OpenLog Skill
Aodusk-Future/openlog-skill
Diário de dev por módulo — arquitetura, log e breakpoint em <1k tokens
Ver no GitHubUma skill para Claude Code e Codex CLI que cria um diário de desenvolvimento por módulo, registrando arquitetura, log de mudanças e ponto de interrupção atual. Permite que agentes e sessões futuras retomem o trabalho em menos de 1k tokens, eliminando reexploração de código e derivas arquiteturais.
Instalação
npx convez-ai Aodusk-Future/openlog-skill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Retomar trabalho em projetos grandes sem reexploração de contexto
- Manter histórico arquitetural e decisões de design centralizadas
- Facilitar handoff entre sessões de IA ou entre múltiplos agentes
- Reduzir tokens gastos em re-leitura de código e derivação de estado
Veredito

OpenLog resolve um problema real em fluxos de trabalho com IA: handoff eficiente entre sessões e agentes em projetos complexos. Implementação clara e bem documentada, embora exija disciplina do usuário em manter os registros atualizados.
README
A tradução para o português deste README ainda está na fila.
OpenLog
A lightweight per-module development journal for Claude Code and Codex CLI — keep architecture, dev log, and current-stage notes for every module of a large project, so any future session or agent can pick up the work in under 1k tokens.
一个轻量级的"按模块开发档案"工具,同时支持 Claude Code 和 Codex CLI——为大型项目的每个模块沉淀架构说明、开发日志、当前断点,让未来任何会话 / agent 都能在 1k token 以内接手工作。
Why OpenLog?
When a project grows past a handful of modules, three things start to hurt:
- High handoff cost — a new session (or a new agent) has to re-derive "where did we stop?" and "why is it built this way?"
- Design / implementation drift — architectural decisions live in chat history and slowly disappear.
- Wasted tokens on re-exploration — every fresh context has to grep / read the same files to rebuild understanding.
OpenLog fixes this by writing the three pieces of context that matter — architecture, dev log, current breakpoint — into three fixed files per module, then standardising how Claude reads and updates them.
Quick Start
Works with both Claude Code and Codex CLI — they use the same SKILL.md format.
# For Claude Code
git clone https://github.com/Aodusk-Future/openlog-skill ~/.claude/skills/openlog
# For Codex CLI (identical workflow, same trigger commands)
git clone https://github.com/Aodusk-Future/openlog-skill ~/.codex/skills/openlog
Using both tools? Clone once to ~/openlog-skill/, then symlink to both locations so future updates stay in sync automatically:
git clone https://github.com/Aodusk-Future/openlog-skill ~/openlog-skill
mkdir -p ~/.claude/skills/openlog ~/.codex/skills/openlog
ln -s ~/openlog-skill/SKILL.md ~/.claude/skills/openlog/SKILL.md
ln -s ~/openlog-skill/SKILL.md ~/.codex/skills/openlog/SKILL.md
Then in any Claude Code or Codex session, type:
/openlog
…or just say something like "let's work on the Combat module" / "start the Combat module" — the agent will recognise the trigger and run the OpenLog workflow.
How It Works
Data layout (lives in your project root)
OpenLog/
├── INDEX.md # Minimal table of all modules + one-liner + status
├── <ModuleA>/
│ ├── ARCHITECTURE.md # Module architecture — kept up to date
│ ├── DEV_LOG.md # Dev log, newest entries on top
│ ├── DEV_LOG.archive/ # Auto-archive when DEV_LOG > 500 lines
│ │ └── YYYY-MM.md
│ └── CURRENT.md # Current breakpoint, <30 lines
└── <ModuleB>/...
The OpenLog/ directory is committed to your project's git, so multiple worktrees / multiple agents / multiple humans share the same journal.
Two-phase workflow
Startup phase (when you invoke the skill):
- Read
OpenLog/INDEX.md(the only full read). - Try to fuzzy-match the module name from what you just said.
- If matched → read ARCHITECTURE (full), DEV_LOG (first ~40 lines = newest ~5 entries), CURRENT (full). Output a 2-3 sentence handoff summary.
- If not matched → ask which module, or create a new one.
Wrap-up phase (only triggered by explicit user signal — "done", "wrap it up", "next task", /openlog commit, etc.):
- Prepend a new entry to DEV_LOG (timestamp + summary + key files + link to CURRENT items).
- Update CURRENT (move "in progress" items, log new blockers, refresh timestamp).
- Update ARCHITECTURE only if interfaces / dependencies / key files changed.
