Docs.

Wargamer mode walkthrough, FAQ, and links to the open-source framework.

Quick start

1. Subscribe from the landing page CTA (Stripe checkout).
2. Open hammerstein.ai/wargamer.
3. Paste three inputs: your board state (positions, units, which side you play); your status report (turn number, recent moves, what's troubling you); and your turn question (what you're trying to decide).
4. Press Issue orders. The response appears below the form, usually within 10–30 seconds.

What makes a good board state

• Include unit positions in whatever notation fits your game — chess: piece + square; hex-and-counter: unit + hex coordinate; abstract games: describe what each side controls.
• State clearly which side you are playing.
• Note constraints that aren't obvious from the layout alone: zone of control, supply lines, morale, special scenario rules, and so on.
• Don't paste the whole rulebook. The framework reasons from doctrine; it is not a substitute for looking up every corner-case rule.

What makes a good status report

• Turn number and what happened last turn (or the last few exchanges that matter).
• What you are trying to accomplish strategically — objectives, victory conditions, tempo.
• What is threatening you and what you suspect about the opponent's plan.
• Be honest about where you feel stuck or uncertain; that is where the orders tend to be most useful.

What makes a good turn question

• Ask for a specific decision, not an open-ended “what do I do?” For example: “Where should I commit my reserve this turn?” or “Should I push the flank or hold the center?”
• Bound the scope: “this turn,” “the next two moves,” or a similarly concrete window. The output is framed as near-term orders, not a multi-session grand campaign plan.

Example query

Below is a minimal chess midgame example using the same shape as the smoke-test paste: three fields filled, then a short note on what comes back.

The reply is typically framework-shaped: a short structural read of the position (what matters tactically and positionally), then specific moves or lines phrased as orders you could give, plus counter-observations — what to watch for if the opponent deviates. Tone and structure follow the Hammerstein / Auftragstaktik register used in Wargamer mode.

FAQ

What model is behind Wargamer mode?

Anthropic Claude Sonnet 4.6 via OpenRouter, running with the Hammerstein system prompt and corpus from the open-source framework repo.

Why isn't there a board-photo upload?

The hosted MVP is text-input only. Image input and persistent campaign memory are full-product features — self-host the open-source product to use them, or wait for the next iteration of the hosted site.

Does it remember previous turns?

On the hosted site, each query is independent for now — the browser-local Session log shows your demo session's turns but doesn't replay them as context to the model. Persistent campaign memory is a full-product feature.

What if it gives me bad orders?

Email ray@hammerstein.ai with your inputs and the model output. I'm the founder and I read everything; bug reports and bad outputs directly inform what gets fixed next.

Can I run the model locally instead?

Yes. The framework and distilled 7B weights are open source: github.com/lerugray/hammerstein and huggingface.co/lerugray/hammerstein-7b-lora. The hosted product runs the same framework on frontier Sonnet for higher quality and zero local setup.

What's the difference between the hosted Wargamer mode and the open-source hp CLI?

Wargamer mode uses Sonnet 4.6, the Hammerstein system prompt, and a kriegspiel–Auftragstaktik output register tuned for wargame orders in the browser. The CLI is a general-purpose Hammerstein wrapper: you supply your own API key and choose whichever model you want to call.

Is my data stored?

Cloudflare Pages Functions do not persist Wargamer queries to a database in the MVP. The Session log on the Wargamer page is stored only in your browser's localStorage and never sent to the server. OpenRouter processes requests under their privacy policy. We don't sell data, train models on user input, or share queries with third parties beyond what OpenRouter needs to route the request.

How do I cancel?

Through Stripe's customer portal; the link is in your subscription confirmation email. A customer-portal button on the landing page is planned for the next iteration.

Open-source references

• github.com/lerugray/hammerstein — framework, corpus, and benchmarks.
• huggingface.co/lerugray/hammerstein-7b-lora — distilled 7B local model (QLoRA).
• In the framework repo, the eval/ directory documents behavior characterization runs, including RESULTS-v0.1.md, RESULTS-v0.4.md, RESULTS-gibson-v3-2026-05-11.md, RESULTS-tuesday-battery-2026-05-11.md, and RESULTS-fra-2026-05-11.md.