create-senpai
by zhanghaichao520create-senpai is a Claude Code skill that turns chat exports, meeting notes, screenshots, and photos into a reusable senpai persona with group memory, correction flow, and ongoing updates.
该技能得分 78/100，属于值得收录的实用型条目：目录用户可以较快判断它是一个用于在 Claude Code 中生成和持续更新“师兄数字分身”技能的 meta-skill，而且仓库已经提供了触发词、工具调用、安装方式、材料准备和分阶段流程，明显比一段泛化提示词更易触发和复用。扣分点主要在于 SKILL.md 内未直接给出安装命令、支持文件虽多但缺少更强的快速上手与结果示例闭环，用户首次落地仍需自行拼接部分步骤。
- 触发性强：`SKILL.md` 明确列出 `/create-senpai`、`/update-senpai {slug}`、`/list-senpais` 及多种自然语言触发说法，便于代理正确启动。
- 执行路径较清晰：仓库同时提供 `INSTALL.md`、输入引导提示词、纠正流程，以及 `tools/wechat_parser.py`、`qq_parser.py`、`photo_analyzer.py` 等具体解析工具。
- 有持续演化价值：不仅描述首次生成，还覆盖补充材料、对话纠正、版本管理与回滚，体现出真实工作流而非一次性角色设定。
- 安装与运行信息分散：安装命令在 `INSTALL.md`/README 中而非 `SKILL.md`，目录用户理解可安装性没问题，但代理执行时要跨文件找关键信息。
- 仓库的实践性证据偏文档导向：虽然有工具文件和流程说明，但缺少更直接的端到端示例输出、测试样例或最小演示材料，首次采用仍有一定试错成本。
Overview of create-senpai skill
What create-senpai skill actually does
create-senpai is a meta-skill for Claude Code that turns a departed lab senior’s traces into a reusable AI skill. In practice, it ingests chat exports, meeting notes, screenshots, photos, and your own descriptions, then generates a structured persona plus group memory that can be updated over time. The job-to-be-done is not generic roleplay; it is creating a repeatable “senpai-like” collaborator for internal lab workflows, meetings, and shared context.
Who should install create-senpai
The best fit for this create-senpai skill is a team already using Claude Code and willing to organize source material locally. It is especially useful for labs, student groups, or small technical teams that want a persistent voice with project lore, critique habits, and mentoring style. If you only want a one-off character prompt, this repo is probably heavier than you need.
Why it stands out from ordinary prompts
The main differentiator is structure. The repo separates factual memory from behavioral persona, includes parsers for WeChat, QQ, social text, and photo metadata, and supports correction and versioning instead of treating the first output as final. That makes create-senpai for Skill Authoring more maintainable than a long hand-written prompt, especially when you want to revise tone without losing factual context.
How to Use create-senpai skill
Install context and dependencies
The repository is designed for Claude Code skill directories, with project-local install preferred. The practical create-senpai install path is:
mkdir -p .claude/skills
git clone https://github.com/zhanghaichao520/senpai-skill.git .claude/skills/create-senpai
Global install is also documented under ~/.claude/skills/create-senpai. Optional Python dependency comes from requirements.txt; today that is mainly Pillow for EXIF extraction from photos. If you will not analyze group photos or whiteboard shots, you can skip it.
What input create-senpai needs to work well
The skill can start from only a nickname, but output quality depends heavily on evidence. Best inputs, in order, are: exported chat logs, meeting notes or issue comments, then photos/screenshots, then your subjective description. The repo’s prompt flow in prompts/intake.md shows the minimum intake: alias, lab role/background, and persona sketch. Stronger inputs look like: “WeChat export from 2022 rebuttal week, 3 meeting notes, 2 screenshots of issue comments, and one sentence on how he criticizes slides.” That gives the skill enough material to separate recurring facts from speaking style.
How to prompt create-senpai effectively
For real create-senpai usage, do not stop at “make me a senpai skill.” Provide a target, sources, and intended use. A better request is:
- “
/create-senpai gpu-brofrom these WeChat logs and two meeting notes; emphasize how he reviews experiments and responds to broken environments.” - “Create a senpai skill from this QQ export plus my corrections: he is sarcastic, but always ends with concrete TODOs.”
This works because the repo is built around extraction and merging, not freeform invention. If you care about authenticity, explicitly mark uncertain memories versus hard evidence.
Files to read first and suggested workflow
Start with SKILL.md for trigger phrases, tools, and command expectations. Then read INSTALL.md for environment assumptions, README_EN.md for the intended workflow, and docs/PRD.md for the two-layer architecture. After that, inspect prompts/intake.md, prompts/memory_builder.md, prompts/persona_builder.md, and prompts/correction_handler.md. The toolchain in tools/ matters if you plan to feed exported logs instead of manual summaries.
Suggested workflow:
- Install the skill in Claude Code.
- Prepare one high-signal batch of source material.
- Run
create-senpaiwith a clear alias. - Review the generated memory/persona split.
- Immediately apply corrections before wider use.
- Add more material later through the update flow rather than restarting from scratch.
create-senpai skill FAQ
Is create-senpai better than a normal persona prompt?
Usually yes, if you want persistence and revision. A normal prompt can imitate tone quickly, but create-senpai usage is better when you need source-backed memory, correction handling, and versioned evolution. The tradeoff is setup time: you must organize materials and install in a supported skill environment.
Is this beginner-friendly?
Moderately. The concept is easy, but the create-senpai guide assumes you can place files in Claude Code skill directories and, if needed, install optional Python dependencies. Beginners can still use it if they stick to text inputs first and avoid parser-heavy ingestion on day one.
When should you not use create-senpai skill?
Skip this create-senpai skill if you need a general chatbot persona, do not have enough source material, or are not using Claude Code or a compatible skill setup. It is also a poor fit if your goal is a public-facing assistant rather than an internally recognizable collaborator with team-specific memory.
How to Improve create-senpai skill
Give create-senpai richer and cleaner source material
The fastest quality improvement is better evidence, not more adjectives. Chat logs reveal rhythm and recurring phrases; meeting notes reveal judgment patterns; screenshots and photos add timeline anchors. If inputs are messy, pre-label them by source and date. Even a simple note like “private chat, 2023 spring, debugging-heavy” helps the create-senpai skill infer what is stable behavior versus a one-off moment.
Fix the most common failure modes early
Typical failures are: too generic a voice, correct tone but wrong facts, or exaggerated harshness without helpful follow-through. The repo already includes a correction path in prompts/correction_handler.md; use it fast. Strong corrections are specific: “He does not start by mocking; he first asks for logs, then criticizes the method.” That updates behavior much better than “make him more accurate.”
Iterate after the first draft instead of rebuilding
Treat the first output as calibration, not the finish line. The repo supports update/evolution mode, and that is where create-senpai for Skill Authoring becomes more valuable than a static prompt. Add one new evidence batch at a time, check whether memory or persona changed, and keep subjective corrections separate from factual additions. Small, traceable revisions produce a more believable result than dumping everything into the first run.