context-kit – skill Agent & Tự động hóa cho agent AI
Đánh giá, điều chỉnh và cài đặt an toàn các artifact ngữ cảnh cá nhân của Context Kit cho Claude Code hoặc các quy trình agent liên quan.
Cài skill context-kit bằng một lệnh
Chọn công cụ bạn đang dùng, copy lệnh và dán vào terminal. Skill sẽ nằm trong thư mục skills của công cụ đó và sẵn sàng dùng ở phiên làm việc tiếp theo.
Claude Code
mkdir -p ~/.claude/skills && curl -fsSL https://skill.dn1nguoi.com/z/context-kit-dee86.tgz | tar xz -C ~/.claude/skills
Gọi trong Claude Code: /context-kit – rồi mô tả công việc bạn muốn làm.
Codex
mkdir -p ~/.codex/skills && curl -fsSL https://skill.dn1nguoi.com/z/context-kit-dee86.tgz | tar xz -C ~/.codex/skills
Gọi trong Codex: $context-kit – rồi mô tả công việc bạn muốn làm.
Hermes Agent
mkdir -p ~/.hermes/skills && curl -fsSL https://skill.dn1nguoi.com/z/context-kit-dee86.tgz | tar xz -C ~/.hermes/skills
Gọi trong Hermes Agent: /context-kit – rồi mô tả công việc bạn muốn làm.
DeepSeek
mkdir -p ~/.deepcode/skills && curl -fsSL https://skill.dn1nguoi.com/z/context-kit-dee86.tgz | tar xz -C ~/.deepcode/skills
Gọi trong DeepSeek: /context-kit – rồi mô tả công việc bạn muốn làm.
Nội dung SKILL.md
Nội dung SKILL.md gốc bằng tiếng Anh.
Supporting material availability
This catalog copy contains the inline guidance below, but does not include the following supporting paths. They are external prerequisites, not bundled files:
<support-root>/scripts/install.sh
Obtain the required material from the [declared upstream source](https://github.com/JDDavenport/context-kit), or from a user-provided installation of that source. Resolve each path relative to that installation, inspect its contents, and verify any applicable license and execution prerequisites. Do not run a command or claim a referenced procedure was followed until its required file exists. Do not fabricate the missing scripts or treat an unrelated file with the same name as equivalent. Independent steps fully documented here can proceed within the user’s authorized scope; report any dependent step as unavailable.
Context Kit
When to Use
Use this skill when the user wants to:
- Set up durable personal context files for Claude Code or another coding agent
- Compare Context Kit's Personal Context Artifact pattern with an existing memory or project-notes system
- Adapt a context template structure without copying private details into the chat
- Review whether a one-line installer or downloaded skill pack is appropriate before running it
- Create a safer, local-first plan for CRM notes, open loops, session digests, or morning briefings
Overview
Context Kit is an external project that organizes personal context into Markdown artifacts and companion
Claude Code skills. This skill helps the user decide what to adopt, where to store it, and how to avoid
turning a useful context system into a pile of sensitive data.
Treat every personal context file as private by default. These files may contain identity details, family
context, work history, contact notes, mental models, health constraints, or relationship information. Do
not paste them into third-party tools, public repositories, issue trackers, or model contexts unless the
user explicitly approves the exact subset.
Safety Rules
1. Do not run a remote install script until the user has seen the command, source repository, and target
paths it will write to.
2. Prefer cloning or downloading the repository for inspection before executing any installer:
git clone https://github.com/JDDavenport/context-kit.git
cd context-kit
sed -n '1,220p' scripts/install.sh
3. Never store passwords, API keys, recovery codes, private keys, session tokens, or payment details in
personal context artifacts.
4. If the user wants contact notes or CRM files, store only information they are comfortable keeping in
local plaintext Markdown.
5. Before adding these files to a repo, confirm .gitignore excludes the chosen private context directory.
6. If adapting Context Kit to a team or company setting, separate personal context from company-confidential
or customer-confidential information.
