.

The decision paths step of a rehearsal: each path runs as its own cell and is described as an option to compare, not a recommendation, with a hold baseline, seats that can voice concerns but never change the decision, and an option to review each Red inject before it lands.

The last step before a rehearsal runs. The paths are options to compare, and the hold baseline is on by default.

Adaptive Simulations now play graph-first with fog of war: the steps you've visited solid, the next hop as untitled options, everything beyond as silhouettes. Preppr is the game master. It maps your written or spoken response to a route and judges it pass, partial or fail with a rationale, citing the design brief, the exercise's sources, your knowledge base and the physics grounding stored at build, and a facilitator can confirm or override every verdict live. SimCell is where all of this started: Red, White and Blue cells in parallel runs that stress a plan before people do.

Four products, one principle. The model plays the world and the crews. The engine owns the truth. You keep the decision, and the clock makes you keep it on time.

Video, and the fence I was on

Preppr can now generate video, and I was on the fence about it for most of a year.

Two reasons. The first was quality. I did not want to put a clip in front of a room of fire chiefs that looked like a dream sequence and made them laugh. An exercise lives or dies on whether the room takes it seriously, and one bad clip would cost more than a hundred good injects earned. The second reason was the opposite worry. If the quality got good, we'd be handing emergency managers a machine that makes convincing footage of disasters, and I have spent three years telling this field that the most dangerous thing about AI is believing it. Our own demo exercise has a presidential deepfake in it as an inject. I was not eager to build the thing that makes the inject.

The quality question answered itself this year. A short shot of a flooded intersection at night, with a narrator and the right ambient sound, now reads the way a news clip reads. That meant the second worry was the only one left, and the design followed from it.

Here is how the Exercise Media Studio works. Every scenario, module, inject and question in the designer has a Media button. A director agent reads the exercise and proposes a shot, a narration and a sound cue as priced cards with editable prompts. Nothing renders and nothing is charged until you click Generate on a card with its credit price on it. Corrections like make it night edit the clip you have rather than starting over. Every attempt is kept as a version. You pick one picture, one narration and one sound, set their volume and timing, and Preppr merges them into one clip with captions. Approve and attach, and delivery plays it to participants at that moment of the exercise. The same pipeline makes clips for the moments that matter in an Operation Center: the opening, decision points, phase changes and severe injects, each suggested by Preppr and ordered by the host, one at a time or all at once. Nothing renders unless a person asks for that clip. Media is billed at a premium rate you see before you generate, and it's on for selected accounts.

And every frame of it is labeled. Every generated clip and still plays under a persistent AI-generated, for training label. It is not a tooltip and it cannot be dismissed. Images Preppr generates carry a synthetic, exercise use only strip burned into the picture and the same disclosure written into the file's metadata, so even if someone crops the strip out, the file still says what it is. Anything that could pass for an official product, a mock news screenshot or a map, gets a diagonal watermark across the whole image on top of that. The rules every prompt is written under are not negotiable: training-appropriate only, no injuries, no real or identifiable people or officials, no real logos or news-network branding, no real addresses, preset voices only, and never anything designed to pass as real footage outside the exercise. Someone will ask whether the label can be turned off for a more immersive experience. No. The label is the feature.

The loop you can see, and the page that closes it

Every post on this blog eventually says the same thing: preparedness resets to zero after every exercise, and the Preparedness Cycle was always drawn but never ran. The two pages I'm proudest of this month are the ones that let you check whether that's still true of you.

The first is a card on the dashboard called your preparedness loop, which takes over from the workflow map once setup is done. Each run you finish closes a pass: a tabletop, a SimCell rehearsal, a Simulate Ops run, an adaptive simulation. Each pass is a row of cells, one for everything that fed it since the run before. Each document you added. Each question you asked Preppr. Each intel pull, each campaign, each design. Then the run itself as the deep tile at the end. Pass four, delivered October 21. Pass five, in progress since October 23. It uses the same visual grammar as the rest of the product now: bright green is what changed since last night, warm yellow is a gap or a contradiction, a ring is something a person set by hand, a dashed cell is a source Preppr hasn't read. Above it sit what Preppr has read, your readiness, one cell per capability that lights when a run tests it, and what Preppr noticed last night. There's one button, which opens a conversation already seeded with where you stand: since the chlorine release tabletop on the twenty-first you've added two documents, an Ask Preppr session and a new design, so design the next exercise around what they changed. Or, more pointedly: nothing new has come in since that run, add findings or documents, or schedule the next one. Nobody has to remember to run the loop. The loop tells you it's waiting.

