feishu-paper-reading git:20260729.6b1edc9A

Immutable. This exact content is served forever at /api/v1/blob/4d72b4df2bb03088.

---
name: feishu-paper-reading
description: Search, screen, deeply read, synthesize, and publish recent research papers for a user-specified topic, time window, count, quality/attention balance, and language. Use for recent or high-quality paper lists, literature digests, cross-paper comparisons, original-language evidence excerpts, Chinese explanations, or Feishu research reports.
license: MIT
---

# Feishu Paper Reading

Turn a request for recent literature into an evidence-grounded reading artifact,
not a ranked list of titles. Search broadly, select deliberately, read the full
papers, preserve source anchors, synthesize across papers, and publish only after
the report is complete.

## Load The References

Read these files before acting:

- `references/evidence-policy.md` before searching or selecting papers.
- `references/report-schema.md` before extracting evidence or drafting.
- `references/feishu-publishing.md` only when the requested destination is
  Feishu or another connected document surface.

Use `scripts/validate_digest.py` on a Markdown report before publishing when a
local runtime is available.

## Resolve The Brief

Infer harmless defaults and ask only when a missing choice would materially
change the result.

Default to:

- topic: derive it from the request;
- window: previous 30 calendar days;
- count: 5 deeply read papers;
- selection: technical quality first, then observable attention;
- language: Chinese explanation with original paper titles, terminology, and
  short original-language evidence excerpts;
- depth: full paper plus relevant appendix, figures, and tables;
- destination: a new Feishu document when an authorized compatible connector
  exists, otherwise a complete Markdown report.

Accept user overrides for topic boundaries, dates, paper count, venues, source
types, quality/attention balance, desired depth, compute budget, reading goal,
language, output schema, and destination.

Write a one-sentence operational topic boundary before searching. Ask only when
two plausible boundaries would materially change the candidate pool.

Always state an absolute start and end date and timezone. Interpret "previous N
calendar days" as an inclusive interval from `today - (N - 1)` through today in
the user's timezone, or UTC when it is unknown. For arXiv-only or newly posted
work, use the v1 submission date for inclusion. For newly accepted or published
work, use the official decision or publication date. For a general "recent
work" request, accept either route but label it per paper. Read the latest
available version and record both dates when they differ. Do not silently widen
the window to fill a quota.

## Build The Candidate Pool

1. Search at least `max(5 * requested_count, 25)` candidates when coverage
   permits.
2. Search several query families: exact topic terms, adjacent terminology,
   benchmark or dataset names, and key method names.
3. Prefer primary sources: official proceedings or OpenReview, arXiv, DOI or
   publisher metadata, author project pages, and official code repositories.
4. Deduplicate by DOI, arXiv ID, and normalized title. Treat versions of one
   paper as one candidate.
5. Mark publication status precisely as preprint, under review, accepted, or
   published. Never infer acceptance from an arXiv category.
6. Record queries, sources, candidate count, unavailable sources, and coverage
   gaps.

When available, use `$research-evidence` for metadata and claim verification.
Use `$last30days` only for current attention signals and community discovery.
Social posts, aggregators, GitHub activity, and ranking sites never establish a
technical claim.

## Select Deliberately

Assess each candidate on separate evidence-backed dimensions:

- fit to the user's question;
- novelty and conceptual value;
- rigor and strength of evidence;
- reproducibility: code, data, checkpoints, and implementation detail;
- observable attention, including metric, source, and retrieval date.

Do not turn missing evidence into a numeric zero or fabricate precision.
Preserve methodological and research-group diversity when quality is similar.
Exclude abstract-only candidates from the deep-reading set; place promising
ones in a labeled watchlist. If fewer than the requested count pass the quality
bar, return fewer papers and explain why.

