claude-thinking

claude-thinking

A Claude Code plugin that enforces deliberate thinking, eliminates sycophancy, and improves response quality.

What it does

When enabled, this plugin changes how Claude Code works:

1. Deliberation checklist — Before every substantive response, Claude works through 5 questions: What is being asked? What does a complete solution require? What do I not know? Does this conflict with anything? How will I verify this works?

2. Zero sycophancy — Blocks reflexive agreement phrases ("you're right", "good point", "great idea"). These are a canary for shallow thinking — if Claude is agreeing without examining, it hasn't evaluated deeply enough. Note: the sycophancy patterns currently match English phrases only.

3. Structured evidence requirement — Every substantive response must include a structured **EVIDENCE:** or **CONJECTURE:** block. EVIDENCE clauses must cite typed refs (src:, url:, quote:, knowledge:); CONJECTURE clauses must declare a basis and a likelihood (low/medium/high). A parser in the hook rejects malformed blocks. See agents/thinking.md for the full format spec.

4. Behavioral rules — Think before acting, verify your own work, fix all problems (no "low priority" excuses), answer questions before jumping to action, finish everything the user asked for.

How it works

The plugin has two parts:

• Default agent (agents/thinking.md) — Applies the thinking rules as a system prompt to every conversation. This is always active when the plugin is enabled.
• Stop hooks (hooks/) — Python scripts that mechanically block responses containing sycophantic phrases or lacking the _I THOUGHT_ deliberation marker. These catch violations that slip through the system prompt.

Install

claude plugin marketplace add vgeshel/claude-code-thinking --scope project
claude plugin install claude-thinking --scope project

The first command adds the marketplace. The second installs the plugin from it. --scope project stores the configuration in the project so all contributors get it.

Requires

• Python 3 — the hooks are Python scripts

The plugin checks for python3 on session start and prints a warning if it's missing.

Run tests

cd hooks && python3 -m unittest -v

Customization

Disable the thinking marker

If the _I THOUGHT_ marker is too noisy for your workflow, remove the check_thinking.py hook from hooks/hooks.json. The behavioral rules from the agent will still apply.

Add sycophancy patterns

Edit the SYCOPHANCY_PATTERNS list in hooks/check_sycophancy.py. Patterns are regexes checked against the first 300 characters of each response.

Use as CLAUDE.md instead

If you prefer CLAUDE.md-based rules over a plugin, copy the content from agents/thinking.md (everything below the frontmatter) into your project's CLAUDE.md.