study-exemplars

name

study-exemplars

description

Studies exemplar papers from a target venue and produces an original style-and-structure brief — fetches best-paper awardees and top-cited papers on demand (DBLP, Semantic Scholar, Unpaywall, arXiv, open-access ACM DL) and analyzes section architecture, contribution framing, evaluation patterns, and figure/table conventions. Use when the user wants to study best papers or award-winning papers at a venue, find the most-cited papers and how they are structured, learn how successful papers at a conference are written ("what do winning SIGSPATIAL papers look like", "analyze NeurIPS best papers before I draft"), or model a draft on a venue's strongest work. Copyright-safe by design - papers are fetched from legal open-access sources and processed transiently, never bundled, stored, or committed; the output is metadata plus original analysis. Trigger words - exemplar, best paper, award-winning, most cited, top cited, model paper, venue style, paper structure.

Study Exemplars

Turns a target venue's strongest papers — best-paper awardees and top-cited work — into an original style-and-structure brief the user can write against: how winning papers at this venue architect their sections, frame contributions, design evaluations, and use figures and tables. Papers are fetched on demand from legal open-access sources and processed transiently. The deliverable contains metadata and original analysis only — never paper text. Sits between select-venue/parse-cfp (choosing the target) and the writing skills (write-abstract, draft-related-work, tailor-to-venue).

When to use

• "What do best papers at look like?" / "analyze the award winners"
• "Show me the most-cited papers and how they're structured"
• "I'm submitting to for the first time — how do successful papers there frame contributions / run evaluations / lay out sections?"
• Before drafting: build a venue style brief that other writing skills consume
• NOT for finding papers on a topic (find-papers) or reviewing the literature for content (literature-review) — this skill studies form

Inputs

• A target venue, ideally with a profile venues/conferences/<id>.yml (schema: venues/schema.yml); otherwise resolve aliases per references/finding-exemplars.md
• Optional: year window (default: the last 3–4 completed proceedings years), exemplar count (default 5–8), the user's paper type (research / short / demo) so analysis targets the right track
• CONTACT_EMAIL env var — required by every script (polite-pool identity); scripts prompt interactively if unset, or exit nonzero with instructions
• Optional: S2_API_KEY env var for reliable Semantic Scholar access

Process

1. Resolve the venue and lock the target

• Read the venue profile if one exists. Take aliases.s2_venue and aliases.dblp_key from its aliases: block; note the track, page limit, and template the user will write for.
• Re-verify critical facts against the live cfp_url before the user relies on them — the brief will state format conventions (page budget, template, required sections), and profiles go stale. If the profile's verified.date is older than the current CFP cycle, fetch the CFP and reconcile first.
• No profile? Resolve aliases via the find-papers skill's venue-aliases table or live DBLP venue search before any query — a wrong S2 venue string silently returns zero papers.

2. Build the exemplar set (two complementary lists)

Best-paper awardees (the venue's own quality signal):

• Find award pages live — the venue/SIG awards page, year-site news posts, or the jeffhuang.com aggregator. Source map and verification protocol: references/finding-exemplars.md.

• Awards exist in no API. Never assert a winner from memory. Every award claim needs (a) a source URL fetched this session AND (b) a DBLP metadata match:

  python3 scripts/lookup_exemplar.py --title "Exact Title From The Award Page"
  

  If either is missing, drop the paper or label it explicitly unverified.

Top-cited (the community's quality signal):

# S2 venue string from the profile aliases — NOT the acronym
python3 scripts/rank_top_cited.py --venue "SIGSPATIAL/GIS" --year 2020-2023 --top 10
# or read the alias straight from a profile:
python3 scripts/rank_top_cited.py \
    --venue-profile venues/conferences/sigspatial-2026.yml --year 2020-2023 --top 10

One polite request ranks the whole venue-year window by citation count. Rank a window ending 2–3 years back — current-year counts are near zero and meaningless. More selection caveats (survey inflation, influential citations, DBLP cross-checks): references/finding-exemplars.md.

Target 5–8 papers total: 3–4 verified awardees + 3–4 top-cited, spread across years, matching the user's track (don't study 10-page research papers to write a 4-page demo). Confirm the final set with the user before fetching.

3. Fetch each exemplar on demand — transiently

• One paper at a time, never in bulk. Resolve the OA copy with the fetch-paper skill (scripts/resolve_oa.py <DOI> --json there), or use the OA hints both scripts here print (S2 openAccessPdf, arXiv HTML, dl.acm.org/doi/pdf/<doi> for post-2026 open-access ACM papers — that host blocks scripted downloads, so open it in a browser).
• Read the paper, extract observations, discard the file. Never write the PDF, its text, or its abstract into the repo or any committed file.
• No legal OA copy (Unpaywall is_oa: false, no arXiv version)? Skip the paper and say so — list it in the brief as "not analyzed (no open copy)". Never bypass a paywall or use shadow libraries.

