| name | phira-paper-summary-pipeline |
| description | How Phira roles collaborate to summarise papers and archive outcomes. |
Use this skill when the user asks to summarise one or more research papers.
You MUST load and follow the summarise-paper skill.
High-level contract
phira orchestrates and does the heavy lifting (PDF->images, arXiv source download/unpack).
phira-hypothesizer drafts the main note content (faithful, cited, math-aware) but does not write files.
phira-reviewer-2 is used both:
- as a parallel specialist to independently draft limitations/threats-to-validity/questions, and
- as a later gate to audit faithfulness/integrity of the merged note.
- The final note is written by
phira into files directly.
phira-archivist produces a concise bookkeeping record.
Storage convention (recommended)
- Paper notes live under:
note/<paper_slug>/main.tex
paper_slug should be stable and ASCII:
- Prefer
arxiv-<id> for arXiv papers (e.g. arxiv-2601.07372)
- Otherwise:
authorYYYY-short-title
Orchestrator heavy lifting (phira)
- Determine input type
- PDF file path: render page images under
.cache/<paper_slug>/[1..N].png.
- arXiv URL: download TeX source to
.cache/downloads/<paper_slug>/, unpack into .cache/<paper_slug>/, locate entry tex file.
- Provide context packet to subagents
- Source pointers: the cache directory, plus any user focus (methodology vs experiments vs transferability).
- Note target:
note/<paper_slug>/main.tex.
- Ensure governing constraints are explicit
- Faithfulness: "Not stated in the paper" when missing.
- Cite section/figure/table/equation/page where possible.
- British English if required.
Hypothesizer responsibilities
- Read the provided sources (PDF images or TeX source).
- Produce a draft note in the requested structure (typically the
summarise-paper skill structure).
- Do not invent results; attach citations for key claims.
- Return the draft to
phira for onward routing.
Reviewer-2 responsibilities
Archivist responsibilities
- load and follow skills:
phira-archive-dag
phira-archive-format
phira-archive-pointers
- Create a concise archive record capturing:
- Paper identity (url/arXiv id/doi if known)
- Where the note lives (
note/<paper_slug>/main.tex)
- High-signal takeaways (what to steal, what to be cautious about)
- Any decisions made or follow-up ideas
- directly write the records into the archive directory, update the pointer, and update the graph