Query deadlines by keyword or idea (portable)

Published copy (for agents)

This file is deployed with the AI Conference Deadlines site. Agents may fetch the source markdown from GitHub (same text as in this repository):

Raw SKILL.md (for agents): https://raw.githubusercontent.com/mlciv/ai-deadlines/gh-pages/static/skill/SKILL.md
(Replace mlciv/ai-deadlines with your fork’s user/repo if needed.)

Human-readable page (built by Jekyll): /ai-deadlines/static/skill/SKILL.html on the site.

Dataset backing this deployment (when applicable): conference YAML under _data/conferences/ and _data/types.yml in the source repository.

Audience

This file is agent-agnostic: any assistant (Cursor, OpenClaw, CLI agents, IDE plugins, etc.) can follow it if the instructions are loaded into context or referenced as a skill URL. No editor-specific APIs are required.

Goal

Produce accurate, sourced answers to: “What conferences / deadlines match this keyword or idea?”

Use either or both:

Structured data — YAML files listing conferences (see Data layout).
Web search — official call-for-papers (CFP) pages when data is missing, predicted, or stale.

Never fabricate deadlines.

Data layout (when a repo or dataset exists)

Typical ai-deadlines–style trees:

types.yml (or equivalent): maps human-readable area names to short codes (sub), e.g. ML, CV, NLP.
conferences/*.yml (or _data/conferences/*.yml): one or more YAML documents per file; each conference is a list item with fields such as:

title, year, id, link, deadline, abstract_deadline, timezone, place, date, start, end, sub, note.

If the user’s workspace has no such paths, skip file search and rely on Web-only mode.

Convention: sub may be a string or a list. Match if any tag fits the user’s topic.

Subject tags (`sub`)

Map user language to tags using the project’s types file when present; otherwise use this table:

User might say	Typical `sub`
machine learning, deep learning	`ML`
computer vision, images, video	`CV`
NLP, LLMs, ACL-style	`NLP`
robotics, embodied	`RO`
speech, audio	`SP`
data mining, recsys	`DM`
planning	`AP`
knowledge representation	`KR`
HCI	`HCI`
AI + education	`EDU`
graphics	`CG`

For narrow topics, also search title, full_name, and note text (e.g. “dialogue”, “3D”, “medical”).

Workflow

Normalize the query
Extract 1–3 keywords; assign one or more sub values. If ambiguous, ask one short question or cover the most likely tags.
Search structured data (if available)
Grep or search all conference YAML for matching sub: and optional text fields. Respect the workspace root the user provides.
Collect fields
Per match: title, year, id, deadline, abstract_deadline, timezone, place, date, link, note.
Sort by urgency
Prefer deadlines after “today” (use user-provided or system date). Sort by main submission deadline unless only abstract/ARR dates exist—then state that clearly.
Flag uncertainty
If note (or absence of official link) suggests Predicted, Estimated, TBA, or when available, label the row unconfirmed and recommend the official CFP URL or a search: "[conference acronym] [year] call for papers".
Web verification
For unconfirmed rows or user requests for “latest” dates, open official domains (conference or society site) and reconcile with YAML.

Web-only mode

If there is no local YAML:

Infer venues from keywords (e.g. vision → CVPR/ICCV/ECCV; ML → NeurIPS/ICML/ICLR; NLP → ACL/EMNLP/NAACL).
Use web search for current year and next year CFP pages.
Present deadlines with source URL; prefer pages hosted by the conference or ACM/IEEE/CVF as appropriate.

Output format

Use a compact table or bullets:

Conference (year)
Deadlines (abstract + full if both exist) and timezone
Venue / conference dates (short)
CFP or home link
Confidence: official dataset official web predicted / unconfirmed

Examples

“Vision + multimodal deadlines” → tags CV (+ NLP if needed); grep multimodal/MM; sort by date.
“Spoken dialogue” → grep dialogue/speech; include NLP and SP venues from data or web.
“NeurIPS-style next cycle” → ML venues; separate official vs predicted in notes.

Loading this skill in other agents

HTTPS (raw markdown): Fetch the raw GitHub URL for static/skill/SKILL.md and inject the response into context.
Repository: Clone the repo and read static/skill/SKILL.md (or .cursor/skills/query-deadlines-by-keyword/SKILL.md if present).
Content: The agent may use the markdown body (after the first YAML front matter block); the name and description fields help discovery where supported.
Repo coupling: When this skill ships inside a clone of ai-deadlines, default data paths are _data/types.yml and _data/conferences/*.yml. When used standalone, ignore those paths and use Web-only mode or ask the user for their dataset root.

Display notes

Strip HTML from note when speaking plain text to the user. Sorting in a static site UI may differ from agent-side date sorting; always compute order from parsed deadline strings and timezone when possible.