Parse CFP

Turn a live Call for Papers into the machine-readable venue profile that the rest of this repo runs on. Output is a venues/conferences/<venue>-<year>.yml file conforming to venues/schema.yml, plus a human-readable requirements card. Downstream consumers: preflight-check, tailor-to-venue, write-rebuttal, prepare-camera-ready, select-venue.

When to use

• The user pastes a CFP, author-guidelines, or submission-instructions URL.
• "What are the requirements / deadlines / page limits for VENUE YEAR?"
• A profile in venues/conferences/ is missing or a year out of date.

Inputs

• A CFP URL. If the user only names a venue, check venues/conferences/<venue>-<year>.yml for a cfp_url first; an older year's profile usually links the venue website where the new CFP lives.
• venues/schema.yml — read it before writing the profile; it defines every field and the mandatory verified: provenance block.
• The matching venues/families/<family>.yml for family-level defaults.
• CONTACT_EMAIL env var (used in the polite fetch User-Agent; the script prompts if unset).

Process

1. Scope one venue + one year. Read venues/schema.yml. If a profile for this venue-year already exists, treat it as hints to re-verify, never as ground truth.

2. Fetch the CFP page. Run:

   python3 scripts/fetch_cfp.py "<cfp-url>"
   

   The script fetches ONE page per invocation, rate-limits to 1 request/second per host, backs off exponentially on HTTP 429, caches under .cache/ (gitignored) for 24h, and prints readable text with link targets as [url]. It exits nonzero with a clear message on failure. Then fetch the handful of follow-up pages you actually need — per-track submission pages, the deadlines page, the camera-ready/author-kit page — one invocation each, never a site crawl.

   • Exit code 3 = non-HTML content (a PDF CFP, a JSON API response); the body is saved under .cache/ — read that file directly.
   • "very little text extracted" warning = JavaScript-rendered page; ask the user to paste the page text and mark the profile needs-verification. Never fill gaps from memory.

3. Verify the year. Confirm the page heading and dates match the requested year before extracting — stale CFPs survive at near-identical URLs. Wrong year: tell the user and locate the current page.

4. Extract every schema field following references/extraction-guide.md. The non-negotiables:

   • Numbers (page limits, char limits, dates) only from fetched text; missing fact = null + an explanatory note. Never infer, never recall.
   • Page limits must capture what is EXCLUDED (references? appendix? checklist?) with the CFP's sentence quoted verbatim in the track notes.
   • Deadlines: record the stated timezone — AoE is common, not universal (SIGSPATIAL uses Pacific Time). Abstract registration is a separate, earlier deadline at most venues.
   • llm_policy and dual_submission: verbatim quotes, never paraphrased.

5. Normalize the template invocation, submission system, rebuttal format, and camera-ready rail using references/templates-and-systems.md. Derive \documentclass options from the family table (add review,anonymous only for double/triple-blind venues) and comment when an invocation is family-derived rather than printed in the CFP.

6. Emit the YAML to venues/conferences/<venue>-<year>.yml:

   • id equals the filename stem; family references an existing family file.
   • Fill the mandatory verified: block — today's date, every URL fetched (with a comment on which facts each supplied), and an honest confidence: (verified-live only if all critical facts came off live pages today).

7. Validate. Run python3 tools/validate_venues.py from the repo root (needs PyYAML; if unavailable, check the file manually against venues/schema.yml field by field).

8. Present the requirements card to the user: a short table of deadlines (with timezone), per-track limits with their exclusions, blind level, template invocation, submission URL, rebuttal format, and LLM policy. End with the verification notice: "Profile extracted from <cfp_url> on . Re-verify deadlines and page limits against the live CFP before submitting — CFPs change mid-cycle and a stale limit can cause a desk reject."

Worked mini-example

User pastes https://sigspatial2026.sigspatial.org/research-submission.html. After fetching, the Research-track section reads: "papers are limited to 10 pages (excluding references), with up to 2 additional pages after the references to be used for appendices", deadlines are stated as "11:59 PM Pacific Time", and the submission link points at easychair.org/conferences/?conf=acmsigspatial2026. That extracts to:

id: sigspatial-2026
deadlines:
  abstract: 2026-05-29
  paper: 2026-06-05
  timezone: PT          # CFP says "Pacific Time", NOT AoE — recorded verbatim
tracks:
  - name: Research
    page_limit: 10
    page_limit_excludes: [references, appendix]
    notes: >
      CFP verbatim: "papers are limited to 10 pages (excluding references),
      with up to 2 additional pages after the references ... for appendices."
review:
  blind: single         # CFP lists author names — single, not double
  submission_system: easychair
verified:
  date: 2026-06-21
  source_urls:
    - https://sigspatial2026.sigspatial.org/research-submission.html  # limits, deadlines, blind
  confidence: verified-live

Two traps this example surfaces and the guide handles: the timezone is Pacific, not the AoE default, and the page limit excludes both references AND the appendix — encoded explicitly so preflight-check can enforce it. The page heading is confirmed to say 2026 before any of this is extracted.

Output

• venues/conferences/<venue>-<year>.yml — schema-conformant, year-versioned, with provenance.
• An inline requirements-card summary with the re-verify warning.

Guardrails

• Never fabricate a deadline, limit, or policy. null + note beats a guess.
• Quote LLM-use and dual-submission policies verbatim; paraphrase can invert compliance meaning.
• GET-only. Never log in to, register with, or submit to any submission system on the user's behalf.
• Never fabricate citations. This skill emits venue metadata only; if a paper reference is ever needed downstream, route it through verify-citations.
• One page per script invocation; no bulk crawling. The cache lives under .cache/ and stays out of git — do not commit fetched CFP text to the repo.
• Always include the re-verify-against-live-CFP notice in the final answer.

Source verification (mandatory)

This skill originates venue facts, so source verification is mandatory without exception:

• The venue is the user's choice; build its rules on demand, never from memory or a prior year.
• Search the web for the live CFP / author-instructions page and read the primary source — never state a fact from a search snippet you didn't open, and never from a deadline aggregator without confirming on the official site.
• Double-check every desk-reject-class fact (deadline, page limit + exclusions, blind level, template, submission system) against a second source; if sources disagree, surface the conflict — do not pick silently.
• Record a clickable reference link and the date for every fact in the profile's verified.source_urls; set confidence to verified-live/corroborated only when sourced, else needs-verification. A fact with no link is emitted as unverified, never asserted.

Bundled files

• references/extraction-guide.md — field-by-field extraction rules, cue phrases, and the trap list.
• references/templates-and-systems.md — template invocations by family, submission-system URL patterns, rebuttal taxonomy, camera-ready rails.
• scripts/fetch_cfp.py — polite single-page fetcher (run it; do not reimplement fetching by hand).