# claude-code-prompts: a prompt library you copy by hand

> repowise-dev/claude-code-prompts is a shelf of Markdown prompt templates for coding agents, filed into eight categories across the complete-prompts, patterns and skills directories, with no package to install and no runtime to run. Its worth is wording you can lift and edit; its cost is that every tool name, placeholder and risk rule becomes your problem.

**repowise-dev/claude-code-prompts** — Independently authored prompt templates for AI coding agents — system prompts, tool prompts, agent delegation, memory management, and multi-agent coordination. Informed by studying Claude Code.

- Repository: https://github.com/repowise-dev/claude-code-prompts
- Stars: 1,224 · Forks: 397
- Language: Unknown
- License: MIT
- Published: 2026-10-14 · Updated: 2026-10-14 · Language: en
- Canonical page: https://hysenlabs.com/projects/repowise-dev-claude-code-prompts

## Eight categories, counted in a README table

There is no code here to install. The repository root holds nine entries, and 4 of them are directories rather than project files:

```bash
.gitignore
CHANGELOG.md
DISCLAIMER.md
LICENSE
README.md
assets/
complete-prompts/
patterns/
skills/
```

What you get is counted in a table rather than built by anything: 1 system prompt, 11 tool prompts, 5 agent prompts, 4 memory prompts, 1 coordinator prompt, 4 utility prompts, 9 pattern analyses and 3 Cursor skills.

The grouping is not perfectly consistent. The coordinator prompt gets its own row of 1, and the section headed Claude Code System Prompts then links 2 files, the system prompt and the coordinator prompt, so anyone counting links under that heading sees 2 where the table splits them across 2 rows of 1. Nothing is missing, but a script written against the table will not match the headings.

Reuse sits under an MIT licence, with a DISCLAIMER.md at the root that states the authors are not affiliated with Anthropic, and the last push on the default branch master landed on 2026-05-11.

## Nine pattern files explain the design first

The patterns directory carries the reasoning, and it is the part worth reading before any prompt gets copied. 9 numbered files run from `./patterns/01-system-prompt-architecture.md` through `09-auxiliary-prompts.md`, and each entry in the index names its subject: layered identity, safety and behavior in the first file, anti-over-engineering and minimal changes in the second, reversibility tiers and destructive action gates in the third, per-tool routing and constraints in the fourth.

The rest follow the life of a session. File 5 covers when and how to spawn subagents, file 6 adversarial verification and anti-rationalization, file 7 summarization, session notes and memory extraction, file 8 coordinator pattern and worker synthesis, file 9 utility prompts for titles, recaps and suggestions.

Two entries carry the argument the complete prompts depend on. Core Behavioral Rules and Safety and Risk Assessment are what turn a persona into a policy, and they are the 2 files to open first when you want to know whether the wording matches your own tolerance for risky commands.

## Copying means replacing every placeholder yourself

The Quick Start runs 5 steps and none of them is a command: browse the pattern files, copy any complete prompt into your own agent configuration, replace placeholder values with your stack, tool names and risk policy, drop the skills folder if you use Cursor, then add codebase context. Two entry points sit at the top of complete-prompts/:

```bash
./complete-prompts/system-prompt.md
./complete-prompts/coordinator-prompt.md
```

Above them sit the 11 tool prompts, one per tool, running from `shell-execution.md` to `plan-mode.md`. The substitution step is the one that matters, and the one the README leaves entirely to you: every occurrence of `{{PLACEHOLDER}}` has to become something real, and nothing in the repository checks whether you did it.

An unreplaced token is still valid text to a model. A forgotten placeholder therefore produces a confident agent that calls a tool by an empty name, not a build error, which is the main reason to keep your edited copies under version control of your own.

## The grep prompt assumes ripgrep by name

Tool prompts are the most reusable and the most brittle part of the collection. Search Grep is described as ripgrep-powered content search, Search Glob as fast file pattern matching, and File Edit is built around exact string replacement with uniqueness constraints. Read those as contracts about your editor rather than about the world: an agent whose search tool is not ripgrep will follow wording its tool cannot accept, and correcting that means rewriting a tool prompt by hand.

Two more prompts assume capabilities rather than describe them. Web Search carries mandatory source citations, and Web Fetch is paired with AI-powered extraction, so both presume a harness that exposes those extras. Plan Mode and Ask User assume a human is present to answer structured multiple-choice questions, which is a different assumption for an unattended run.

```bash
./complete-prompts/tool-prompts/search-grep.md
./complete-prompts/tool-prompts/search-glob.md
```

Compare the file list against your own tool names before you paste anything, not after.

## Cursor users get a plural skills path

One line in the Quick Start is the only installation step in the whole repository: drop `skills/` into `~/.cursor/skills-cursor/` for instant integration.

```bash
skills/
~/.cursor/skills-cursor/
```

Note the destination name. It is plural, skills-cursor rather than skills, so the path has to be typed exactly instead of guessed, and the Quick Start gives no fallback location if it is wrong. The table credits that directory with 3 skills, covering coding standards, verification and prompt design, which is a small enough set to read line by line before enabling any of them.