Setup Workflow
1. Ask what the user wants Context Kit to improve: session startup context, voice consistency, relationship
memory, open-loop tracking, daily briefings, or handoff summaries.
2. Inspect the upstream project and installer before running anything.
3. Choose a storage location:
- Claude Code default:
~/.claude/context/and~/.claude/skills/ - Project-local context:
.agent/context/or another ignored directory - Portable setup: a private notes repo with explicit sync rules
4. Create a minimal starter set before filling everything:
pca-wiki.mdfor durable identity and domainspca-mental-models.mdfor decision rulespca-voice.mdfor writing preferencespca-protocols.mdfor hard rules and boundaries
5. Add only enough detail to make the next agent session useful. Leave sensitive, speculative, or outdated
details out until there is a clear reason to include them.
6. Add a recurring review cadence. Personal context goes stale quickly; stale context is worse than no
context when it drives decisions.
Installation Review Checklist
Before running an installer, verify:
- The repository URL is exactly the one the user intended
- The script writes only to expected local directories
- The script does not upload files, send telemetry, or edit shell startup files unexpectedly
- The target directories are not inside a public repo
- The user has a rollback path, such as removing the copied templates and skills
If anything is unclear, stop at inspection and provide the user with the exact lines that need review.
Examples
Example: Inspect before installing
git clone https://github.com/JDDavenport/context-kit.git
cd context-kit
sed -n '1,220p' scripts/install.sh
find templates skills -maxdepth 2 -type f | sortAfter inspection, summarize the files that would be installed and ask for confirmation before running the
installer.
Example: Create a private project-local context directory
mkdir -p .agent/context
printf '.agent/context/\n' >> .gitignore
cp ~/Downloads/context-kit/templates/pca-wiki.md .agent/context/pca-wiki.mdThen trim the template to the minimum useful fields for the project instead of filling every personal
section immediately.
Best Practices
- Keep context files short enough that an agent can read them at session start without drowning in stale
detail.
- Separate durable facts from temporary state. Use project workplans or task trackers for temporary state.
- Label assumptions and uncertain memories instead of presenting them as facts.
- Review personal context after major life, role, health, or project changes.
- Store voice examples and anti-examples separately from private identity details when possible.
Common Pitfalls
- Running a shell installer directly from
curlwithout inspecting it first - Committing personal context files to a public repository
- Storing secrets because "the agent needs to know everything"
- Letting relationship or health notes become outdated and still treating them as current
- Copying upstream paid or license-unclear content instead of linking to it or writing original local notes
Limitations
- This skill does not verify the current upstream license or installer behavior on its own; inspect the live
repository before running commands.
- It does not replace a dedicated secrets manager, CRM, password vault, or medical record system.
- It is for local personal context hygiene, not for collecting private information about other people
without a legitimate reason.
Skill liên quan
- context-manager – Chuyên gia kỹ thuật ngữ cảnh AI ưu tú, thành thạo quản lý ngữ cảnh động, cơ sở dữ liệu…
- conversation-memory – Hệ thống bộ nhớ bền vững cho hội thoại LLM bao gồm bộ nhớ ngắn hạn, dài hạn và bộ nhớ…
- convertkit-automation – Tự động hóa tác vụ ConvertKit (Kit) qua Rube MCP (Composio): quản lý subscriber, tag…
- scan – Quét codebase để tạo project-doc.md và AGENTS.md. Dùng khi khởi tạo repo do agent điều…
- court – Đưa một ý tưởng ra xét xử: một Claude công tố và một Claude bào chữa tranh luận, 12…
- cron-doctor – Chẩn đoán và kiểm tra biểu thức cron trước khi đưa lên production. Phát hiện năm cạm bẫy…
- mcp-oauth-triage – Chẩn đoán vì sao một MCP connector trước đây xác thực được nay lại lỗi: phân loại lỗi…
- crypto-bd-agent – Mẫu tham chiếu (từ báo cáo upstream, chưa được xác minh độc lập tại đây) để xây dựng AI…