Give your AI coding tools one consistent brain. BOT-CODE-MODS installs a persistent system prompt into both Claude Code and VS Code Chat / GitHub Copilot so every chat turn, in every project, follows the same rules you set. It also ships an optional token-compression mode.
It's a handful of small scripts and editable Markdown — no framework, no telemetry, no account.
AI assistants forget your preferences between chats and behave differently across tools. Claude Code
reads a CLAUDE.md; VS Code Chat / Copilot read instruction files. BOT-CODE-MODS writes the same
prompt into both, globally, so you get consistent behavior (your rules, your tone, your guardrails)
without repeating yourself.
- One prompt, both tools, every turn — Claude Code loads it each session; Copilot/Chat attaches it
to every request via
applyTo: '**'. - Editable, not hard-coded — the prompt lives in
system-prompt.md. Edit it, re-run the installer, done. You never touch installer code to change what the agents are told. - Cross-platform — Windows (PowerShell) and macOS/Linux (bash).
- VS Code with Claude Code and/or GitHub Copilot Chat (whichever you use).
- Node.js ≥ 18 — only needed for the optional caveman mode. The core prompt install works without it.
- git — only if you want to
git clonethe repo (you can also download the ZIP).
The core prompt install has no dependencies beyond your shell.
git clone https://github.com/<your-account>/BOT-CODE-MODS.git
cd BOT-CODE-MODSWindows (PowerShell):
powershell -ExecutionPolicy Bypass -File .\install.ps1macOS / Linux (bash):
bash install.shThen reload VS Code (Developer: Reload Window) and/or restart Claude Code. To confirm it's
active, open chat and ask: "what standing instructions are you following?"
| Goal | PowerShell | bash |
|---|---|---|
| Prompt only (default) | install.ps1 |
install.sh |
| Also install caveman compression | add -Caveman |
add --caveman |
| Use a different prompt file | -PromptFile .\my-prompt.md |
--prompt ./my-prompt.md |
| Target | Path | Notes |
|---|---|---|
| Claude Code | ~/.claude/CLAUDE.md |
loaded at the start of every Claude Code session |
| VS Code Chat / Copilot | <VS Code User>/prompts/bot-code-mods.instructions.md |
applyTo: '**' attaches it to every chat request |
<VS Code User> is %APPDATA%\Code\User (Windows), ~/Library/Application Support/Code/User
(macOS), or ~/.config/Code/User (Linux). VS Code Insiders and VSCodium are detected and
handled too. Any file it would overwrite is backed up to *.bak-<timestamp> first.
The Copilot/Chat instruction file lives in VS Code's prompts area, which VS Code Settings Sync can sync across your machines.
~/.claude/CLAUDE.mdis outside VS Code and does not sync — install it per machine (or keep it in your own dotfiles).
Open system-prompt.md and make it yours — your voice, your rules, your do's and don'ts. Then
re-run the installer to push the change to Claude Code and Copilot/Chat everywhere. That's the whole
workflow; the prompt is data, not code.
A second file, bootstrap-prompt.md, is a paste-in prompt that sets up a lightweight per-project
"agent system" (a .agent/MEMORY.md, an AGENTS.md, standing rules) in any new or existing repo.
It's optional and self-contained — paste it into your AI chat from inside a project.
-Caveman / --caveman installs caveman, a third-party,
open-source "compressed communication" mode that reduces token usage. It's independent of this kit and
entirely optional; omit the flag if you don't want it.
- Prompt: delete
~/.claude/CLAUDE.mdand<VS Code User>/prompts/bot-code-mods.instructions.md(or restore their*.bak-*backups). - Caveman:
npx -y github:JuliusBrussee/caveman -- --uninstall.
system-prompt.md— the system prompt (edit this).bootstrap-prompt.md— optional paste-in prompt to scaffold a project's agent system.install.ps1/install.sh— the installers.AGENTS.md— a pointer so an AI agent opening this folder knows how to install it.LICENSE— MIT.
Does the prompt really apply on every turn? For Claude Code, CLAUDE.md is loaded per session and
stays in context. For VS Code Chat / Copilot, the instruction file uses applyTo: '**', which attaches
it to every request. Verify by asking chat what instructions it's following.
I don't use Copilot / I only use Copilot. That's fine — the installer writes both targets; the one you don't use is simply inert.
Is anything sent anywhere? No. The scripts only write local files. The optional caveman flag fetches that separate open-source project from GitHub; nothing else leaves your machine.
Windows script won't run. Use -ExecutionPolicy Bypass as shown, or unblock the file. The scripts
are plain text — read them first if you like.
MIT — see LICENSE. Contributions welcome.