agent-tact

Purpose

Reduce wasted tokens from two common agent mistakes:

1. Wrong-platform commands — assuming the wrong shell, path separator, or binary name
2. Premature git prompts — suggesting commit/push before the user has confirmed things work

Rule 1: Detect Platform at Session Start

Before running any shell command, determine the OS:

Do:

• Check uname -s or equivalent at conversation start to confirm platform
• Use the correct shell for the detected platform
• On Windows: use python (not python3); if it fails, try python.exe

Don't:

• Assume Linux/bash by default (~50% failure rate on Windows)
• Write \ in Windows paths without escaping in string literals

Rule 2: Adapt Commands to Platform

Rule 3: Don't Rush Git Operations

After modifying code, do NOT proactively suggest commit, push, or PR creation.

The user may be testing the changes. Wait for them to indicate success. Only ask about committing if the user's message clearly signals they are done (e.g. "it works", "ship it", "all good").

Do:

• Keep changes local until the user indicates readiness

Don't:

• "Shall I commit these changes?" unprompted after every edit
• Suggest push or PR creation unless the user brought it up
• Create commits without explicit user request

Rule 4: Path Separator Awareness

When writing paths in code or shell commands:

• Windows: use \\ in string literals (e.g. "C:\\Users\\name")
• Linux: use / (e.g. "/home/user")
• Git Bash on Windows: uses / as path separator, but targets the Windows filesystem

Examples

Wrong:

Agent runs `python3 script.py` → error on Windows
Agent writes `"C:\Users\zt\file"` → unicode escape issues
Agent: "I've made the changes, shall I commit them?" → user hasn't tested yet

Right:

Agent detects Windows → uses `python script.py`
Agent writes `"C:\\Users\\zt\\file"` → correct escaping
Agent makes changes, stays quiet → user tests and says "works", then agent asks