redundancy

name	redundancy
description	Audit bilingual formal deliverables for low-information repetition, empty contrast, sequencer scaffolds, adjacent restatement, and repeated comparison-axis labels. Use when the user asks whether a draft feels redundant, wordy, low-information, or says the same thing twice.

Redundancy

Mission

Own Redundancy diagnostics for ai-detect.

This subskill is the dedicated owner for:

• low-information repetition
• empty contrast
• sequencer scaffolds used in place of content labels
• adjacent restatement
• repeated comparison-axis labels

It is not a general style critic and it does not decide whether text is AI-generated.

Scope

In scope:

• bilingual formal deliverables
• .tex
• .md
• .js / .ts deck-authoring source when the strings are user-visible

Out of scope:

• ordinary chat
• email drafts
• free-form brainstorming notes
• .jsx / .tsx
• binary .pptx
• OCR / image text

Relationship To ai-detect

• ai-detect remains the parent owner for overall AI-smell auditing.
• redundancy owns only the Redundancy diagnostics surface.
• When ai-detect scans a file, it should delegate redundancy work to this subskill rather than reimplementing the heuristics inline.

Detection Categories

Use exactly these categories in v1:

• redundant_filler
• empty_contrast
• sequencer_label
• adjacent_restatement
• axis_repetition

Do not invent additional category names in v1.

Confidence Policy

Balanced precision means:

• confirmed rules and high-confidence heuristics go straight to findings
• borderline cases stay in the review queue
• do not silently promote review-queue items into confirmed findings

Canonical Workflow

1. Bind The Surface

Before scanning, bind one surface:

• tex
• md
• js
• ts

If the source is unknown:

• infer it from the file suffix
• default pasted text without a file suffix to md

2. Extract Only User-visible Text

• For .tex, scan document-body prose and ignore comments, inline math, display math, macro noise, and structural commands.
• For .md, ignore code fences, inline code, and raw URLs; keep headings, lists, and body prose.
• For .js / .ts, use the bundled AST helper and extract only user-visible string surfaces such as title, subtitle, heading, caption, body, text, content, label, and footer, plus common slide-builder text arguments.
• Do not scan arbitrary string literals, file paths, style keys, or implementation comments.

3. Run Three Layers

Run in this order:

1. confirmed rules
2. pending review-queue candidates
3. heuristics

Heuristics must stay category-bounded to the five allowed categories.

4. Return Structured Findings

Each confirmed finding should include:

• kind
• line_no
• optional end_line_no
• surface
• language
• text
• optional next_text
• confidence
• fix

Pending review hits should stay separate from confirmed findings.

CLI

Primary scanner:

python3 ~/.codex/skills/ai-detect/redundancy/scripts/scan_redundancy.py <file>
python3 ~/.codex/skills/ai-detect/redundancy/scripts/scan_redundancy.py <file> --json

The scanner also supports:

• --text
• stdin
• --surface

References

Load only when needed:

• references/historical-signals.md
• data/rules.json
• data/review_queue.jsonl

Acceptance Bar

The subskill is ready only when:

• it can run standalone on .tex, .md, and .js/.ts
• ai-detect can call it and preserve the Redundancy diagnostics section name
• pending redundancy candidates do not leak into confirmed output
• at least one positive and one negative case pass for each supported surface

Validation And Checkpoints

• Before final handoff, validate the requested artifact or decision against this skill's output contract and report the verification result explicitly.
• Before any local mutation, pass the recoverability gate: create a rollback point when the change is reversible, and request confirmation when backup cannot cover the risk.
• Use an explicit checkpoint when required input is missing, tool evidence conflicts, or repeated attempts fail; wait for approval or route to the named owner instead of guessing.
• For multi-session work, update a progress or HANDOFF artifact with current state, verified result, and next executable step.