Likely — AI skill for Claude Code

likely

A Claude Code hook that catches AI hedging, making ungrounded assumptions, and forces verification as a way for self-healing.

The Problem

LLMs hedge constantly. Words like "likely", "potentially", "probably" appear in **35% of sessions** — not because the model is uncertain, but because it's trained to sound non-committal. This is dangerous in a coding context: when the AI says "this will likely work," it might be guessing rather than verifying.

After analyzing 14,000+ messages across Claude Code, Codex, and Cortex sessions:


Word	% Sessions	Median Turn (first use)	Avg Turn
likely	34.5%	6	17.5
potentially	14.2%	2	10.2
probably	11.5%	26	59.9
possibly	3.5%	108	86.4
presumably	2.7%	44	45.5

The AI uses "potentially" by **turn 2** (median). "Likely" shows up by **turn 6**. These aren't genuine expressions of uncertainty — they're verbal tics that mask whether the AI actually verified its claims.

The Solution

Two Claude Code hooks forming a feedback loop:

1. **`hedge-detector.js`** (Stop hook) — After each AI response, scans for hedging words. If found, writes a session-scoped signal file.

2. **`hedge-enforcer.js`** (UserPromptSubmit hook) — Before the next turn is processed, reads the signal and injects corrective context forcing the AI to **verify its assumptions using tools** (read files, grep, run tests) before responding.

AI responds: "this will likely work"
         │
         ▼
    Stop hook detects "likely"
    Writes signal file
         │
         ▼ (next user message)
    UserPromptSubmit hook reads signal
    Injects: "VERIFY your claims before answering"
         │
         ▼
    AI now reads files, checks code, states evidence

The AI doesn't just rephrase with more confidence — it's forced to **do the research** and ci