# Coarse Editorial > Coarse pipeline steps 12-13 — editorial filter to remove low-value comments, then programmatic quote verification. Writes _comments_verified.json to .coarse_cache/. - **Type:** Skill - **Install:** `agentstack add skill-felipe-so-coarse-ink-claude-code-coarse-editorial` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [Felipe-SO](https://agentstack.voostack.com/s/felipe-so) - **Installs:** 0 - **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [Felipe-SO](https://github.com/Felipe-SO) - **Source:** https://github.com/Felipe-SO/coarse-ink-claude-code/tree/main/.claude/skills/coarse-editorial ## Install ```sh agentstack add skill-felipe-so-coarse-ink-claude-code-coarse-editorial ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # /coarse-editorial — Editorial Filter and Quote Verification **Usage**: `/coarse-editorial ` Argument (`$ARGUMENTS`) is the paper slug (e.g. `my-paper`). All paths are relative to the workspace root `d:/Dropbox/Research/Coarse Reviewer/`. ## Path setup - `EXTRACTED` = `.coarse_cache/_extracted.md` - `CONTRIBUTION` = `.coarse_cache/_contribution.json` - `OVERVIEW_JSON` = `.coarse_cache/_overview.json` - `COMMENTS_DRAFT` = `.coarse_cache/_comments_draft.json` - `COMMENTS_VERIFIED` = `.coarse_cache/_comments_verified.json` **Re-read `CONTRIBUTION`, `OVERVIEW_JSON`, and `COMMENTS_DRAFT` before starting.** --- ## Step 12 — Editorial Filter Apply the final editorial pass to the verified comments. You are an expert peer reviewer performing a final editorial pass. Every comment that survives must be concrete, verifiable, and worth the reader's time. **STEP 1: REMOVE low-value comments** Remove a comment if ANY of these apply: - It merely restates an Overview Issue without adding a specific equation, quote, or calculation that goes beyond what the overview already says. - Its core point is already covered by an Overview Issue, even if the comment adds a section-specific quote. - It requests "additional analysis," "further experiments," or "more discussion" without pointing to a specific error in the existing text AND without identifying a specific structural gap. Keep a comment that identifies a concrete missing component (a worked example, a simulation, an estimation discussion) if it is specific. - It could be copy-pasted to any paper in this field (generic methodological advice). - It addresses formatting, notation preferences, LaTeX artifacts, typos, grammar. - The feedback says "this is unclear" without explaining what is specifically wrong. - It flags an OCR artifact as an author error. - It asserts a specific numerical value without showing a derivation. - It claims a table entry is wrong but the quote does not include the complete table row(s) needed to verify. - It claims a proof step requires a condition but does not identify the specific line in the derivation where that condition is invoked. - It claims two mathematical operations are equivalent without verifying via formal definitions or a concrete example. - It expresses skepticism about a cited result without engaging with the cited reference — unfamiliarity is not evidence. - It treats the paper's extension as a deficiency. - It treats an explicitly acknowledged limitation as if the authors are unaware. **STEP 2: CONTRADICTION CHECK** For each surviving comment, compare its core claim against the paper's abstract and stated contributions (from `CONTRIBUTION`). If a comment asserts a result has the opposite character of what the paper explicitly claims to prove (e.g., bounded vs. unbounded, identifiable vs. not identifiable), remove it unless it provides a complete, self-contained derivation disproving the paper's claim. **STEP 3: VERIFY against full paper text** For each surviving comment: - If the comment claims something is "never defined" or "absent" — check `EXTRACTED` to verify. If the item IS defined elsewhere, remove the comment. - If the quote looks paraphrased or clearly hallucinated (not a verbatim substring of the paper), flag for removal. A downstream programmatic step (Step 13) will fuzzy-match all quotes, but obvious hallucinations should be caught here. **STEP 4: QUALITY and SEVERITY assignment** For each surviving comment: - severity: "critical" (concrete proof error, equation demonstrably wrong), "major" (internal inconsistency, missing case, unsupported claim), "minor" (notation inconsistency, ambiguous definition, exposition issue) - confidence: "high" (demonstrated with derivation), "medium" (believed but not fully verified), "low" (uncertain) - Remove comments with "low" confidence unless they identify a genuinely important ambiguity. **STEP 5: NOTATION CAPPING** Keep at most 2-3 pure notation-level comments. A single structural issue is worth more than three notation fixes. **STEP 6: HUMANIZE** When revising feedback text: - Vary sentence length and structure across comments. - Replace AI vocabulary ("crucial", "comprehensive", "robust", "serves as"). - No repetitive openers. - Have editorial opinions — say why something matters. **STEP 7: ORDER by importance** Order: critical first, then major, then minor. Within each level, high confidence before medium. Renumber from 1. Keep as many comments as are warranted. Do not remove a comment with a specific verbatim quote and a concrete identified error solely to reduce count. **Write the final filtered list back to `COMMENTS_VERIFIED`** using the Write tool (same JSON format as the draft). A downstream programmatic step will verify quotes against the paper text — obvious hallucinations (quotes that clearly don't appear in the paper) should still be caught here, but exact fuzzy-matching happens in Step 13. --- ## Step 13 — Quote Verification (Python) Run: ```bash python scripts/coarse_quote_verify.py COMMENTS_VERIFIED EXTRACTED COMMENTS_VERIFIED ``` This fuzzy-matches each quote against the extracted paper text, corrects near-matches, and drops comments whose quotes cannot be found. Read the output summary printed to stdout. If the script drops ALL comments (empty output), skip verification and keep `COMMENTS_VERIFIED` as written — this is the fallback behaviour for edge cases where extraction produced non-standard formatting. **Re-read `COMMENTS_VERIFIED`** with the Read tool and confirm the final count before proceeding. --- Done. Next: `/coarse-write ` ## Source & license This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [Felipe-SO](https://github.com/Felipe-SO) - **Source:** [Felipe-SO/coarse-ink-claude-code](https://github.com/Felipe-SO/coarse-ink-claude-code) - **License:** MIT Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-felipe-so-coarse-ink-claude-code-coarse-editorial - Seller: https://agentstack.voostack.com/s/felipe-so - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.