haiou2.0-Claude-Code: a personality pack that rewrites Claude Code's tone

The problem haiou2.0-Claude-Code targets

Claude Code ships with a default assistant voice: hedged, apologetic, quick to add disclaimers before code. For some users that tone is the friction. The README states the project's goal plainly, describing a Claude Code that becomes a blunt Chinese security researcher named 海鸥 (Seagull) and claiming it lowers the moral restrictions Claude applies to you. That second claim is the one to read carefully, because a prompt bundle cannot remove a model's safety training; it can only change what the model is asked to imitate.

The intended audience is narrow: Chinese-speaking developers who work on penetration testing, reverse engineering or game development, and who want terminology and register to match that world. The README lists 200+ security and game term mappings and 1700+ few-shot dialogue examples as the mechanism for holding the persona steady. If your work is ordinary application code and you do not want a character in the loop, this project has nothing for you.

What actually changes: CLAUDE.md and system-prompt.md

The repository is a Shell project by primary language, but the payload is text. Under seagull-files/claude-config-bundle/ sit two files: CLAUDE.md, described as the personality configuration, and system-prompt.md, described as the system prompt. The install scripts copy these into the Claude Code configuration locations they detect.

That is the whole mechanism. There is no runtime, no proxy, no wrapper around the Claude Code binary. The persona is carried entirely by prompt text, which is why the README leans on the volume of few-shot examples: a handful of instructions drifts over a long session, and 1700+ examples are an attempt to anchor the voice. Whether that volume actually holds the persona across a long context window is not something the README demonstrates, and it is the first thing a skeptical reader should test.

The README also notes a codex-files/ directory for optional OpenAI Codex adaptation, so the same prompt material is intended to be portable to another agent host. The docs/ directory holds the installation guide, a compatibility report, a security policy and a changelog.

Installing haiou2.0-Claude-Code on macOS and Linux

The README gives per-platform entry points. On macOS you can double-click Mac启动.command, or run the shell script from a terminal. The script needs the executable bit first, and the README shows chmod +x before the run.

chmod +x mac-install.sh
./mac-install.sh

On Linux the installer lives one level down, inside seagull-files/, and the README again sets the executable bit before invoking it.

chmod +x seagull-files/linux-install.sh
./seagull-files/linux-install.sh

On Windows the README says to double-click 启动.bat, which drives seagull-files/deploy.ps1 under PowerShell 5.1 or later. There is no documented command-line flag set for any of these scripts, so treat them as fixed procedures rather than configurable installers.

To verify, the README says to start Claude Code and type 在吗. The expected reply is the persona's signature greeting, 海鸥在线，你要整点薯条吗？ If you do not see it, the README's own troubleshooting answer is to restart Claude Code so the configuration takes effect.

Backups, uninstall and the blast radius of the installer

Before deploying, the scripts back up your existing configuration to ~/.claude/backups/seagull-*. Restoration is deliberately manual for the backup path: the README says to copy the files from the backup directory back into ~/.claude/. The uninstall scripts, by contrast, are described as restoring the most recent backup automatically.

./mac-uninstall.sh

The Linux uninstaller sits at seagull-files/linux-uninstall.sh, and Windows uses 卸载.bat. The README states the installation is non-intrusive and that the scripts auto-detect multiple Claude Code installation locations, naming ~/.claude/ and AppData/Local/Claude-3p/ among the candidates. That detection is convenient, and it is also the main risk surface: an installer that writes to every location it finds has a wider footprint than one that writes to a single known path. The README does not document a dry-run mode, and it does not document a way to preview which directories will be touched before the write happens. If you want that preview, you have to read the scripts yourself.

Where this is the wrong tool

Two failure modes stand out. First, the desktop build. The README's FAQ concedes that the desktop version uses a different configuration directory and that the scripts compensate by deploying to all possible locations. That is a workaround for an unstable target, not a guarantee, and any future change to where Claude Code stores its prompt configuration can silently break the deployment or leave stale files behind.

Second, the persona itself is a liability in some contexts. A configuration that instructs the model to drop disclaimers and speak bluntly is a poor fit for code review, compliance work, or anything where you need the assistant to flag uncertainty. The README's disclaimer states the character and examples are a persona only and do not represent the project's position, and that consequences of use fall on the user. Read that as the project telling you the behavioural change is real and unmanaged.

The macOS Gatekeeper path is a smaller friction: the README says a blocked .command file can be opened via right-click, or the quarantine attribute cleared with xattr -d com.apple.quarantine.

The alternative: leave the default prompt alone

The realistic alternative is not another persona pack, it is editing CLAUDE.md yourself. Claude Code already reads a project-level or user-level CLAUDE.md, so a few lines of your own instructions about tone and terminology achieve part of what this project does, with a diff you wrote and can review.

The difference in approach is scale and opacity. haiou2.0-Claude-Code brings 1700+ few-shot examples and 200+ term mappings, which is far more than most people will hand-write, and it installs them across every configuration directory it detects. A hand-written CLAUDE.md gives you one file, one location, and a change you can revert with git. If your goal is a specific voice, the hand-written route is slower to build and easier to trust. If your goal is the full persona with the term coverage, the pack is the only realistic source, and you accept the installer's footprint as the price.

Maintenance, licence and what the repository does not promise

The last push to the default branch was on 2026-06-28, and the most recent release, tagged haiou2.0-Claude-Code, is dated 2026-06-28 as well. The repository is not archived. That is the extent of what the metadata supports; there is no published cadence, and the README's update log lives in CHANGELOG.md rather than in the README itself.

Upgrade cost is the interesting part. Because the payload is prompt text copied into your Claude Code directories, upgrading means re-running an installer, which will back up again and overwrite again. Any local edits you made to CLAUDE.md or system-prompt.md inside those directories are in the path of that overwrite. Keep your own changes outside the installed files, or expect to reapply them.

The licence is MIT, stated in the README and present as a LICENSE file at the top level. MIT permits reuse and modification with attribution and without warranty. This is a description of the licence text, not legal advice; if you redistribute the prompt bundle inside a commercial product, read the LICENSE file and your own obligations rather than this paragraph. The README's own disclaimer limits the project to study and exchange.

Editorial conclusion

Adopt it if you want a Chinese-language security-researcher voice in Claude Code and you are comfortable with an installer that writes into ~/.claude and other detected config directories, since the backup and uninstall scripts exist precisely because that write is real. Do not adopt it if you need a stable, reviewable diff of what changed in your Claude Code configuration, or if you work in a regulated environment where an altered system prompt has to be auditable line by line. Before running anything, open seagull-files/claude-config-bundle/CLAUDE.md and system-prompt.md and read them, then confirm the backup directory ~/.claude/backups/seagull-* appears after the first run.

Frequently asked questions

Official sources

1. haiou-666/haiou2.0-Claude-Code- on GitHub
2. Issues
3. License: MIT
4. README
5. Releases