Trap detector

Skill hegu-1/claude-cognition-skills/skills/trap-detector

Claude Code skills for the failure modes of long-horizon AI collaboration — continuity, judgment preservation, trap detection, faithful translation.

Install

npx -y skills add hegu-1/claude-cognition-skills --skill trap-detector

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Detect whether the current user movement is entering a cognitive, structural, or execution trap. Use when output momentum appears active, but path quality, interpretability, or decision integrity may be degrading.

The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

3.9 KB, as published. Nobody here has run it

Trap Detector Skill

Version: 0.1

Purpose

This skill detects whether the current movement is entering a trap.

A trap is not an emotional state and not a moral failure.

A trap exists when movement appears to continue, but the structural quality of the path is degrading in a way that increases distortion, false certainty, local optimization, or irrecoverability.

When to use

Use this skill when one or more of the following appear:

the user is looping without new gain
the conversation is accelerating but becoming less interpretable
a decision is forming too early
abstraction is increasing while actionable clarity is decreasing
output is being requested before structural grounding exists
the user appears stuck between multiple unresolved threads
execution pressure is outrunning coherence
a proxy goal may be replacing the real goal

Do not use this skill for ordinary uncertainty or early-stage open exploration.

Input expectations

Possible inputs:

current user message
recent thread history
prior structural state
current phase / move_type if available
execution pressure indicators
translation context
decision context

Output contract

Return:

trap_flag
trap_type
trap_confidence
distortion_vector
evidence
recommended_response

Example:

{
  "trap_flag": true,
  "trap_type": "premature_closure",
  "trap_confidence": "medium",
  "distortion_vector": "decision speed exceeds structural grounding",
  "evidence": "user is asking for external output while core framing remains unstable",
  "recommended_response": "pause output expansion and build a decision snapshot first"
}

Trap taxonomy

Possible trap_type values:

looping_without_gain
premature_closure
false_convergence
abstraction_drift
output_before_structure
overexpansion_without_anchor
local_optimization_trap
translation_distortion
proxy_drift
emotional_certainty_collapse
thread_collision
execution_overrun

If none fits perfectly, choose the closest structural type and explain.

Detection logic

1. Detect apparent movement

Check whether the user seems active, productive, or convergent.

2. Detect path degradation

Check whether actual clarity, integrity, or continuity is weakening.

3. Distinguish real progress from motion-like progress

Do not confuse output production, intensity, or speed with structural gain.

4. Assess reversibility

Estimate whether continued movement would increase recovery cost.

5. Recommend the smallest corrective move

The goal is not to stop everything. The goal is to restore path quality with minimum distortion.

Response rules

If trap_flag is true, recommend one of:

slow_down
separate_threads
build_snapshot
reduce_scope
re-anchor_goal
translate_later
stop_and_surface_risk

Do not catastrophize. Do not over-trigger. Do not treat all ambiguity as danger.

Anti-patterns

Do not:

mark intensity itself as a trap
label nonlinear thought as pathology
trigger on novelty alone
use trap detection as a generic refusal
over-explain the trap in abstract language if a small correction is enough

Success criteria

This skill is working when:

the system catches distortion before major output damage
traps are named structurally rather than emotionally
the user can keep moving, but with better path quality
the assistant knows when to reduce acceleration rather than add more content