4. Analyze each paper against the rubric

Work through references/analysis-rubric.md — the dimensions are: identity card, title/abstract patterns, section architecture, contribution framing, method presentation, evaluation patterns, figure/table conventions, related-work positioning, reproducibility apparatus, writing micro-style. Record facts and original observations (section names, counts, orderings, framing moves), not prose. Quotes: at most one short attributed fragment (<25 words) per paper, only when the exact wording is the observation.

As you record each paper's counts, also log the measurable ones (pages, references, figures, tables, abstract words, teaser/badge booleans, section skeleton) into a small per-paper JSON — these feed step 6.

5. Synthesize the style-and-structure brief

• Cross-paper synthesis first (what ≥ half the exemplars do = the venue convention; splits = noted as variants), then one exemplar card per paper. Templates for both are at the end of the rubric.
• Reconcile with the venue profile: if exemplars contradict the current CFP (e.g. older 8-page exemplars vs. a 10-page limit today), the live CFP wins — flag the delta so the user doesn't imitate an outdated rule.
• Cite every exemplar by verified metadata (title, authors, year, DOI). If any entry will land in the user's bibliography, route it through verify-citations.

6. Cache a measured exemplar bundle (data hygiene)

The session's measurable facts are the on-family distribution downstream skills score drafts against. When a downstream skill's live exemplar fetch is skipped or rate-limited, it falls back to the venue/family profile's exemplar_distribution: block — which for most venues is hand-estimated, never measured. Cache yours so that fallback rests on real exemplars:

python3 scripts/build_exemplar_bundle.py measurements.json --out block.yml

This aggregates your per-paper measurements (from step 4) into a schema- conforming block: density bands (never fabricated single points), rates, and the modal skeleton, each stamped with measured: true, n, recency, and as_of: <date>. Paste it into the relevant profile under review (venues/conferences/<id>.yml, or venues/families/<family>.yml when the set spans the family), replacing any hand-estimated block. Bands from fewer than 3 papers are left null and the block is marked measured-low-confidence rather than overclaiming. Full rules and the input schema: references/analysis-rubric.md §14.

Because the block carries measured + as_of, every score a consumer derives from it is labelled cache-vs-live (live / family-prior (measured, as_of <date>) / family-prior (hand-estimated) / none) — a cache-derived score never reads as if measured live.

Output

A markdown brief (default exemplar-brief-<venue>.md in the working directory, or wherever the user asks) containing:

1. Exemplar roster — the 5–8 papers with metadata, selection reason (award + source URL / citation rank + count), and OA link used
2. Venue conventions — the cross-paper synthesis across all rubric dimensions, each claim tagged with which exemplars exhibit it
3. Exemplar cards — one compact per-paper analysis each
4. Deltas & caveats — exemplar habits that conflict with the live CFP, papers skipped for lack of OA copies, unverified award claims dropped
5. Provenance — scripts run, award-page URLs, date, and the note that citation counts are a snapshot (Semantic Scholar, ODC-BY, attributed)

Plus, when measurements were taken, a cached exemplar_distribution: block (step 6) pasted into the relevant venue/family profile — the measured fallback downstream skills use when a live fetch fails, stamped measured, n, and as_of.

The brief contains only metadata and original analysis — no abstracts, no reproduced passages, no extracted figures.

References

• references/analysis-rubric.md — the analysis dimensions, what to record per paper, copyright line for outputs, synthesis + exemplar-card templates, and §14 caching the measured exemplar bundle (input schema + provenance labelling)
• references/finding-exemplars.md — award sources and the verification protocol, top-cited selection methodology and caveats, OA resolution order, alias gotchas
• scripts/build_exemplar_bundle.py — aggregates per-paper measurements into a provenance-stamped exemplar_distribution: block (offline, stdlib; invents nothing, suppresses thin bands)

Guardrails

• Never bundle paper content. No paper text, abstracts, figures, or PDFs in the repo, the brief, or any committed file — fetch on demand, process transiently, keep metadata (DOI, title, BibTeX fields) and original analysis only. Quotes ≤25 words, attributed, at most one per paper.
• Never fabricate exemplars. Every award claim needs a live source URL plus a DBLP match; every citation count comes from a script run this session; anything entering a bibliography goes through verify-citations.
• Legal OA sources only; single polite fetches (the scripts enforce ≤1 req/s per host, contact-email User-Agent, 429 backoff, caching under .cache/study-exemplars/); never bulk-harvest a proceedings.
• Venue profiles are a starting point, never ground truth — re-verify page limits, templates, and required sections against the live cfp_url before the user relies on them.
• Cache measurements, not text, and never overclaim them. The cached exemplar_distribution: block is counts/bands/section-names only (safe to commit); never put paper text in it. It is a fallback, not ground truth — a live corpus for the target venue still wins. Emit bands only from ≥3 papers (thinner → null / measured-low-confidence), never a fabricated single point, and label every cache-derived score cache-vs-live so it never reads as measured live this session.
• Studying exemplars means learning conventions, not copying — never reproduce a specific paper's text, structure verbatim, or ideas without attribution. Never submit anything to any system on the user's behalf.