review-writing-slop

Review Writing Slop

Use the analyzer as editorial evidence, not as an authority or optimization target.

Setup

The sst binary must be on PATH. If it is missing, install it from the project releases page or with go install github.com/lemon07r/SlopScoreTool/cmd/sst@latest. Inside the source repository, build with make build if the binary is absent. If the installed version exposes a different interface, inspect sst --help rather than guessing.

Analyze

1. Establish the draft's purpose, audience, genre, voice, and hard constraints.

2. Analyze a substantive draft. Do not compose from a prohibited-word list.

3. Request bounded, agent-optimized output:

   sst --format agent --max-findings 20 [FILE...]
   

   With no files, pass the draft on stdin. Use --format brief instead when a structured pipeline consumes the JSON; it is the same evidence with provenance, calibration detail, and excerpts stripped. Reserve --format json for the full-fidelity contract.

4. Read warnings first: short samples, few words, invalid UTF-8, and truncated findings all limit confidence. When output reports truncated findings, rerun with --full before concluding a pattern is absent.

5. Inspect the strongest repeated findings in their surrounding prose. The report gives byte offsets (@start:end) into the original text; re-read the source around them rather than judging phrases in isolation. Treat isolated low-confidence matches lightly. Use --exclude-term only for vocabulary that is necessary to the subject (names, domain terms), never to make findings disappear.

For a before/after question, prefer sst --compare BEFORE AFTER over two separate runs; it reports score, component, and finding deltas without declaring a winner.

Judge

• Ask whether each finding reveals repetition, empty emphasis, vague abstraction, redundant explanation, canned structure, or another weakness a reader would notice without the metric.
• Keep language that is precise, intentional, genre-appropriate, rhythmic, quoted, or necessary to the subject.
• Give repeated patterns and concentrated habits more weight than a single flagged phrase. In multi-file runs, shared findings mark author-level habits, not topic vocabulary.
• Never infer authorship, competence, honesty, or overall writing quality from a score.

Revise

Revise only findings that independently improve clarity, specificity, economy, rhythm, or voice. Prefer fixing the underlying sentence or paragraph over mechanically swapping a flagged word. Preserve useful texture and unevenness.

For a meaningful revision, optionally compare drafts directly:

sst --format agent --compare BEFORE.txt AFTER.txt

Read the changed findings in context rather than the score delta. Stop when the prose is better. Do not chase a target score, zero findings, or endless passes.

Return the improved text when revision was requested. Summarize only the editorial choices that help the user; mention important findings deliberately retained when that explains the result.