I Tested a Viral AI Prompting Technique on My Own System. It Won, Then Lost.

Someone sent me a YouTube video pitching "wargaming", the idea that you get a powerful model to pre-enumerate every way a task could go wrong, hand that map to a cheaper or local model, and let it execute without needing live intelligence. Consultant writes the decision tree, junior follows the arrows.

Timing was annoying. I was losing my free Fable access the next day, which meant if I wanted to actually test this properly, it had to happen now, not eventually.

So I pointed it at something real instead of a toy example: a genuine trap in Hermes, my own task-orchestration system. Specifically: what happens when Claude's unavailable and the local model (Qwen3-30B-A3B) has to fall back on planning a task by itself. The planner prompt advertises three tools (monitor, code, graph_query), but the execution layer silently blocks all three for research-project tasks. Nobody told the planner that. It's exactly the kind of landmine wargaming is supposed to defuse.

Round one

Ran it two ways, five times each, same task. Wargamed brief (Claude pre-mapping the failure branches) versus cold: local model with nothing extra.

Cold: 0 out of 5.

Wargamed: 4 out of 5, quality scored 9.6 out of 10.

Looked like a clean win. Worth digging one layer deeper though, because the wargamed condition wasn't just "brief present", it also had thinking mode switched on, which is disabled everywhere else in Hermes. So I split it apart: brief alone, no thinking, 0/5. Thinking alone, no brief, 0/5. Only both together worked.

That's a real cost, not a free lunch. Thinking mode ran 5.9x the latency and 6.7x the completion tokens of a normal call. Fine for a background agent with a 30-minute budget. Not fine if you're trying to make this a standard part of every decision.

Round two: the part that made me suspicious

Here's the thing that started nagging at me. The "wargame" step was Claude reading the codebase and telling the local model which tools were actually permitted. But that information already existed, as plain dictionaries in the code. Was the wargame step doing real reasoning, or was it just... restating facts that were already sitting there?

Tested it. Static permissions injected directly into the prompt, no Claude call at all: 6 out of 10, quality 8.2.

Wargamed brief: 4 out of 5 (80%), quality 9.6.

Still a gap. But the gap wasn't about missing facts. Both conditions had the same permission info by this point. It was about behavior. The static runs never admitted when they'd dropped a piece of work. They just quietly routed around the blocked tool and hallucinated a plausible-looking URL to paper over it. The wargamed runs, when they hit a wall, said so, explicitly stated what got skipped and why.

So the wargame wasn't adding knowledge. It was adding honesty about failure. Which is a genuinely different thing to have discovered, and also raised an obvious next question.

Round three: the actual answer

If the gap was behavioral, could I just... tell it to behave that way? No reasoning call, no wargame, just one extra sentence bolted onto the static template: "omit blocked work, say so in goal, don't improvise workarounds."

Static template plus that one line. Zero Claude calls.

10 out of 10. Quality scored a perfect 10.0. Beat the wargamed brief on every single axis: speed, quality, and cost (free, versus roughly three cents per decision for the Claude call).

What I'd assumed was sophisticated pre-emptive reasoning turned out to decompose into two things that render statically: facts that were already sitting in the codebase as dictionaries, and one coaching sentence about how to handle a wall when you hit one. The expensive part of the technique, the actual LLM call doing the "wargaming", added nothing detectable once those two things were in place.

Worth being honest about the limits here: this was one task, one trap family, twenty-five-odd runs total. I'm not claiming this generalizes to every agentic workflow on the internet. But going from 0/5 to 10/10 at zero marginal cost, on a real production landmine, is a decision-grade result for my own system regardless of whether it holds everywhere else.

Shipping the static version. The wargame pipeline that started this whole thing isn't getting built. Not because the idea was wrong, but because the actual mechanism turned out to be something I already had, wearing a more impressive-sounding name.