When the user supplies quality/attention weights, first apply a minimum
technical-quality gate. Rank the survivors independently by evidence quality
and observed attention, combine normalized rank positions with the requested
weights for ordering, and break ties by relevance then reproducibility. Use the
combined rank only as an internal selection aid, not an objective published
score. If attention coverage is too incomplete for a fair ranking, disclose
that and fall back to quality-first selection.

## Read The Full Evidence

For every selected paper:

1. Read the abstract, introduction, method, main results, limitations, and
   conclusion.
2. Inspect the figures, tables, appendix, proofs, and supplementary material
   that affect the main claims.
3. Record verified metadata, canonical links, version, and publication status.
4. Build an evidence ledger containing the research question, contribution,
   design choices, datasets, baselines, metrics, central results, failure
   cases, assumptions, limitations, and reproducibility assets.
5. Anchor important claims to a page and section, figure, table, theorem,
   equation, or appendix.
6. Separate `source fact`, `author claim`, and `Codex interpretation`.
7. Record uncertainty and contradictory evidence rather than smoothing it
   away.

Never claim a full read when only an abstract, metadata page, or secondary
summary was available.

## Preserve Original Language

Use original language to let the reader inspect the authors' wording while the
surrounding explanation does the teaching.

- Include 2 to 4 short, high-information fragments per paper when useful.
- Keep all direct quotation from one paper within 25 source-language words by
  default; for Chinese, Japanese, or Korean source text, also keep it within 50
  CJK characters.
- Preserve wording and punctuation exactly.
- Attach a page, section, figure, table, or appendix anchor to every fragment.
- Follow every fragment with an explanation of what it means, why it matters,
  and how strongly the paper supports it. Use Chinese by default or the
  requested output language.
- Prefer claims, definitions, limitation statements, or experimental findings;
  do not reproduce abstracts or long passages.

Retain original titles, method names, dataset names, variable names, and metric
labels. These are not substitutes for evidence anchors. When the user asks for
"more original sentences," explain that the report will use more micro-excerpts
within the per-paper quotation limit rather than reproduce full passages.

## Synthesize, Do Not Stack Summaries

Follow `references/report-schema.md`. The report must include:

1. an executive reading guide;
2. search and selection provenance;
3. a comparison matrix;
4. one deep-reading section per selected paper;
5. agreements, conflicts, trends, and missing evidence across papers;
6. user-specific research opportunities and a recommended reading order;
7. a coverage and confidence statement.

Treat recent preprints as provisional. Distinguish measured results from
author explanations and from your own hypotheses. Prefer a smaller report with
traceable evidence over a larger report built from abstracts.

Describe the final set as the strongest papers within the disclosed candidate
pool. Never imply that a bounded search proves a global "best papers" ranking.

## Validate

Before publishing:

1. Check canonical links and bibliographic metadata against primary sources.
2. Confirm that every selected paper has full-text evidence and locators.
3. Check paper count, duplicate IDs, unresolved placeholders, required report
   sections, quote anchors, and conservative quote limits.
4. Run:

   ```bash
   python "<skill-directory>/scripts/validate_digest.py" report.md \
     --expected-count <actual-selected-count>
   ```

5. Resolve failures or disclose intentional deviations. Do not weaken the
   validator merely to make a report pass.

## Publish And Read Back

Create one new document per run unless the user explicitly asks to update a
specific existing document. Follow `references/feishu-publishing.md`.

After publication, read the document back and verify:

- exact title, date window, and requested paper count;
- canonical paper and code links;
- comparison, per-paper, synthesis, and confidence sections;
- original fragments and their source anchors;
- no placeholders, truncation, duplicated sections, or missing media.

Return the document link and disclose any formatting downgrade. If no
authorized document connector exists, return the same complete report as
Markdown; publication failure must not erase the research result.

## Report Completion

Report the selected count, candidate-pool size, absolute time window, sources
searched, important coverage gaps, destination link or fallback artifact,
readback result, and any papers treated as abstract-only or provisional.

Do not save credentials, request broad drive permissions, overwrite unrelated
documents, or delete existing documents or local sources.