O Journal
Notas sobre modelos, protocolos e cobranças transparentes
Notas de trabalho para quem desenvolve com modelos de linguagem: como os agentes se comportam, qual modelo escolher para cada tarefa e como se compõe uma cobrança por token.
DeepSeek V4 Flash 与 GPT-6.1 Sol 摘要对比实操
从版本核验、样本准备到盲评和成本核算，逐步比较两款模型的摘要表现，避免把接口别名、文字流畅度或宣传指标误当成可靠结论。
Qwen3.7 Plus 大规模批处理：从试跑到可靠交付
分步搭建 Qwen3.7 Plus 批处理流程，覆盖接口核验、数据分片、思考预算、任务提交、失败重试与成本对账，并区分官方批量接口和中转服务能力。
用 Haiku 级小模型搭建可靠分类流水线
从标签定义、评测集和结构化输出，到异常分流、成本核算与灰度上线，逐步搭建可验证、可回滚的小模型分类流水线。
哪个模型最擅长处理脏乱电子表格：六步实测指南
从合并表头、混合日期到跨表公式，逐步建立可复现的测试流程，用正确率、结构保留和实际成本选择适合处理脏乱电子表格的模型。
Gemini 4 Argon 与 Claude Opus 5：分步完成工程选型
从输入能力、代码修复、工具调用到缓存账单，逐步比较 Gemini 4 Argon 与 Claude Opus 5，用可复现测试确定适合项目的模型。
逐步算清：何时值得为Opus级模型的额外质量付费
从验收标准、对照评测到缓存账单与升级策略，逐步计算Opus级模型能否通过减少返工、重试和人工复核收回额外成本。
Claude Sonnet 5 与 GPT-6 Astra：结构化输出对比实操
从统一数据契约、配置原生接口到异常处理与成本核算，逐步搭建可复现的结构化输出对比流程，以业务正确率而非格式观感选择模型。
客服机器人该选哪个模型：从真实工单到上线验收
按任务拆分、模型初筛、真实工单评测、工具验证、成本核算与灰度上线六个步骤，为客服机器人选择合适的模型。
代码生成与代码审查如何选模型：六步建立可靠流程
从任务拆分、真实仓库评测到缺陷验证和成本核算，逐步为代码生成与代码审查选择合适模型，避免只看榜单或单次报价。
Building Small, Quick Assistants on MiniMax M3, One Step at a Time
Build a focused MiniMax M3 assistant with bounded context, controlled reasoning, safe tools, and measurable latency and cost.
Qwen3.8 Max vs Closed Models: A Practical Walkthrough
Compare Qwen3.8 Max with frontier closed models through benchmark checks, controlled coding tests, latency measurements, and cost analysis.
Grok 4.7 for Current Events: A Step-by-Step Guide
Build a repeatable Grok 4.7 workflow for current events with explicit dates, verified sources, careful citations, and controlled API usage.
Gemini 3.8 Flash When Speed Matters More Than Depth
A practical walkthrough for configuring, testing, and deploying Gemini 3.8 Flash when responsiveness matters more than exhaustive reasoning.
DeepSeek V4 Pro: A Mathematical Reasoning Walkthrough
A practical walkthrough for solving mathematics with DeepSeek V4 Pro, configuring reasoning, checking answers, and tracking API usage.
Gemini 3.1 Pro for Document-Heavy Work
Build a repeatable Gemini 3.1 Pro workflow for document intake, evidence extraction, cross-document analysis, validation, and cost tracking.
GPT-5.6 Sol vs GPT-6.1 Sol: Everyday Prompts
A practical walkthrough for comparing GPT-5.6 Sol and GPT-6.1 Sol on writing, summaries, planning, and code without inventing benchmark results.
GPT-6.1 Sol in Practice: Strengths and Limits
A step-by-step guide to evaluating GPT-6.1 Sol for coding, tool use, reliability, latency, and real production costs.
What Claude Fable 5.1 Is Good At: A Walkthrough
Explore Claude Fable 5.1 through practical steps for coding, research, document work, tool integration, and quality evaluation.
GPT-6 Astra: A Careful Reasoning Walkthrough
A practical, step-by-step workflow for GPT-6 Astra: define goals, organize evidence, choose reasoning effort, verify results, and track costs.
How to Keep API Keys and Passwords Away From Coding Agents
A practical walkthrough for isolating credentials, restricting agent access, sanitizing tool output, and testing your security boundaries.
Claude Code or Codex for Your Backend Service?
A practical walkthrough for comparing Claude Code and Codex on backend correctness, permissions, review effort, and total cost.
Gemini CLI and Claude Code on the Same Task
A step-by-step walkthrough for comparing Gemini CLI and Claude Code on one Python task, with isolated worktrees, shared tests, and cost tracking.
Using a Coding Agent for Database Migrations
Walk through a PostgreSQL migration with a coding agent, from schema inspection and safe batching to validation, deployment, and recovery.
Using Codex CLI for Everyday Shell Chores
Build and verify a safe cleanup script with Codex CLI, from defining the task to testing filenames, dry runs, and repeatable execution.
Reviewing Coding Agent Changes Line by Line
A practical walkthrough for reviewing coding-agent changes carefully, from diff scope to tests and follow-up notes.
Turn Long Coding Agent Sessions Into Checkpoints
A practical walkthrough for dividing coding agent work into verifiable checkpoints, preserving decisions, and resuming safely after interruptions.
A Reliable Coding Assistant Documentation Workflow
A step-by-step workflow for using coding assistants to update technical documentation accurately and safely.
Use Gemini CLI to Navigate Large Repositories
A practical walkthrough for asking Gemini CLI focused repository questions, tracing behavior across files, and checking answers against code.
Make Your Coding Agent Explain Every Diff
A practical walkthrough for getting a coding agent to explain changes, prove its claims, and expose risks before you merge.
Give a Coding Agent a Bug It Can Fix
Build an actionable bug report with a reliable reproduction, clear evidence, bounded scope, and checks that prove the fix works.
Codex in VS Code: A Careful First Setup
Configure Codex in VS Code with deliberate authentication, limited permissions, verified provider settings, and a small first task.
Debug a Failing CI Job With a Coding Agent
Follow a practical workflow to reproduce a CI failure, guide a coding agent toward the cause, and verify a focused fix.
How Claude Code Decides Which Files to Read
Follow Claude Code from your prompt through project instructions, file searches, targeted reads, and verification-driven exploration.
Gemini CLI Setup With a Native-Protocol Base URL
Configure Gemini CLI with a native Gemini API relay, select API-key authentication, test routing, and troubleshoot common setup mistakes.
Write a CLAUDE.md That Changes Agent Behaviour
Build a focused CLAUDE.md with actionable rules, verified commands, scoped instructions, and repeatable tests of agent behaviour.
Run Codex CLI Through an OpenAI-Compatible Relay
Configure Codex CLI for BestRelay, protect your API key, verify Responses API compatibility, and troubleshoot your first request.
用 Claude Code 从头到尾审查一次拉取请求
从确认审查范围、读取完整上下文，到复现缺陷、整理评论与复查提交，逐步建立有证据、可追溯的 Claude Code 拉取请求审查流程。
Quer experimentar os modelos apresentados aqui? Cadastrar