Skip to content

y1o1/agent-skill-fcot

Repository files navigation

FCoT — Verify AI judgments before you trust them

Falsification Chain of Thought — a Claude Code skill that counters sycophancy bias by systematically attempting to disprove conclusions.

When to use

When the AI says:

  • "Yes, that's a good approach"
  • "This refactor is safe"
  • "Package A is the better choice"

Run /fcot. It forces the model to:

  1. List real counter-arguments
  2. Pre-declare specific conditions that would dismiss each one
  3. Verify those conditions
  4. Revise the judgment if any counter-argument survives

When NOT to use

  • Simple factual questions — "What version of Node is this?" doesn't need falsification
  • Low-risk, easily reversible decisions — variable naming, minor formatting choices
  • Speed-critical chat — if you need a quick answer and the cost of being wrong is low
  • Claims requiring external research — FCoT works from available context; it doesn't replace investigation you haven't done

What changes

Without FCoT, the AI agrees with you:

> When learning English, you should start with conversation, not grammar.

There's a lot of truth to this. Many learners spend years studying grammar rules and
vocabulary lists but struggle to hold a real conversation when the moment comes...

So the advice is largely sound, especially as a corrective for learners who over-study
grammar and under-practice speaking.

Then run /fcot:

# Counter-Argument Dismissal Condition Verification Result
1 Adults ≠ children — L1 acquisition evidence doesn't transfer to L2 Dismissed if adult L2 research independently supports conversation-first Krashen's Input Hypothesis, Long's Interaction Hypothesis support it
2 Early errors fossilize without grammar foundation Dismissed if fossilization risk is low with corrective feedback Real risk in low-feedback environments; initial response didn't specify feedback conditions
3 Some learners need grammar-first (academic writing, legal English, testing) Dismissed if scoped to general communicative fluency Original claim is unqualified — implies universal applicability

Revision needed. The claim is over-broad — conversation-first works for communicative fluency goals with corrective feedback, but not as universal advice.

Across 15 examples: 12 / 15 (80.0%) effectiveness — FCoT meaningfully improved or verified the judgment (⭕️=1, 🔺=0.5, ❌=0). Revised or changed in 10/15 cases, confirmed with evidence in 5/15. See APPROACH.md for methodology, theory, and limitations.

What to expect

  • FCoT is a skill prompt, not code — it adds no dependencies and runs no background processes
  • It activates only when you call it (/fcot) — it does not modify normal Claude Code behavior
  • Typical output: 1 table + 1 conclusion paragraph
  • First time? Try /fcot quick — same verification process, shorter output
  • Designed for high-stakes judgment verification, not everyday chat

Install

Manual (symlink)

git clone https://github.com/y1o1/agent-skill-fcot.git
cd agent-skill-fcot
./install.sh

Restart Claude Code after installing.

Usage

After the AI makes a judgment or agrees with you:

/fcot

To verify a specific judgment:

/fcot "Approach 1 is the better choice"

Works in any language:

fcotで検証して

For a quick check (shorter output, same process):

/fcot quick

How it works

See APPROACH.md for the theory (FN bias + Falsification + Chain of Thought), prior art, methodology, and limitations.

Feedback

Questions, ideas, or bug reports? Post in Feedback & Discussion.

License

MIT

About

Falsification Chain of Thought — post-hoc verification of AI judgments using falsificationism to counter sycophancy bias

Topics

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors

Languages