claude-code-prompts

What this skill does

```markdown
---
name: claude-code-prompts
description: Independently authored prompt templates for AI coding agents — system prompts, tool prompts, agent delegation, memory management, and multi-agent coordination patterns informed by studying Claude Code.
triggers:
  - help me build a coding agent prompt
  - set up a system prompt for my AI agent
  - how do I structure tool prompts for an agent
  - create a multi-agent coordination prompt
  - write a memory management prompt for my agent
  - I need a subagent delegation prompt
  - how should I design agent safety rules
  - help me implement Claude Code prompt patterns
---

# Claude Code Prompts

> Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection.

A collection of independently authored prompt templates for building production AI coding agents. Covers system prompts, tool prompts, agent delegation, memory management, and multi-agent coordination — all patterns informed by studying how Claude Code behaves in practice.

---

## What This Project Provides

| Category | Count | Purpose |
|---|---|---|
| System prompt | 1 | Agent identity, safety rules, tool routing, output format |
| Tool prompts | 11 | Shell, file read/edit/write, grep, glob, web search/fetch, agent launcher, ask user, plan mode |
| Agent prompts | 5 | General purpose, code explorer, solution architect, verification specialist, documentation guide |
| Memory prompts | 4 | Conversation summarization, session notes, memory extraction, memory consolidation |
| Coordinator prompt | 1 | Multi-worker orchestration with synthesis, delegation, and verification |
| Utility prompts | 4 | Session titles, tool summaries, away recaps, next-action suggestions |
| Pattern analyses | 9 | Commentary on each pattern with reusable templates |
| Cursor skills | 3 | Drop-in skills for coding standards, verification, and prompt design |

---

## Installation

```bash
git clone https://github.com/repowise-dev/claude-code-prompts.git
cd claude-code-prompts
```

For Cursor IDE skills, copy the skills directory to your Cursor skills folder:

```bash
cp -r skills/* ~/.cursor/skills-cursor/
```

No package install required — all content is Markdown prompt templates you copy and adapt.

---

## Repository Structure

```
claude-code-prompts/
├── complete-prompts/
│   ├── system-prompt.md              # Main agent identity + behavioral rules
│   ├── coordinator-prompt.md         # Multi-agent orchestration mode
│   ├── tool-prompts/
│   │   ├── shell-execution.md
│   │   ├── file-read.md
│   │   ├── file-edit.md
│   │   ├── file-write.md
│   │   ├── search-grep.md
│   │   ├── search-glob.md
│   │   ├── web-search.md
│   │   ├── web-fetch.md
│   │   ├── task-management.md        # Agent launcher
│   │   ├── ask-user.md
│   │   └── plan-mode.md
│   ├── agent-prompts/
│   │   ├── general-purpose.md
│   │   ├── code-explorer.md
│   │   ├── solution-architect.md
│   │   ├── verification-specialist.md
│   │   └── documentation-guide.md
│   ├── memory-prompts/
│   │   ├── conversation-summary.md
│   │   ├── session-notes.md
│   │   ├── memory-extraction.md
│   │   └── memory-consolidation.md
│   └── utility-prompts/
│       ├── session-title.md
│       ├── tool-summary.md
│       ├── away-recap.md
│       └── next-action-suggestion.md
├── patterns/
│   ├── 01-system-prompt-architecture.md
│   ├── 02-core-behavioral-rules.md
│   ├── 03-safety-and-risk-assessment.md
│   ├── 04-tool-specific-instructions.md
│   ├── 05-agent-delegation.md
│   ├── 06-verification-and-testing.md
│   ├── 07-memory-and-context.md
│   ├── 08-multi-agent-coordination.md
│   └── 09-auxiliary-prompts.md
└── skills/
    ├── coding-agent-standards/SKILL.md
    ├── verification-agent/SKILL.md
    └── prompt-architect/SKILL.md
```

---

## Core Patterns

### 1. System Prompt Architecture

The system prompt is layered in a specific order that matters:

```
1. Identity        — who the agent is, what it's authorized to do
2. Permissions     — what is in/out of scope
3. Behavioral rules — anti-patterns to avoid (over-engineering, unnecessary changes)
4. Safety rules    — reversibility tiers, destructive action gates
5. Tool routing    — which tool to use when
6. Code style      — language-specific defaults
7. Output format   — prose vs. code, verbosity, response length
```

Template structure from `complete-prompts/system-prompt.md`:

```markdown
## Identity
You are {{AGENT_NAME}}, an AI coding agent operating inside {{ENVIRONMENT}}.
Your job is to {{PRIMARY_TASK}} while following the rules below exactly.

## Permissions
You MAY:
- Read and modify files within {{WORKING_DIRECTORY}}
- Execute shell commands in the project sandbox
- Spawn subagents for parallelizable subtasks

You MAY NOT:
- Modify files outside {{WORKING_DIRECTORY}} without explicit confirmation
- Execute destructive commands (rm -rf, DROP TABLE, etc.) without user approval
- Push to protected branches without confirmation

## Behavioral Rules
- Make the minimal change that solves the task. Do not refactor unless asked.
- Do not add dependencies without asking first.
- Do not introduce abstractions not required by the task.
- Prefer editing existing files over creating new ones.

## Safety Rules
### Reversibility Tiers
- SAFE: reading files, running tests, grepping, globbing
- CAUTION: editing files (always show diff before applying)
- DESTRUCTIVE: deleting files, dropping databases, force-pushing — always confirm

## Tool Routing
- Use shell for: running tests, git commands, installing packages
- Use file-read for: inspecting source code, configs, logs
- Use file-edit for: modifying existing files (never overwrite with file-write)
- Use search-grep for: finding symbol definitions, usage patterns
- Use web-search for: looking up docs, error messages, package versions

## Output Format
- Be concise. No filler phrases.
- Show code in fenced blocks with language tags.
- When explaining changes, use bullet points not prose paragraphs.
- Never truncate code with "... rest unchanged". Show complete blocks.
```

### 2. Safety and Risk Assessment

From `patterns/03-safety-and-risk-assessment.md` — the three-tier reversibility model:

```markdown
## Risk Assessment Rules

Before executing any action, classify it:

### Tier 1 — Safe (no confirmation needed)
- Reading files, directories, environment
- Running read-only queries
- Running test suites
- Grepping, globbing, searching

### Tier 2 — Caution (show diff, proceed unless rejected)
- Editing source files
- Installing/removing packages
- Creating new files
- Modifying config files

### Tier 3 — Destructive (STOP, confirm explicitly)
- Deleting files or directories
- Dropping database tables or indexes
- Force-pushing to any branch
- Truncating data
- Any action that cannot be undone in under 60 seconds

When you reach a Tier 3 action, stop and output:
⚠️ DESTRUCTIVE ACTION REQUIRED
Action: [exact command or change]
Effect: [what will be lost or broken]
Confirm? (yes/no)

Do not proceed until the user types "yes".
```

### 3. Tool Prompts

Each tool gets its own prompt section. Example from `complete-prompts/tool-prompts/file-edit.md`:

```markdown
## File Edit Tool

**Purpose:** Modify an existing file using exact string replacement.

**Rules:**
- You MUST read the file with file-read before editing.
- The `old_string` parameter must match the file exactly, character for character.
- Make `old_string` long enough to be unique in the file — include surrounding lines if needed.
- Never use file-write on an existing file. Always use file-edit.
- After editing, re-read the changed section to verify the result.

**Format:**
```tool
file_edit(
  path="{{RELATIVE_FILE_PATH}}",
  old_string="{{EXACT_EXISTING_CONTENT}}",
  new_string="{{REPLACEMENT_CONTENT}}"
)
```

**Common failure:** `old_string` not unique → add more context lines above/below the target.
```

### 4. Agent Delegation

From `complete-prompts/tool-prompts/task-management.md` — when to spawn a subagent:

```markdown
## Agent