The Your preparedness loop card: five passes, each a row of cells for the documents, Ask Preppr sessions, intelligence sessions and designs that fed it, ending in a darker cell for the run that closed it, with the fifth pass still in progress.

Four passes closed and a fifth under way. Every cell is something that fed the run at the end of its row.

The second is Improve, which is where everything lands. Across the top, the numbers: exercises run, including the ones we only know about from after-action reports you uploaded, exercises designed, capabilities exercised, findings, priority gaps. Then capability coverage, one panel per framework you answer to, planned against tested, laid out as a map by mission area with one cell per capability. Hover a cell and it names the exercise that tested it. Click a grey one and it starts designing an exercise around it. A strip underneath shows how coverage has built up quarter by quarter, which is the first picture of compounding I've been able to put in front of a leadership team.

The top of Improve: four exercises run of five designed, eight of 32 FEMA core capabilities exercised, four open priority gaps, the capability coverage map with one cell per capability in mission-area rows, and how coverage has built up by quarter.

Improve, filtered to FEMA Core Capabilities. One cell per capability, and the quarter-by-quarter build-up underneath. Then the priority gaps, each with its provenance: from an exercise, from an uploaded AAR, from a Collaborate campaign, from the people and plan lanes of an Operation Center, or from a rehearsal, marked as synthetic. Re-synthesize them, add your own, close one with Ask Preppr, or send the whole list off to develop an improvement plan. Beside that, the hazards in your intelligence feed you have never exercised, and your knowledge base's health, where an aging plan has a Review button that asks Preppr what's out of date and an after-action report has an Exercise button that turns its open corrective actions into the next exercise. And as of this week, you can just ask. Ask Preppr reads your improvement status and your preparedness metrics directly, so a dashboard of current improvement actions is a sentence, not a search.

I still say no to the AAR button. Improve is what we built instead of it. Not a document a machine wrote and nobody argued with, but a living list of what's been tested, what hasn't, and what's open, where every item says where it came from and hands off to the next run.

Giving some of it away

Ask Preppr's Actions are now called Skills. Only the name changed: the THIRA, the CPG 101 plan evaluation, the after-action input analysis and the rest are right where they were. Two things around them are new. Your own AI assistant can now run them, whether that's Claude, ChatGPT, Copilot or anything else connected to Preppr through MCP, asking you the same questions Ask Preppr would and producing the same deliverable. And we've started giving them away.

There is now a free, public Preppr Skills repository on GitHub. The first skill in it is the one I've written about most: alert and warning template generation, built with researchers and risk communicators. It walks you through five alert types, Imminent Threat, Public Safety, AMBER, Missing Person and Developing Situation, each asking for what that type needs in the order the public expects to read it. Before anything is written, it checks your answers against plain-language, certainty and content-order standards, and anything bound for IPAWS or Wireless Emergency Alerts has to carry a clear protective action or it comes back to you to fix. Then it writes templates only for the channels you actually use: 90 and 360 characters for WEA, mass notification, email, voice and social. It installs as a native skill in Claude, GitHub Copilot, Gemini CLI and OpenAI Codex, or as a Gemini Gem or a ChatGPT custom GPT. It needs no Preppr account. It's licensed CC BY-SA 4.0, so use it, change it, and share what you change.

Why give it away? Both reasons are in the last post. Access is settled. Practitioners are already using general-purpose AI because it came with the license, and the real question is whether what they use is grounded in anything. If you're going to draft alerts in a general-purpose model anyway, I'd rather you did it inside a structure built from the research than from a blank prompt. And the measured improvement in alert quality has come from structure: templates written in advance, with experts reviewing, not live generation at incident tempo. A skill that builds the template before the incident is the version of this use case I think has earned its place. The problems in this field are bigger than any one company. This is a small way to act like it.

The fine print

Simulate Ops launches at IAEM26, where you can try it in person. Rehearse, the Media Studio and the preparedness loop card are on for selected accounts while we watch them under real use. Operation Centers run synthetic incidents only. The Preppr Skills repository is free and public and needs no account. The screenshots in this post come from demo accounts with synthetic data. Generated media costs more than anything else in the product and says so before you spend. The changelog has the full account of what shipped, including the fixes, and it will keep having it.

The homepage says preparedness that compounds. This month is the first time I can point at a screen and show you yours doing it.