返回技能列表

write-continue-here

pjt222
更新于 2 days ago
8 次查看
17
2
17
在 GitHub 上查看
文档ai

关于

This Claude Skill creates a structured CONTINUE_HERE.md file to preserve session context for seamless handoffs between coding sessions. It captures the current state including objectives, completed work, in-progress tasks, and next steps when ending with unfinished work. Use it to maintain continuity where git alone falls short, ensuring fresh sessions can immediately resume productive work.

快速安装

Claude Code

推荐
主要方式
npx skills add pjt222/agent-almanac -a claude-code
插件命令备选方式
/plugin add https://github.com/pjt222/agent-almanac
Git 克隆备选方式
git clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/write-continue-here

在 Claude Code 中复制并粘贴此命令以安装该技能

技能文档

Write Continue Here

Write structured continuation file so next session starts with full context.

When Use

  • Ending session with work still in progress
  • Handing off complex task between sessions
  • Preserving intent, failed approaches, next steps that git cannot capture
  • Before closing Claude Code when mid-task

Inputs

  • Required: Active session with recent work to summarize
  • Optional: Specific instructions about what to emphasize in handoff

Steps

Step 1: Assess Session State

Gather facts about recent work:

git log --oneline -5
git status
git diff --stat

Review conversation context: what was objective, what completed, what partially done, what tried and failed, what decisions made.

Got: Clear understanding of current task state — completed items, in-progress items, planned next steps.

If err: Not in git repository? Skip git commands. Continuation file can still capture conversational context, task state.

Step 2: Write CONTINUE_HERE.md

Write file to project root using structure below. Every section must contain actionable content, not placeholders.

# Continue Here

> Last updated: YYYY-MM-DDTHH:MM:SSZ | Branch: current-branch-name

## Objective
One-paragraph description of what we are trying to accomplish and why.

## Completed
- [x] Finished item with key file paths (e.g., `src/feature.R`)
- [x] Decisions made and their rationale

## In Progress
- [ ] Partially complete work — describe current state (branch, file:line)
- [ ] Known issues with partial work

## Next Steps
1. Immediate next action (most important)
2. Subsequent actions in priority order
3. **[USER]** Items needing user input or decision

## Context
- Failed approaches and why they did not work
- Key constraints or trade-offs discovered
- Relevant issue/PR links

Guidelines:

  • Objective: Capture WHY — git log shows what changed, not why
  • Completed: Mark items clear done to prevent re-work
  • In Progress: Highest-value section — partial state hardest to reconstruct
  • Next Steps: Number by priority. Prefix user-dependent items with **[USER]**
  • Context: Record negative space — what tried and rejected, why

Got: CONTINUE_HERE.md file at project root with all 5 sections populated with real content from current session. Timestamp and branch accurate.

If err: Write fails? Check file permissions. File should be created in project root (same directory as .git/). Verify .gitignore contains CONTINUE_HERE.md — if not, add it.

Step 3: Verify File

Read back CONTINUE_HERE.md. Confirm:

  • Timestamp current (within last few minutes)
  • Branch name matches git branch --show-current
  • All 5 sections contain real content (no template placeholders)
  • Next Steps numbered and actionable
  • In Progress items describe current state specifically enough to resume

Got: File reads as clear, actionable handoff fresh session could use to immediately resume work.

If err: Edit sections containing placeholder text or too vague. Each section should pass test: "Could fresh session act on this without asking clarifying questions?"

Check

  • CONTINUE_HERE.md exists at project root
  • File contains all 5 sections with real content (not placeholders)
  • Timestamp and branch accurate
  • .gitignore includes CONTINUE_HERE.md
  • Next Steps numbered and actionable
  • In Progress items specify enough detail to resume without questions

Pitfalls

  • Write placeholders instead of content: "TODO: fill in later" defeats purpose. Every section must contain real information from current session.
  • Duplicate git state: Do not list every file changed — git already tracks that. Focus on intent, partial state, next steps.
  • Forget Context section: Failed approaches most valuable thing to record. Without them, next session will retry same dead ends.
  • Overwrite without reading: CONTINUE_HERE.md already exists from prior session? Read first — may contain unfinished work from earlier handoff.
  • Leave stale files: CONTINUE_HERE.md ephemeral. After next session consumes, delete. Stale files cause confusion.

See Also

  • read-continue-here — complement: reading and acting on continuation file at session start
  • bootstrap-agent-identity — cold-start identity reconstruction that consumes continuation file this skill produces
  • manage-memory — durable cross-session knowledge (complements this ephemeral handoff)
  • commit-changes — save work to git before writing continuation file
  • write-claude-md — project instructions where optional continuity guidance lives

GitHub 仓库

pjt222/agent-almanac
路径: i18n/caveman/skills/write-continue-here
0
agentsagentskillsai-assisted-developmentclaude-codeskillsteams

相关推荐技能

railway-docs

文档

Railway Docs Skill可实时获取最新的Railway官方文档,确保回答的准确性。当开发者询问Railway功能特性、工作原理或分享docs.railway.com链接时,应优先使用此技能。它通过专门的LLM优化文档源提供最新信息,避免依赖过时记忆来回答技术问题。

查看技能

n8n-code-python

文档

该Skill为在n8n平台的Python代码节点中编写代码提供专家指导,特别适用于需要使用_input/_json/_node语法、Python标准库或了解n8n中Python限制的场景。它强调JavaScript应作为首选方案,仅当需要特定Python功能或对Python语法更熟悉时才使用Python。Skill提供了快速入门模板和关键注意事项,帮助开发者在n8n中高效编写Python代码。

查看技能

archon

文档

Archon Skill为开发者提供了基于RAG的语义搜索和项目任务管理功能,可通过REST API访问知识库。它支持文档搜索、网站爬取、文件上传和版本控制,适用于技术文档查询和项目管理场景。首次使用时需要配置Archon主机地址,建议在处理外部文档时优先使用该Skill。

查看技能

n8n-code-javascript

文档

这个Skill为n8n工作流中的JavaScript代码节点提供专业指导,涵盖数据处理、HTTP请求和日期操作等核心场景。它详细解释了如何正确使用n8n特有的`$input`/`$json`语法、`$helpers`工具以及DateTime对象,并包含关键的错误排查和模式选择建议。开发者通过该Skill能快速掌握Code节点的正确返回格式、数据访问方法和常见陷阱解决方案。

查看技能