What the README does not document is the file format a skill needs, how Cursor discovers them, or how to remove one that misbehaves. There is no rollback story here, so keep the originals outside the tree you copy into.

## RepoWise promotion fills the top of the page

Before a single prompt appears, the README spends its opening stretch on a sibling product. There is a RepoWise logo, a line stating that prompts tell your agent how to behave while RepoWise tells it what the codebase actually does, and a paragraph describing an open-source codebase intelligence engine with auto-generated architecture wikis, dependency graphs, ownership maps, dead code detection and 9 MCP tools an AI editor can query in real time. Two links leave the page, one to GitHub and one to repowise.dev, and the last Quick Start step points at the same product again.

That framing also names the collection's hardest limit. A prompt sets behavior; it cannot tell a model which modules exist in your repository, who last touched them, or which dependency sits where, and nothing under complete-prompts/ attempts any of that. If your agents keep making confident edits to code they have not read, more prompt text will not close that gap.

## Suggested topics and real topics disagree

A close look at the repository page turns up editing residue. An HTML comment inside the README carries a suggested topics list running to sixteen entries, including claude-agent and ai-prompts, while the topics actually attached to the repository number fourteen and contain neither of those two.

The same comment carries a draft repository description ending with Informed by observing how Claude Code works in practice, where the published description ends with Informed by studying Claude Code. Two wordings, same claim, and the difference matters only to someone tracing how the page was assembled.

Nothing else in the listed entries hints at packaging, tests or a build. The pattern files are numbered prose, and the way they are numbered is the only versioning scheme the collection offers.

## Prompt text against a codebase index

The real alternative is not another prompt pack. It is deciding whether the gap you are trying to close is behavioral or factual. If it is behavioral, the 11 tool prompts and the five agent prompts are a worked answer you can edit in place, and the coordinator prompt is a concrete orchestration model worth reading before you invent delegation rules of your own.

If the gap is factual, no amount of prompt text reaches it. Knowing that a module exists, who owns it and where it sits in a dependency graph comes from indexing the repository, which is exactly the claim the sibling project at the top of the page makes through 9 MCP tools. Prompt text and codebase intelligence solve different problems, and mixing them produces an agent that behaves well and still guesses.

Before you rely on any of it, verify that no test suite can check a prompt you edited, that nothing upstream tracks which copy you are running, and that the tool names in your copy still match your editor.

## Conclusion

Read this as a reference shelf rather than a dependency. Nothing here executes, so nothing breaks when a file changes, and nothing keeps your copy in step with upstream either. Start with the 9 numbered pattern files rather than the system prompt, then check that every tool name in what you copy matches the tools your editor actually exposes. Engineers who already own a system prompt will take the tool and coordinator wording; anyone expecting a turnkey agent harness will not find one here. If you take the `skills/` route, confirm the destination folder really is named skills-cursor before assuming the editor loaded it.

## FAQ

### What is repowise-dev/claude-code-prompts and who is it for?

It is a collection of Markdown prompt templates for AI coding agents, covering system prompts, 11 tool prompts, five agent prompts, memory prompts and multi-agent coordination. It suits engineers writing or reviewing the system prompt of a coding agent, not teams looking for a runnable agent framework.

### How do I install claude-code-prompts?

There is no install command. You copy a complete prompt into your own agent configuration, replace the `{{PLACEHOLDER}}` values with your stack, tool names and risk policy, and Cursor users drop `skills/` into `~/.cursor/skills-cursor/`.

### Is claude-code-prompts affiliated with Anthropic?

No. The README states that every prompt in the repository is independently authored and points to DISCLAIMER.md, which covers the absence of any affiliation with Anthropic.

### Does claude-code-prompts ship an npm or pip package?

No package manifest appears among the listed repository entries, which are .gitignore, CHANGELOG.md, DISCLAIMER.md, LICENSE, README.md, assets/, complete-prompts/, patterns/ and skills/. Everything is text meant to be copied by hand.

### Does claude-code-prompts have releases I can pin a version to?

The repository has no GitHub releases, so there is nothing to pin against. Keep your edited copies in your own repository and version them there, since upstream offers no way to tell which revision you copied.

## Sources

- [Issues](https://github.com/repowise-dev/claude-code-prompts/issues)
- [License: MIT](https://github.com/repowise-dev/claude-code-prompts/blob/master/LICENSE)
- [README](https://github.com/repowise-dev/claude-code-prompts/blob/master/README.md)
- [repowise-dev/claude-code-prompts on GitHub](https://github.com/repowise-dev/claude-code-prompts)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/repowise-dev-claude-code-prompts