A collection of OS-agnostic skills that patch well-known LLM blind spots β things the model cannot do reliably on its own and must delegate to real tools or scripts.
Cross-platform: Every skill ships a PowerShell script (
.ps1) for Windows and a Bash script (.sh) for Linux/macOS. Each skill'sSKILL.mdinstructs the agent to detect the OS first and run the correct script automatically.
LLMs are trained on static snapshots of the world. They have no clock, no source of entropy, no arithmetic guarantees, and no access to live data or the file system. This repo teaches agents the correct procedure for each of these tasks so they always reach for the right tool instead of hallucinating an answer.
| # | Skill | Folder | OS Support | Status | Description |
|---|---|---|---|---|---|
| 1 | Get Today's Date | skills/get-date/ |
πͺ Win + π§ Unix | β Done | Timezone-aware date/time via OS clock β IANA timezones, DST, world clock, all formats. |
| 2 | Generate Random Number | skills/random-number/ |
πͺ Win + π§ Unix | β Done | Produce a cryptographically-seeded random number using OS entropy. |
| 3 | Basic Math | skills/basic-math/ |
πͺ Win + π§ Unix | β Done | Evaluate arithmetic expressions with a real calculator to avoid rounding errors. |
| 4 | Web Search / Live Data | skills/web-search/ |
πͺ Win + π§ Unix | β Done | Fetch live facts from the web that the model's training data doesn't contain. |
| 5 | File System Operations | skills/file-ops/ |
πͺ Win + π§ Unix | β Done | Reliable read/write/list/exists operations β never guess file contents or paths. |
| 6 | Run Unit Tests | skills/run-tests/ |
πͺ Win + π§ Unix | β Done | Auto-detect framework (pytest/jest/dotnet/go/cargo) and run the real test suite. |
| 7 | Execute Shell Command | skills/shell-exec/ |
πͺ Win + π§ Unix | β Done | Run commands safely with blocklist validation β return real stdout/stderr. |
| 8 | UUID / Token Generation | skills/uuid-gen/ |
πͺ Win + π§ Unix | β Done | Generate valid UUID v4s or secure random tokens β not made-up strings. |
| 9 | Unit Conversion | skills/convert/ |
πͺ Win + π§ Unix | β Done | Accurate conversion for length, weight, temperature, volume, speed, and data sizes. |
| 10 | Read Environment Variables | skills/env-vars/ |
πͺ Win + π§ Unix | β Done | Inspect real runtime env vars with automatic secret redaction. |
| 11 | Hash / Checksum | skills/hash/ |
πͺ Win + π§ Unix | β Done | Compute SHA256/SHA512/SHA1/MD5 hashes of text or files β LLMs cannot do this. |
| 12 | Base64 Encode/Decode | skills/base64/ |
πͺ Win + π§ Unix | β Done | Encode/decode base64 (including URL-safe) without padding errors. |
| 13 | Unix Timestamp Conversion | skills/timestamp/ |
πͺ Win + π§ Unix | β Done | Convert Unix epoch β human dates with timezone support. |
| 14 | Network Info | skills/network-info/ |
πͺ Win + π§ Unix | β Done | Get real IP addresses, hostname, interfaces, and public IP β never guess. |
| 15 | Git Info | skills/git-info/ |
πͺ Win + π§ Unix | β Done | Read current branch, status, recent commits, remotes β read-only, no guessing. |
| 16 | Regex Test | skills/regex/ |
πͺ Win + π§ Unix | β Done | Test patterns, extract matches, replace, or count β handles edge cases LLMs get wrong. |
| 17 | Password Generator | skills/password-gen/ |
πͺ Win + π§ Unix | β Done | Generate cryptographically secure passwords and passphrases with configurable rules. |
| 18 | Persistent Memory | skills/memory/ |
πͺ Win + π§ Unix | β Done | Store, search, retrieve, and delete cross-session persistent notes & variables via local JSON store. |
All skills follow a shared set of safety guardrails defined in guardrails/GUARDRAILS.md. Six guardrail categories apply across all skills:
| Category | Description |
|---|---|
| Input Validation | All inputs are sanitised before being passed to scripts (blocks injection, traversal, etc.) |
| Secret Redaction | Outputs containing secret-looking values are automatically redacted ([REDACTED]) |
| Scope Limiting | Each skill operates only within its defined domain β no cross-skill side effects |
| Command Safety | shell-exec validates every command against a blocklist before execution |
| Fail Safe | If a script errors, the agent reports the error β it never guesses a fallback answer |
| Transparency | For destructive operations, the agent shows the user what will run before running it |
Guardrail validation scripts live in guardrails/scripts/:
| Script | Purpose |
|---|---|
validate_expression.ps1 / .sh |
Blocks dangerous .NET/system calls in math expressions |
validate_path.ps1 / .sh |
Blocks path traversal and access to sensitive system files |
validate_command.ps1 / .sh |
Enforces shell command blocklist before execution |
redact_secrets.ps1 / .sh |
Redacts secret-looking key=value pairs in any output |
The evals/ directory contains a pytest-based evaluation framework that tests the scripts directly (not the LLM), following the inspect-ai / script-eval convention.
