# The Wargame Protocol: Everyone Ships Their Plan. Nobody Attacks It First.
I used to launch every plan straight at production — no dress rehearsal, no resistance, nothing.
Every single blowup was something I could have caught on paper, before it cost real hours and
real trust. So now nothing I build ships until it's been through five passes designed to break
it before reality does. This page hands you The Wargame Protocol — the exact 5-pass process,
the 8-criteria checklist that gates every plan, and a worked example showing all five phases
finding real failures in a toy plan.
The kit is free. On Instagram you can also comment WARGAME and the bot will DM it to you;
everywhere else the download button above is it.
Why a plan that "sounds done" isn't done
A plan that's only been read once, by the person who wrote it, has never met resistance. It reads
as finished because nothing has tried to break it yet — not because it's actually solid. The gap
between "sounds right in the doc" and "survives contact with a real, weaker executor" is exactly
where builds die: the missing file nobody checked for, the step that's obvious to the planner and
a guess to everyone else, the rollback that fires on a healthy deploy because nobody asked "what
if the checker itself is broken." None of that shows up from reading a plan. It only shows up
from attacking one.
The full step-by-step
1. RECON — kill ghosts before writing a move. Check every "should exist" against the real
system: real paths, real files, real interfaces. Every assumption the plan quietly made gets
verified live. A plan written on one unverified assumption fails at execution time, not at
review time — that's the single most expensive failure class this catches.
2. STYLE-BIBLE — copy, never paraphrase. If the plan extends something that already exists,
pull verbatim excerpts of the real code/config it has to match. Paraphrase is exactly where
drift enters when a cheaper or faster executor runs the plan later.
3. MISSIONS — write every step with a fail case. Every move gets an exact command, an
observable EXPECT (what success actually looks like), and a FAIL→ counter-move for the most
likely way it breaks. No step is left for the executor to improvise.
4. RED-TEAM — attack it on purpose. A fresh reviewer with no memory of writing the plan tries
to break it: wrong assumptions, order dependencies, irreversible steps with no gate. Patch,
log, repeat until an attack pass finds nothing material.
5. EXECUTOR-SIM — hand it to a weaker model, blind. A fresh, mid-tier model reads the plan
with zero access to your reasoning and reports every guess it would have to make, every step
that would error, and a final yes/no: would this succeed, run blind, with no hand-holding?
The 8-criteria gate
A plan is wargame-ready only when all eight of these hold — the SUCCESS.md checklist in the kit,
copy-once-adapt-never:
| # | criterion |
|---|---|
| 1 | Every load-bearing fact was verified live, not assumed |
| 2 | Every style reference is a verbatim excerpt, not a paraphrase |
| 3 | Every move has an observable EXPECT and a named FAIL→counter-move |
| 4 | Every data transform ships as tested code with a real measured result |
| 5 | Every irreversible step has an explicit ABORT condition before it |
| 6 | Every mission has a fresh-checker VERIFY loop with a capped round count |
| 7 | A red-team pass found nothing material on its last round |
| 8 | A mid-tier model, reading it blind, reports it would succeed |
Criterion 8 is the one that actually matters — everything above it exists to make that one true.
What most people get wrong
They review their own plan one more time and call that "checking it." Rereading isn't attacking.
The whole method only works because each phase is a genuinely different adversary: RECON attacks
assumptions, RED-TEAM attacks the system, EXECUTOR-SIM attacks the executor. Skipping straight to
"looks good to me" is exactly how a plan that reads as finished still explodes on first contact
with reality.
What's in the kit
- wargame-SKILL.md — the drop-in 5-pass process, sanitized and ready for any Claude-based
agent setup
- The 8-criteria SUCCESS.md checklist printed inside it, copy-once-adapt-never
- worked-example.md — a real toy plan run through all five phases, with the exact 3 failures
each phase catches
Pay the one skill install once, and every plan gets stress-tested before it can fail in
production, forever. Comment WARGAME on the video or grab the kit above — it's free either
way.