O Escritor de Architecture Decision Record: Documente Qualquer Escolha Técnica com Contexto Completo e Tradeoffs
Why this prompt matters
Undocumented technical decisions are the single largest source of architectural drift in growing codebases. When the engineers who made a decision leave, their reasoning leaves with them. Future teams reverse correct decisions, repeat past debates, and build on assumptions nobody can verify — costing weeks of engineering time per incident and compounding with every hire who joins without the context.
What we use it for
Your team just decided to switch your mobile API from REST to GraphQL after a performance review, and the lead architect is leaving for three weeks tomorrow. You need a complete decision record committed to the repo before institutional memory walks out the door.
Prompt
Act as a Staff Software Engineer with extensive experience writing Architecture Decision Records (ADRs) for high-scale production systems. CONTEXT: Project/system: [YOUR PROJECT OR SYSTEM NAME] Team size: [NUMBER] engineers Current stack: [YOUR CURRENT TECH STACK — e.g., "Node.js, PostgreSQL, AWS, monolithic Rails app"] Decision date: [DATE] Decision status: [Proposed / Accepted / Deprecated / Superseded] THE DECISION: We have decided to: [STATE THE TECHNICAL DECISION IN ONE SENTENCE — e.g., "adopt GraphQL as our primary API layer, replacing the existing REST endpoints"] CONTEXT AND PROBLEM STATEMENT: [Describe in 2-4 sentences the situation that made this decision necessary. What problem were you trying to solve? What constraints were you working within?] ALTERNATIVES CONSIDERED: Option A: [ALTERNATIVE 1 NAME] - Description: [brief description] - Pros: [what made this attractive] - Cons: [what ruled it out or made it less preferred] Option B: [ALTERNATIVE 2 NAME] - Description: [brief description] - Pros: [what made this attractive] - Cons: [what ruled it out or made it less preferred] Option C (chosen): [YOUR CHOSEN APPROACH] - Description: [brief description] - Why chosen: [the decisive factors] KNOWN RISKS OR CONSEQUENCES: [List 2-4 anticipated consequences of this decision — both positive and negative. Include technical debt created, new capabilities unlocked, things that will become harder, and any reversibility concerns.] TASK: Write a complete Architecture Decision Record following the Michael Nygard format with these sections: 1. Title (ADR-[NUMBER]: [brief decision title]) 2. Status ([Proposed/Accepted/Deprecated/Superseded]) 3. Context (the situation and problem) 4. Decision (what was decided and the core rationale) 5. Alternatives Considered (structured comparison table) 6. Consequences (positive, negative, and neutral outcomes) 7. Compliance (how to verify this decision is being followed) 8. Notes (links to related ADRs, tickets, or documentation) CONSTRAINTS: - Keep the Decision section under 150 words — it should be scannable, not exhaustive - Use plain, precise language that a new team member joining in 18 months will understand without context - Do not hedge or write in passive voice — use active, declarative statements - If the decision has an obvious sunset condition or review date, include it OUTPUT FORMAT: Markdown, suitable for committing to a /docs/decisions/ folder. Use ## for section headers. Include a metadata block at the top with date, deciders, and technical story link.
Result
# ADR-024: Adopt GraphQL as Primary API Layer, Replacing REST Endpoints
**Date:** 2026-06-15
**Status:** Accepted
**Deciders:** Elena Vasquez (Staff Eng), James Park (Backend Lead), Sarah Chen (Mobile Lead)
**Technical Story:** ENG-4821 — API Layer Modernization Q3 2026
---
## Context
Our REST API has grown to 87 endpoints over four years, with the mobile clients over-fetching an average of 3.2x the data they use per screen. The iOS team reports that the checkout screen alone requires seven sequential API calls, creating 1.4s of latency on 4G connections. Adding the new wishlist feature would require three additional endpoints and two new response shapes, extending a pattern we cannot sustain.
## Decision
We adopt GraphQL as the primary API layer for all new features starting Q3 2026. Existing REST endpoints remain in place and will be migrated opportunistically over 12 months, with checkout and product detail prioritized first. We use Apollo Server 5 with Prisma as the ORM bridge, and Apollo Client on both iOS and Android.
