Analyze Variation

Reports rhythm (within-post sentence-length variation) and length spread (across a sequence of posts), and flags monotone posts and too-similar neighbors.

Why this is a skill with a script, not a prompt: length and rhythm metrics must be deterministic. The numbers are computed by the bundled Node script — this skill just tells the agent to run it and how to read the output. Do not eyeball the counts; run the script.

When to use

• Before scheduling a batch of social posts, to check the feed won't read as a same-length content machine.
• Reviewing journal/blog drafts or any prose for metronomic cadence.
• Whenever a house style asks to "vary sentence and post length" (the measurement is voice-agnostic; the style rules stay in the project's own voice doc).

How to run

Always execute the bundled script — never compute the metrics yourself.

The script lives at scripts/analyze-variation.mjs inside this skill's own directory. Resolve that directory however your harness exposes it, then run the absolute path with Node:

• Claude Code: $CLAUDE_PLUGIN_ROOT/skills/analyze-variation/scripts/analyze-variation.mjs
• opencode / other Agent Skills hosts: use the absolute path to this SKILL.md's directory + /scripts/analyze-variation.mjs (the env var above is Claude-Code-only).

# pick whichever resolves on your harness; SCRIPT points at the bundled .mjs
SCRIPT="$CLAUDE_PLUGIN_ROOT/skills/analyze-variation/scripts/analyze-variation.mjs"

# one post per file → sequence analysis with neighbor flags
node "$SCRIPT" posts/*.md

# a single post inline or piped
node "$SCRIPT" --text "Short. Then a noticeably longer sentence."
pbpaste | node "$SCRIPT"

# machine-readable, or write a version-tracked report
node "$SCRIPT" posts/*.md --json
REPORT_DATE=2026-06-07 node "$SCRIPT" posts/*.md --write report.md

Markdown frontmatter and JSX/HTML tags are stripped before analysis. The script is CJK-aware (Chinese characters count as tokens; wide chars weighted ×2 for length), so it works for English and mixed CJK copy alike. Dependency-free (Node built-ins).

Reading the output

• cv (per post) = stdev/mean of sentence length. Higher = more varied rhythm. Below 0.35 with ≥3 sentences is flagged monotone.
• spread(cv) (sequence) = variation of post lengths across the batch. Target ≥ 0.35; lower means the posts are all about the same size.
• too-similar neighbors = adjacent posts whose lengths are within 15%.
• wt = X-weighted length; anything over 280 sits past the in-timeline "Show more" fold (informational, not a cap on Premium accounts).

Fixing flags (the script measures; the writer fixes)

Mix short and long sentences within a post; let post length rise and fall across the schedule. Apply the project's own voice rules for how to cut or expand — this tool only tells you where the rhythm is flat.

Thresholds (tunable)

Defined as named constants at the top of the script:

• MONOTONE_CV_THRESHOLD = 0.35 — a post with ≥3 sentences and a lower sentence-length CV is flagged monotone.
• NEIGHBOR_SIMILARITY_THRESHOLD = 0.15 — adjacent posts within this length fraction are flagged too-similar.

These are first-guess heuristics; tune them to your corpus.