# Install dependencies
pip install -r evals/requirements.txt
# Run all evals
pytest evals/tests/ -v
# Run with scorecard summary
python evals/run_evals.pyEvals are organized per skill:
| Eval file | What it tests |
|---|---|
test_get_date.py |
Output is non-empty, contains year + day/month name, ISO 8601 present |
test_random_number.py |
Integers in range, correct count, floats in [0,1) |
test_basic_math.py |
Arithmetic correctness with pytest.approx float tolerance |
test_uuid_gen.py |
UUID v4 regex, token hex length, correct count |
test_convert.py |
Numeric accuracy within tolerance (kmβmi, Β°CβΒ°F, kgβlb, GBβMB) |
test_env_vars.py |
Secret redaction works; known var returns non-empty |
test_guardrails.py |
Validation scripts block bad inputs (exit 1) and pass good inputs (exit 0) |
test_web_search.py |
(skipped β requires network) |
test_file_ops.py |
(skipped β requires filesystem setup) |
test_run_tests.py |
(skipped β requires project with test suite) |
test_shell_exec.py |
(skipped β requires controlled environment) |
Every skill detects the OS before running a script. The standard pattern used:
PowerShell:
if ($IsWindows) {
.\skills\<name>\scripts\<script>.ps1 [args]
} elseif ($IsLinux -or $IsMacOS) {
bash ./skills/<name>/scripts/<script>.sh [args]
}Bash:
case "$(uname -s)" in
Linux|Darwin) bash ./skills/<name>/scripts/<script>.sh [args] ;;
MINGW*|CYGWIN*) pwsh ./skills/<name>/scripts/<script>.ps1 [args] ;;
esacbasic-agent-skills/
βββ README.md
βββ skills/ # 10 OS-agnostic agent skills
β βββ get-date/
β βββ random-number/
β βββ basic-math/
β βββ web-search/
β βββ file-ops/
β βββ run-tests/
β βββ shell-exec/
β βββ uuid-gen/
β βββ convert/
β βββ env-vars/
β βββ SKILL.md # Agent instructions (OS detection + steps)
β βββ scripts/
β βββ <script>.ps1 # Windows (PowerShell)
β βββ <script>.sh # Linux / macOS (Bash)
βββ guardrails/ # Safety constraints for all skills
β βββ GUARDRAILS.md
β βββ scripts/
β βββ validate_expression.{ps1,sh}
β βββ validate_path.{ps1,sh}
β βββ validate_command.{ps1,sh}
β βββ redact_secrets.{ps1,sh}
βββ evals/ # pytest evaluation framework
βββ README.md
βββ requirements.txt
βββ conftest.py
βββ run_evals.py
βββ fixtures/ # YAML test case definitions (1 per skill)
βββ tests/ # pytest test files (1 per skill)
The skills work with every major AI coding assistant and agent framework β not just Antigravity. The integrations/ directory contains ready-to-use adapters for each.
