CodeBuddy Cookbook

For people who can already sign in. Each recipe is one job: goal, official commands, then the pitfall.

If you cannot install yet, read the tutorial. Command lookup is the cheatsheet.

Run /init before you edit a repo

Goal: build project context in the CLI before a large change.

Source: Quick start

cd /path/to/your-project
codebuddy

> /init

The official page marks /init as "strongly recommended": it builds a knowledge graph so later turns scan less. After a large structural change:

> /clear
> /init

Pitfall: skipping /init and throwing a cross-file task at a cold session. Official copy says that is slower and more error-prone. "The repo is huge so it must know" is not a substitute.

One-shot review in print mode

Goal: run once from a script or pipe, no REPL.

Sources: Quick start, CLI reference

codebuddy -p "Optimize this SQL query"
cat error.log | codebuddy -p "Analyze these error logs"
codebuddy -p "Review code quality of src/utils.js" -y

Official wording: with -p/--print, operations that need file access or command execution must add -y (or --dangerously-skip-permissions). You may also use --permission-mode auto / dontAsk, or pre-set permissions.allow.

Pitfall: -p alone blocks turns that need confirmation. -y skips prompts; official text says use it carefully.

Add a custom slash command

Goal: turn a repeated task into /name.

Source: Slash commands · custom

• Project: .codebuddy/commands/
• User: ~/.codebuddy/commands/

test.md registers /test. Nested dirs use colons: commands/frontend/build.md → /frontend:build.

Official review example:

---
description: "Review the given files"
argument-hint: "[file-paths...]"
allowed-tools: Read
---

Please review the following files for quality, maintainability, and security:

@$ARGUMENTS

Commands that run shell lines need Bash in frontmatter or the official page says they will not run.

Pitfall: custom slash commands are user-triggered. Official Skills are "templates the AI recognizes and invokes". Do not write them as one system.

Add an MCP server

Goal: let the CLI call an external tool.

Source: MCP

codebuddy mcp add --scope user my-tool -- /path/to/tool arg1 arg2
codebuddy mcp add --scope project python-tool -- python /path/to/script.py
codebuddy mcp add --scope user --transport sse sse-server https://example.com/mcp/sse
codebuddy mcp add --scope project --transport http http-server https://example.com/mcp/http

CloudBase's own docs say IDE users authorize Tencent CloudBase in Settings / Integrations instead of writing MCP by hand. That is CloudBase's wording; trust the settings page on your machine.

Migrate from Claude Code

Goal: reuse Claude Code agents / commands / skills / instruction files.

Source: Troubleshooting · migrate from Claude Code

Official mapping:

Path	Role
agents/	Custom agents
commands/	Slash commands
skills/	Skills
CLAUDE.md → CODEBUDDY.md	AI instructions and memory

Option 1 (official recommended, symlinks):

cd ~/.codebuddy
ln -s ~/.claude/agents agents
ln -s ~/.claude/commands commands
ln -s ~/.claude/skills skills
ln -s ~/.claude/CLAUDE.md CODEBUDDY.md

Option 2 (copy, independent configs):

cp -r ~/.claude/agents ~/.codebuddy/agents
cp -r ~/.claude/commands ~/.codebuddy/commands
cp -r ~/.claude/skills ~/.codebuddy/skills
cp ~/.claude/CLAUDE.md ~/.codebuddy/CODEBUDDY.md

Verify:

codebuddy

> /skills
> /config

Official note: Skills from Claude Code plugins can be installed in one click and load automatically.

Pitfall: this copies config. It does not make Claude Code and CodeBuddy the same product. Permission modes, login domains, and quota stay on the CodeBuddy account.

Control tokens: new task, new session

Goal: do not stuff ten unrelated jobs into one chat.

Source: Troubleshooting · cost

Official rules:

• /clear for a new task
• /compact on a long chat
• @filename instead of pasting code

Official comparison (their numbers, not this site's measurements):

Approach	Input tokens	Relative cost
10 tasks in one session	~50,000	High
New session per task	~15,000	Low
Periodic /compact	~25,000	Medium

Official advice: /compact every 20–30 turns.

Ask the plugin about the whole repo

Goal: question the project, not just the current file.

Source: Product overview

Official mentions: @workspace and #Codebase for structure, class relationships, dependencies, and business flow.

Pitfall: this is plugin / IDE repo Q&A. It is not the same switch as CLI /init.

IDE: one sentence to a previewable artifact

Goal: use the standalone IDE for 0-to-1, not a one-line patch in an old repo.

Sources: Product overview, IDE landing

Official chain: natural-language idea → structured PRD → prototype / mock (or sketch / component library) → design-to-code (built-in Figma) → CloudBase / Supabase → CloudStudio / EdgeOne Pages.

Pitfall: this is the IDE's home field. The plugin does not become the same product-design-dev pipeline. Do not invent console paths for deploy or BaaS; use the IDE settings page.

Guardrails

• Every command in a recipe must exist on a linked official page. No "usually you can".
• Do not teach WorkBuddy, Yuanbao, WeChat, or QQ here.
• Do not install a CLI plugin and an editor extension as if they were one step.