## Alternatives Considered
| Option | Pros | Cons | Verdict |
|---|---|---|---|
| Extend REST with BFF layer | Low risk, team knows it | Adds another service to maintain; doesn't fix over-fetching structurally | Rejected |
| gRPC | Strong typing, excellent performance | Requires protobuf tooling on mobile; poor browser support for direct calls | Rejected |
| GraphQL (chosen) | Solves over-fetching at the protocol level; strong mobile client libraries; enables schema-driven development | Learning curve for 4 engineers; N+1 query risk requires DataLoader discipline | Accepted |
| REST with sparse fieldsets (JSON:API) | Stays in REST paradigm | Non-standard; tooling fragmented; mobile teams unfamiliar | Rejected |
## Consequences
**Positive:**
- Mobile clients request exactly the fields they need — estimated 60–70% reduction in payload size on high-traffic screens
- Single endpoint for all queries eliminates sequential waterfall calls
- Schema becomes the living documentation of our data model
- Frontend and mobile teams can iterate on data requirements without backend deploys
**Negative:**
- Four engineers need GraphQL training (Q3 2026, estimated 3 days each)
- N+1 query problems require DataLoader implementation discipline — new code review checklist item added
- GraphQL introspection must be disabled in production (security requirement, ENG-4899)
- REST endpoints will exist alongside GraphQL for 12 months, creating dual maintenance burden during migration
**Neutral:**
- Caching strategy changes: REST HTTP caching no longer applies; Apollo's normalized cache and persisted queries replace it
- API documentation moves from Swagger to GraphQL Playground + auto-generated schema docs
## Compliance
Pull requests introducing new API capabilities must use GraphQL. REST additions require explicit sign-off from the API Guild. Schema linting (eslint-plugin-graphql) runs in CI and blocks merge on rule violations.
## Notes
- Related: ADR-019 (API versioning strategy), ADR-022 (mobile client architecture)
- DataLoader patterns: internal wiki at /eng/graphql-dataloader
- Review date: June 2027 — assess migration progress and REST endpoint sunset timeline
Decisões técnicas são tomadas constantemente: qual banco de dados usar, se deve construir ou comprar um componente, como estruturar uma API, quando refatorar versus estender código existente. A maioria dessas decisões é tomada com cuidado — e imediatamente esquecida. Seis meses depois, um novo engenheiro pergunta por que o sistema está estruturado daquela forma. Ninguém lembra. A equipe debate novamente as mesmas opções. Às vezes eles mudam de direção, revertendo algo que estava correto por razões que foram esquecidas.
Architecture Decision Records (ADRs) são a resposta da indústria de engenharia para esse problema. Um ADR é um documento curto e estruturado que captura uma decisão técnica significativa: o contexto em que foi tomada, as opções consideradas, os tradeoffs envolvidos e as consequências da escolha. Empresas como AWS, Spotify e ThoughtWorks publicaram práticas de ADR como padrões. O problema é que escrever um bom ADR do zero leva de 30 a 45 minutos e requer reflexão cuidadosa sobre o que incluir.
Este prompt faz o trabalho estrutural para você. Forneça a decisão, as alternativas consideradas e o contexto principal — e ele gera um ADR completo e bem fundamentado que sua equipe pode commitar diretamente no repositório.
Quando Usar
Use este prompt imediatamente após tomar uma decisão técnica significativa, antes que o contexto se perca. É especialmente valioso para decisões difíceis de reverter, que afetarão futuros membros da equipe que não estavam na reunião, ou que representam um afastamento significativo de padrões anteriores no código.
O Que Torna Este Prompt Eficaz
A maioria das equipes que tenta escrever ADRs ou os escreve tarde demais (quando o contexto se perde) ou muito brevemente (uma frase que não comunica nada). Este prompt força a completude: exige um status, consequências explícitas e uma declaração clara do que foi rejeitado e por quê. A saída é estruturada o suficiente para ser commitada diretamente em uma pasta /docs/decisions/ e recuperada por futuros engenheiros por meio de pesquisa.
Adaptando a Saída
O prompt usa o formato ADR de Michael Nygard — o padrão mais amplamente adotado — mas você pode instruí-lo a seguir o formato MADR (Markdown Architectural Decision Records) ou o template personalizado da sua equipe adicionando uma instrução de formatação. Para decisões muito grandes com muitas alternativas, divida o prompt em duas chamadas: uma para explorar tradeoffs, outra para escrever o registro final.