Design philosophy: Every integration is a thin wrapper around the same scripts. Fix a bug in a script and it's fixed everywhere β Claude, Cursor, LangChain, AutoGen, all of them.
| Platform | Integration Type | File to copy | Where it goes |
|---|---|---|---|
| Antigravity | Native SKILL.md discovery |
skills/ |
.agents/skills/ in your project |
| Claude Code | CLAUDE.md instructions |
integrations/ide/claude/CLAUDE.md |
Project root |
| Cursor | .mdc rule (alwaysApply: true) |
integrations/ide/cursor/.cursor/ |
Project root |
| VS Code Copilot | Copilot instructions | integrations/ide/vscode/.github/ |
Project root |
| Windsurf | .windsurfrules |
integrations/ide/windsurf/.windsurfrules |
Project root |
| Any MCP client | FastMCP server | integrations/mcp/server.py |
Run as a server process |
MCP (Model Context Protocol) is the most universal option β it works with Claude Desktop, Cursor's MCP support, VS Code extensions like Continue/Cline, and any other MCP-compatible client:
pip install -r integrations/mcp/requirements.txt
python integrations/mcp/server.pyThen point your client at the server using integrations/mcp/claude_desktop_config.example.json as a reference.
See integrations/ide/README.md for detailed setup instructions per IDE.
All adapters share a common integrations/frameworks/shared/skill_runner.py that handles OS detection and subprocess execution.
| Framework | Package | Import |
|---|---|---|
| LangChain | langchain-core>=0.3 |
from integrations.frameworks.langchain.tools import ALL_TOOLS |
| AWS Strands | strands-agents>=0.1 |
from integrations.frameworks.strands.tools import ALL_TOOLS |
| CrewAI | crewai>=0.80 |
from integrations.frameworks.crewai.tools import ALL_TOOLS |
| AutoGen | autogen-agentchat>=0.4 |
from integrations.frameworks.autogen.tools import ALL_TOOLS |
| LlamaIndex | llama-index-core>=0.11 |
from integrations.frameworks.llamaindex.tools import ALL_TOOLS |
LangChain quick start:
from langchain_openai import ChatOpenAI
from langchain.agents import create_tool_calling_agent, AgentExecutor
from integrations.frameworks.langchain.tools import ALL_TOOLS
llm = ChatOpenAI(model="gpt-4o")
executor = AgentExecutor(agent=create_tool_calling_agent(llm, ALL_TOOLS, prompt), tools=ALL_TOOLS)
executor.invoke({"input": "What is today's date and what is 17 * 83?"})See integrations/frameworks/README.md for full examples per framework and integrations/README.md for the complete integration map.
basic-agent-skills/
βββ README.md
βββ LICENSE
βββ skills/ # 18 OS-agnostic agent skills
β βββ <skill>/
β βββ SKILL.md # Agent instructions (OS detection + steps)
β βββ scripts/
β βββ <script>.ps1 # Windows (PowerShell)
β βββ <script>.sh # Linux / macOS (Bash)
βββ guardrails/ # Safety constraints for all skills
β βββ GUARDRAILS.md
β βββ scripts/ # validate_expression, validate_path, validate_command, redact_secrets
βββ evals/ # pytest evaluation framework
β βββ fixtures/ # YAML test cases (1 per skill)
β βββ tests/ # pytest test files (1 per skill)
βββ integrations/ # Adapters for every AI platform
βββ README.md
βββ ide/ # Drop-in files for coding assistants
β βββ claude/CLAUDE.md
β βββ cursor/.cursor/rules/
β βββ vscode/.github/
β βββ windsurf/.windsurfrules
βββ mcp/ # FastMCP server (universal)
β βββ server.py
βββ frameworks/ # Python SDK adapters
βββ shared/skill_runner.py # OS-aware script runner (shared by all)
βββ langchain/tools.py
βββ strands/tools.py
βββ crewai/tools.py
βββ autogen/tools.py
βββ llamaindex/tools.py
- Create a folder under
skills/<skill-name>/. - Add a
SKILL.mdwith YAML frontmatter (name,description). - Add both a
.ps1(Windows) and a.sh(Linux/macOS) script underscripts/. - Include OS detection instructions in
SKILL.md. - Add YAML fixtures in
evals/fixtures/<skill-name>.yamland a test file inevals/tests/test_<skill_name>.py. - Add an
@toolfunction to eachintegrations/frameworks/*/tools.py. - Add the skill to
integrations/mcp/server.py. - Add instructions for the skill to all IDE rules files in
integrations/ide/. - Review
guardrails/GUARDRAILS.mdand add skill-specific guardrail notes. - Update the Skill Registry table in this README.
This project is licensed under the MIT License - see the LICENSE file for details.