Think First
February 20, 2026 ยท View on GitHub
Stop your AI agent from coding before thinking. Enforces a plan-first workflow: read before writing, search before creating, understand before fixing. Reduces backtracking, wasted tokens, and broken changes.
Your agent's #1 problem isn't knowledge - it's process. It edits files before reading them. It creates utilities that already exist. It fixes symptoms without tracing root causes. Think First changes the default from "dive into code" to "understand first, plan second, implement third."
The Difference
Without Think First - "Fix the login bug"
Agent opens auth.ts, changes a condition, proposes a diff. Maybe it works. Maybe it breaks the reset flow. Maybe there's already a fix in a different branch.
With Think First - "Fix the login bug"
Agent reads auth.ts, traces the login flow, searches for similar fixes, identifies affected areas (login, reset, OAuth), proposes a plan with risks and testing strategy, then implements.
Install
Cursor / Claude Code / Windsurf
npx skills add ofershap/think-first
Or copy skills/ into your .cursor/skills/ or .claude/skills/ directory.
What's Included
| Type | Name | Description |
|---|---|---|
| Skill | think-first | 10 rules for read-before-write, search-before-create, plan-before-implement |
| Rule | think-first | Always-on behavioral rule that enforces the plan-first workflow |
| Command | /plan | Create a structured implementation plan before making changes |
| Command | /review-approach | Review the current approach and suggest improvements |
The 10 Rules
- Read before writing - read all relevant files before editing any
- Understand the full picture - trace the issue through the codebase before proposing a fix
- Check if it already exists - search for existing implementations before creating new ones
- Propose a plan for complex changes - outline the approach for changes touching 3+ files
- Identify affected areas - find all files, tests, and features that could break
- Consider alternatives - briefly evaluate 2-3 approaches for non-trivial changes
- Verify assumptions - read the actual code instead of guessing
- Minimal change principle - smallest change that solves the problem, separate refactoring from fixes
- Test awareness - identify how the change will be verified before implementing
- Context preservation - stay focused on the original goal, don't chase tangential issues
Related Plugins
- vibe-guard - Security guardrails (pairs well with plan-first workflow)
- ai-humanizer - Prevent AI-detectable patterns in generated content
If this helped your workflow, a star helps others find it.
Author
License
MIT