- Auto-archive old DEV_LOG entries when the file exceeds 500 lines.
Designed for low token cost
- Resuming a module: typically 700-1000 tokens of reads, no grep, no whole-codebase exploration.
- Creating a module: ~300 tokens.
- Wrap-up: ~150-400 tokens (ARCHITECTURE updates are skipped most of the time).
Customisation
The skill is a single markdown file at SKILL.md. Edit the rules, file templates, or wrap-up triggers to match your team's conventions.
A reference English-language version of the skill body is provided at SKILL.en.md. To use it, replace the YAML body of SKILL.md with the English content, or rename it directly.
License
MIT — do whatever you want with it; attribution appreciated but not required.
中文
为什么需要 OpenLog?
当项目模块数超过几个,就会遇到三个痛点:
- 接手成本高——新会话或新 agent 要重新摸清"上次做到哪""为什么这么设计"
- 设计与实现分裂——架构决策散落在对话里,没有沉淀
- 重复探索浪费 token——每次新上下文都要 grep / read 同一批文件重建理解
OpenLog 把"架构说明 + 开发日志 + 当前断点"沉淀到每个模块固定的三个文件里,并标准化 Claude 读写它们的方式。
快速上手
同时支持 Claude Code 和 Codex CLI——两者用相同的 SKILL.md 格式。
# Claude Code
git clone https://github.com/Aodusk-Future/openlog-skill ~/.claude/skills/openlog
# Codex CLI(工作流完全一致,触发命令也相同)
git clone https://github.com/Aodusk-Future/openlog-skill ~/.codex/skills/openlog
两个工具都用?建议 clone 一份到 ~/openlog-skill/,然后用 symlink 链到两个位置,以后更新只需 git pull 一次:
git clone https://github.com/Aodusk-Future/openlog-skill ~/openlog-skill
mkdir -p ~/.claude/skills/openlog ~/.codex/skills/openlog
ln -s ~/openlog-skill/SKILL.md ~/.claude/skills/openlog/SKILL.md
ln -s ~/openlog-skill/SKILL.md ~/.codex/skills/openlog/SKILL.md
之后在任何 Claude Code 或 Codex 会话里输入:
/openlog
…或者直接说"开始开发 Combat 模块" / "接手战斗模块"等,agent 会识别触发条件并走 OpenLog 工作流。
工作原理
数据布局(位于项目根目录)
OpenLog/
├── INDEX.md # 模块清单(极简表格 + 一句话定位 + 状态)
├── <模块A>/
│ ├── ARCHITECTURE.md # 模块架构(保持最新)
│ ├── DEV_LOG.md # 开发日志(倒序,最新在最上)
│ ├── DEV_LOG.archive/ # DEV_LOG > 500 行时自动归档
│ │ └── YYYY-MM.md
│ └── CURRENT.md # 当前断点(< 30 行)
└── <模块B>/...
OpenLog/ 目录跟随项目 git 仓库提交,团队 / 多 worktree / 多 agent 共享同一份档案。
两阶段工作流
启动阶段(唤起 skill 时):
- Read
OpenLog/INDEX.md(唯一允许的全量读) - 尝试从你的指令里模糊匹配模块名
- 匹配到 → 读 ARCHITECTURE(全文)+ DEV_LOG(前 40 行 ≈ 最近 5 条)+ CURRENT(全文),输出 2-3 句话接手摘要
- 未匹配 → 询问要开发哪个模块,或新建
收尾阶段(仅在用户明确发信号时触发——"做完了""收尾""下一个任务""/openlog 收尾"等):
- 在 DEV_LOG 顶部插一条新记录(时间戳 + 摘要 + 关键文件 + 关联 CURRENT 条目)
- 更新 CURRENT("进行中"移走 / 标记新阻塞 / 刷新时间戳)
- 仅当接口 / 依赖 / 关键文件变化时才改 ARCHITECTURE
- DEV_LOG > 500 行时自动归档老条目
为低 token 设计
- 接手已有模块:~700-1000 tokens,无需 grep,无需全代码库探索
- 新建模块:~300 tokens
- 收尾:~150-400 tokens(多数情况下跳过 ARCHITECTURE 更新)
自定义
整个 skill 就是一个 markdown 文件 SKILL.md,规则、模板、触发口令都可以按团队习惯改。
如需英文版 skill body,参考 SKILL.en.md。把 SKILL.md 的正文替换为英文内容,或者直接改名使用即可。
许可证
MIT——随便用,署名鼓励但不强制。