A ready-to-use prompt that turns any project brief into a war game: phases with optimistic/pessimistic assumptions, failure modes, verifiable exit criteria, and a master risk table — all before writing the first line of code.
It's a single text — in EN and PT-BR — that you paste at the start of a conversation with your planning AI before asking it to execute anything. There's nothing to install.
Each phase must state its objective, optimistic assumption, pessimistic assumption, failure modes and fixes, verifiable exit criteria, alternative path, and stop condition—before any code exists.
A single table covering infrastructure, model API specifics, frontend, conversation quality, security, and the learning loop, with detection and recovery for each scenario.
Design a weekly review that scans conversations for prompt injection and adjusts the agent's style within strict limits—no changes go to production without regression evaluation and explicit approval.
The logic behind the prompt: investing significant reasoning in planning buys low-to-medium effort—and much more predictable—execution.
Nothing beyond a briefing and a model capable of long-term planning — no installation, dependencies, or build.
A description of what you want to build—the more concrete (stack, integrations, project owner), the better the plan.
# what we're building "Um agente de chat com tour guiado + qualificação de lead..."
Choose the language and copy the corresponding file from the repo.
# EN cat doc/war-game-prompt.md # PT-BR cat doc/prompt-war-game-portugues.md
A model with strong reasoning (e.g., Claude Opus/Fable in "high" mode), preferably with access to research tools.
# paste the completed prompt como a 1ª mensagem # without prior context
The prompt was designed for a "cold-start agent" — it works best as the first message in a new conversation, with no prior project context.
The two files are equivalent — choose the one in the language you'll use to talk with the AI.
cat doc/prompt-war-game-portugues.md # or doc/war-game-prompt.md (EN)
Replace each [PROJETO], [Claude Haiku], [Supabase Edge Functions], [workshops/serviços] etc. with your project's actual details — model, infrastructure, stack, expected deliverables.
# before Plan a full build of [PROJECT], but do it as a war game. # after Plan a full build of the Prompt Advisers website chat agent, but do it as a war game.
No preamble, no extra context — the prompt already asks for everything the AI needs to know to plan from scratch.
nova conversa → colar o prompt preenchido # First message, no history
If the AI stops before delivering both formats requested in the prompt, explicitly ask for both.
1. plano em Markdown (preâmbulo, caminho crítico, stack, estrutura, metodologia) 2. versão visual em HTML com diagramas da arquitetura e de cada fase
Stop after each phase, check the exit criteria, and only then move on to the next. If you hit a stop condition, really stop: record the blocker instead of letting the AI improvise a way out.
fase 0 → checar critérios de saída → fase 1 → ...
The prompt above opens one project in depth. There is a second mode in the same repo for when the question is different: I have several stalled projects — which should I do first, which should I simplify, and which should I abandon?
One project, in depth. Answers "how do I build this without taking a beating along the way?" Delivers a phased plan, a master failure table, and a visual dossier in HTML.
Multiple projects, compared. Answers "where do I start?" Delivers an analysis for each project, a ranking by impact/effort/risk, and a 30- and 90-day plan.
Both use the same logic — action → reaction → counteraction — and the same rule: anticipate the obstacle before the first line of code. The difference is where the expensive model is used.
Never combine several into one file — that mixes context and ruins the comparison. Start with the 5-question minimum briefing; whatever you don't know, leave blank, don't guess.
analise-projetos/ ├── projetos/ # project-01.md, project-02.md, ... ├── resultados/ # the AI writes here ├── contexto-geral.md # what applies to everyone ├── criterios.md # the same criteria for everyone ├── ledger.md # blockers └── assumptions.md # assumptions, with the impact if they're wrong
These two files are what make different projects comparable. Shared criteria make ranking possible — project-specific criteria don't.
# criteria.md — the same for everyone substância clareza · problema real · público · valor · diferencial execução viabilidade · complexidade · gargalo · simplificação · automação risco o que impede · segurança · manutenção · escala retorno receita · custo
It generates a first version of all the projects — without polishing any of them. Polishing project 01 before looking at the rest burns the budget on the first in line, which is rarely the most important.
cat doc/war-game-massa-portugues.md # or doc/mass-war-game.md (EN)
The first round almost always comes out generic. The second is what removes the generic parts: look for contradictions, overlooked risks, steps without validation, and actions without stop conditions.
revisar /resultados → aprofundar war-game → atualizar ledger + suposições
Analysis can be done at scale. Execution can't. Read the resumo-geral.md, choose the top priority and deliver its war game with the execution prompt.
modelo avançado pensa e antecipa modelo intermediário organiza modelo barato executa # following the plan that's already laid out
O ledger.md e o assumptions.md are what keep the analysis from turning into fiction: without them, the model makes up whatever is missing so it can continue. Missing information becomes [PRECISA_DEFINIR: x] at the exact point where it would be needed.
Want to see where this leads? The sample output includes a fictional portfolio of 4 projects analyzed from start to finish — a complete individual analysis, o overall summary with ranking and a 30/90-day plan, plus the ledger and assumptions. This is the definition of "done" for this mode.
The two HTML files below (doc/war-game.html in EN and doc/war-game-portugues.html in PT-BR) are the actual output of the prompt applied to the PA-CHAT-V0 case — a chat agent with Claude Haiku + Supabase — and serve as a reference for what "done" looks like.
The prompt is stable in both languages. Possible next steps, with no deadline commitment.