External tools you can run yourself¶
Third-party tools that pair well with EMPN work but are not part of the empn-core plugin. They run from a Mac Terminal, not from inside a Claude Cowork session.
Coarse: deep peer-review of an academic paper¶
Coarse is a free open-source tool that runs a 15-stage peer-review pipeline on an academic paper: extracts the structure, critiques each section, checks internal contradictions, verifies quotes against the source. The output is the kind of detailed report a journal referee produces.
Typical uses: a second opinion before citing a paper heavily; a check on whether a working paper's findings are solid enough to build a brief on.
Runs on a Claude Max subscription. No API key and no per-token cost: every model call routes through the local claude CLI and bills the Max quota. Each review counts against weekly Max usage.
Designed for academic papers with formal structure. On other genres the section-level scaffolding and the overall verdict fit poorly. Line-level comments remain useful regardless.
One-time setup¶
Step 1. Open the Terminal.
Press ⌘ + Space, type Terminal, press Enter.
Step 2. Log into Claude with the Max subscription.
In the browser window that opens, pick the Claude Pro / Max subscription option, not "API key". Sign in with the same account used on claude.ai.
Confirm the login:
The reply should be OK. A "401" or "authentication credentials" error means the login did not complete; repeat claude /login and pick the subscription option.
Step 3. Install the uv package manager.
uv runs coarse without a manual Python install.
Wait for the installer to finish (about 30 seconds), then close the Terminal window and open a new one (⌘ Q, then reopen). Verify:
The output should be a version like uv 0.7.x. command not found means the install did not complete; repeat Step 3.
Every-time usage¶
Step 1. Save the paper in Downloads. Supported formats: .pdf, .docx, .md, .tex, .html, .epub. Non-PDF formats require no additional API key. PDF papers cost about $0.10 per review for the OCR step and require an OpenRouter key (see "PDF papers" in the Troubleshooting section).
Step 2. Open a new Terminal window (⌘ Space, type Terminal, press Enter).
Step 3. Run the review. Replace Paper name.docx with the actual filename:
uvx --python 3.12 --from 'coarse-ink[formats]' coarse-review ~/Downloads/"Paper name.docx" --host claude --output-dir ~/Downloads
On first run, uv downloads coarse (about 30 seconds). Subsequent runs start immediately.
Step 4. Wait. A full review takes 15 to 25 minutes. The Terminal window must stay open for the duration. Progress messages appear as each stage completes. If output stops for more than 5 minutes, cancel with ⌃ + C and re-run.
Step 5. Read the review. The output lands in Downloads as <paper name>_review.md. Any Markdown-aware viewer (VS Code, Obsidian, TextEdit) will open it.
Troubleshooting¶
"Permission denied: '/Users/coarse-output'". The --output-dir ~/Downloads argument is missing from the command in Step 3.
"DOCX extraction requires mammoth" (or similar for .html, .epub). The [formats] extra is missing from the command. Include the square brackets exactly as shown.
"claude -p exited 1" or "401 authentication". The Claude login expired. Repeat one-time-setup Step 2.
Prompt for an OpenRouter key on a PDF paper. Coarse requires a paid OCR key for PDFs (about $0.10 per paper). Two options: (a) convert the PDF to .docx (Word → Save As → .docx) and review the .docx instead. Equations may degrade but no key is needed. (b) Create an OpenRouter account at openrouter.ai, obtain a key, then run uvx --from 'coarse-ink' coarse setup in a Terminal and paste it. Keys are stored locally on the first setup.
About the output¶
The review follows a journal-referee format: an executive summary with an accept / revise / reject verdict, a per-section critique, internal contradictions, quote-verification results, and line-level comments.
Coarse is tuned on academic writing. It expects a methods section, a formal literature review, and on math papers, proof steps. On material without that structure, the section-level scaffolding and the overall verdict fit less well; the line-level comments do not depend on that scaffolding.
Attribution¶
Repository: github.com/Davidvandijcke/coarse. Coarse is maintained externally. EMPN uses it as a design reference for the writing skills in this plugin. See external-references.md.