grounded-paper-skills
CLAUDE.md0 starsChanged 3 months agoMITView on GitHub
Cite the paper, or say it is missing. No silent invention. Procedures: /paper-ingest, /paper-extract, /paper-study, /paper-survey. Forbidden: speculative completion; paraphrases that change meaning; unsupported assertions; inventing page numbers or section anchors; answering without reading extracted/ (or the PDF when extract is insufficient). For questions / research: Do not use _raw/ as evidence (ingest input only; see /paper-extract). Forbidden: loading an entire PDF or every file under extracted/ at once; skipping _index/ and diving into PDF / extracted/ / _raw; citing…
In this repository4 other files it gives its agents.
- paper-extractSkill
- paper-ingestSkill
- paper-studySkill
- paper-surveySkill
The file
We split this CLAUDE.md into the rules it gives an agent
10 rules. About 0 are common, 7 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈7
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 9
Every claim, summary, explanation, or answer must cite an extracted file under sources/
- Very specific: line 20
Before writing a citation, open the target extracted/ file and confirm the exact block ID exists in it.
- Distinctive: line 22
If content is not in the sources, say Not stated in the attached source first
- Distinctive: line 24
Never present an AI-made example as if it appeared in the paper.
- Distinctive: line 49
Only if still insufficient, read adjacent chunks or the matching PDF pages
- Very specific: line 51
Do not use _raw/ as evidence (ingest input only.
- Distinctive: line 82
No plugin-specific syntax (no Dataview, etc.)
- Distinctive: line 83
Do not add frontmatter keys without asking the user
- Distinctive: line 88
Keep technical terms in their original form when useful (e.g., attention)
- Distinctive: line 89
If this vault uses another language, change this file deliberately — do not mix languages silently
≈ “Common” is an estimate: at least 1 in 5 of the 28,029 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Paper vault (always on)
Cite the paper, or say it is missing. No silent invention.
Procedures: /paper-ingest, /paper-extract, /paper-study, /paper-survey.
Grounding
- Every claim, summary, explanation, or answer must cite an extracted file under
sources/ - Citation format:
Claim summary.
> Quote the source text without altering meaning.
— [[sources/<source-id>/extracted/NN-section#^pXX]]- Before writing a citation, open the target
extracted/file and confirm the exact block ID exists in it. If it does not exist, do not write the citation: either add the anchor toextracted/first, or write Not stated in the attached source - One block ID names one Obsidian block. If a quote spans several blocks, split it and cite each block. Do not treat
^p12-14as a page-range pointer - If content is not in the sources, say Not stated in the attached source first
- Illustrative examples, metaphors, or concrete cases that are not in the source must be marked Not stated in the attached source (illustrative example) before the explanation. Do not put them in
extracted/,summary.md, oralgorithm.md - Never present an AI-made example as if it appeared in the paper. Use:
> [!example] Not stated in the attached source (illustrative example)
> Example content…- Web search is allowed afterward. Web-derived notes may use only the callout below, and must not be mixed into source-derived files such as
summary.md,extracted/, oralgorithm.md
> [!web] Supplement (not stated in the attached source)
> Content…
> Source: https://...Forbidden: speculative completion; paraphrases that change meaning; unsupported assertions; inventing page numbers or section anchors; answering without reading extracted/ (or the PDF when extract is insufficient).
Reading order
For questions / research:
- Identify the paper in
_index/MOC.md - Optionally look up terms in
_index/keywords.md - Read
sources/<id>/index.md - From the section table, read only the needed
extracted/chunks - Only if still insufficient, read adjacent chunks or the matching PDF pages
Do not use _raw/ as evidence (ingest input only; see /paper-extract).
Forbidden: loading an entire PDF or every file under extracted/ at once; skipping _index/ and diving into PDF / extracted/ / _raw; citing _raw/ (always cite extracted/).
Chunking (during ingest): about 300–500 lines per extracted/ file; filename NN-<section-slug>.md; figures/tables: captions only, note "See original for figure". Anchor placement: see /paper-extract.
Obsidian conventions
Frontmatter (paper-related notes):
---
source-id: <slug>
title: "<paper title>"
mode: study | survey | both
status: draft | extracting | ready | studying | surveying
tags: []
authors: []
year: null
venue: ""
created: YYYY-MM-DD
updated: YYYY-MM-DD
---source-id: lowercase letters, digits, hyphensstatus:draft/extractingwhile incomplete;readyafter ingest;studying/surveyingwhile those modes runauthors/year/venue: fill only from the paper. Leave empty ([],null,"") if not stated; do not guess- In-vault refs:
[[sources/<id>/extracted/03-method#^p12]] - Block IDs: Latin letters, digits, hyphens only; unique within each
extracted/file. Same-page blocks:^p12-a,^p12-b. Placement:/paper-extract - Per-paper artifacts →
sources/<id>/; machine extract →_raw/(not a citation target); cross-paper notes →notes/; templates live only in_templates/ - No plugin-specific syntax (no Dataview, etc.)
- Do not add frontmatter keys without asking the user
Language
- Notes and answers default to English
- Keep technical terms in their original form when useful (e.g., attention)
- If this vault uses another language, change this file deliberately — do not mix languages silently
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

