analyze

paper-wiki Analyze

Overview

The analyze SKILL takes a single paper (by canonical id, arXiv URL, or title fragment) and produces a deep-analysis wiki entry: research problem, method, key innovations, experimental setup, deep analysis, and a comparison block. The output is written into the user's vault as a paper note that the digest can later cross-link to.

This SKILL is the bridge between "I saw this in the digest" and "I have a durable note about it in my vault".

Status: Phase 6 deliverable. The Python runner backing this SKILL is not yet implemented; for now the SKILL guides Claude through the workflow manually, which still produces a usable note.

When to Use

• The user types /paper-wiki:analyze <id-or-title> or /paper-wiki:analyze after picking a paper from a digest.
• The user says "analyze this paper", "summarize ", "give me a deep dive on <paper>".
• The user asks for "the methods" or "the contributions" of a specific paper.

Do not use when the user wants a list of papers to read — that routes to paperwiki:digest. Do not use for general literature overviews or topic surveys (no SKILL covers that yet).

Process

1. Resolve the paper. Accept arxiv:1234.5678, an arXiv URL, s2:<paperId>, or a fuzzy title. Confirm the resolution before doing the work.
2. Fetch metadata. Use the source plugin matching the namespace: arXiv for arxiv:, Semantic Scholar for s2: or unspecified.
3. Read the abstract and (optionally) the PDF. PDF download is a future capability; for now, lean on the abstract.
4. Build the analysis. Cover six sections in this order: research problem, method overview, key innovations, experimental results, deep analysis, comparison with related work.
5. Write to the vault. Paper notes live under the vault's Sources/ subdir at {vault_path}/Sources/{normalized-title}.md (the constant SOURCES_SUBDIR in paperwiki.config.layout). Frontmatter must carry canonical_id, title, tags, status, and confidence so dedup recognizes the entry on the next digest run and wiki_lint does not flag it as broken. Johnny.Decimal / PARA users may override the subdir name in their recipe; the constant is the single source of truth.
6. Hand off to wiki-ingest. Immediately invoke /paper-wiki:wiki-ingest <canonical-id> so the new source is folded into the user's concept articles. Skipping this step leaves a DANGLING_SOURCE finding the next time wiki-lint runs.
7. Cross-link. If the user came from a digest entry, update the digest line to point its wikilink at the new note.

Common Rationalizations

Excuse	Why it's wrong
"I can guess what's in the paper from the title."	Wrong analyses corrupt the wiki and erode trust. Always anchor in the abstract and explicit metadata.
"Skipping frontmatter is fine; the user knows what they wrote."	The dedup filter relies on paper_id/title frontmatter. No frontmatter = recommendation loop.
"If the PDF download fails, give up."	The abstract alone is enough for a usable first note. The user can deepen it later.

Red Flags

• The user provides only a title and there are multiple matches: surface the candidates and ask, do not pick blindly.
• A paper with the same paper_id already exists in the vault: ask before overwriting; offer to refresh metadata only.
• The note exceeds ~600 lines without the user asking for that depth: trim to the six core sections.

Verification

• The output file lives under {vault_path}/Sources/ (or whatever SOURCES_SUBDIR is configured to in the active recipe).
• Frontmatter parses as YAML and contains at least canonical_id, title, tags, status, and confidence.
• /paper-wiki:wiki-ingest ran on the new canonical id and reported zero BROKEN_LINK findings.
• /paper-wiki:wiki-lint does not flag the new source with DANGLING_SOURCE (a concept now references it).
• The next /paper-wiki:digest run dedupes the analyzed paper out of recommendations (it should not surface again).