--- name: ccfddl-x-posts description: Prepare and, with a verified authenticated publishing route and authorization, publish daily conference-deadline countdown posts for @ccfddl on X/Twitter. Use for the recurring CCFDDL campaign, AI/Data Systems countdown text, email-notification-style social cards, schedule setup, publication checks, or maintaining this workflow. Read the live CCFDDL source, include CCF A and B plus ICLR and MLSys, and link every digest to ccfddl.com. --- # CCFDDL X Posts Create up to two English digests per publication day: one **AI**, one **Data Systems**, with up to six rows each. After the rows and one blank line, include `see details: ` immediately followed on the same line by the most-starred selected conference’s full detail URL, followed immediately on the next line by `#ccfddl #conf_deadline #蓝v`. Omit the entire `see details: ccfddl.com` line. X controls whether a link preview appears; do not attach a generated PNG by default. ## Campaign contract - Use exactly the verified X account **@ccfddl**. Do not treat a display name, brand name, or workspace as account proof. - Include CCF A and CCF B conferences plus explicit ICLR and MLSys exceptions. Classify `sub: AI` as AI and **all remaining source subjects** as Data Systems. This is the requested campaign grouping, not a claim about the source taxonomy. - Publish in the morning in **America/Los_Angeles**, with automatic daylight-saving changes. Prefer flexible scheduling around 08:00 local; do not impose an exact time unless requested. - Use English as the stated operational default. Preserve later user changes in campaign configuration. - Publish at most one digest per category per local calendar day. Choose future abstract or paper deadlines within the configured 120-day horizon, up to six text rows per category. When the user explicitly selects the optional generated-card fallback, show only the first three rows in the image, followed by an ellipsis when more eligible rows exist. Keep all candidates when they fit; on overflow, retain representation of both eligible CCF A and CCF B tiers instead of allowing nearer B deadlines to displace every A conference. Record any omissions in the manifest and surface material omissions rather than silently hiding a tier. Use only actually eligible entries when fewer than the limit exist; never invent or duplicate rows to fill capacity. Configure `max_items_by_category` as `{"AI": 6, "Data Systems": 6}`. Keep the 120-day horizon, nearest-stage selection, sorting and tier representation unchanged. Skip empty categories. - Include no unrelated posts, replies, DMs, follows, or engagement actions. Do not put credentials, account identifiers beyond the public handle, private state, or publication ledgers in the public source repository. - Treat this skill as the workflow, not proof that a schedule or publishing connection exists. Never report live posting as enabled without verifying the actual provider and schedule. Read [references/config.json](references/config.json) for editable campaign settings. Read [references/source-and-publishing.md](references/source-and-publishing.md) for source mapping, identity, conflict handling, provider checks, and state transitions. ## 1. Check the connection and authorization Before creating a scheduled task, perform a harmless read through the intended publishing route and verify an authenticated X account whose exact normalized handle is `ccfddl`. Prefer a suitable connected provider; an authorized authenticated X browser is an alternative when no suitable provider is available. For a provider, discover its tools and supported media/publication operations; do not guess them. For a browser route, inspect the actual signed-in profile and composer and verify that the scheduling environment can use that authenticated session. Never assume the current browser login will be available to a future task. If the chosen route is unconnected, unauthenticated, or points at another account, prepare previews and report the precise blocker. Use the supported plugin or browser login setup flow. Respect cancelled login/setup requests and wait for the user to continue rather than automatically reopening them. Metricool is optional; do not require it or subscribe to a paid service without authorization. For Metricool, a connected plugin alone is insufficient: read brand settings and verify an actual connected X network. Do not guess the `info` JSON schema or media-file format for `createScheduledPost`; inspect current tool documentation and official provider documentation first. Do not request or store an API key or password in chat. Installing this skill does not connect X. Default to `media_mode: link_preview`: submit the approved text with one most-starred-selected-conference detail URL and no separate image attachment or homepage footer. Verify that URL’s reachability, canonical, edition and sharing metadata. Successful metadata validation is not proof X will display a preview. Validate the actual account/provider’s full-post capability and report uncertain rendering. For an explicitly requested generated-card fallback only, apply the public-image gate in the source-and-publishing reference. Use the user's bounded authorization for this recurring campaign. Ask before expanding its audience, topic, account, or scope. Obtain any additional action-specific confirmations required by the platform or applicable policy. Do not publish a preview merely to test the connector. ## 2. Acquire a coherent current source Use the verified canonical source: `https://github.com/ccfddl/ccf-deadlines`, directory `conference/`. Fetch the latest default-branch commit read-only and pin every YAML file to that same full 40-character SHA. Use that latest repository snapshot as the exact publishing source for text, cards, deadline selection and countdowns. Official-site discrepancies may be recorded for a separate data-review task; they must not silently replace repository timestamps/timezones or remove otherwise eligible conferences during publication. Use the bundled snapshot command on a fresh, read-only checkout, or construct the same snapshot with authorized connector results. Do not edit the conference repository as a side effect of publication. ```bash python3 /scripts/countdown.py snapshot --repo --output /snapshot.json ``` The snapshot stores `repository`, `commit`, `fetched_at_utc`, `complete_catalog`, and `files` (`path`, exact UTF-8 `content`, SHA256). When assembling with a connector, set `complete_catalog: true` only after confirming all conference YAML paths at that commit were retrieved. Sample subsets must have `complete_catalog: false` and remain previews. `snapshot` reads committed Git objects, never uncommitted working files. Never convert `TBD`, historical patterns, website estimates, conference dates, decisions, rebuttals, or camera-ready dates into submission deadlines. Preserve source seconds. Resolve AoE as UTC−12, fixed UTC offsets explicitly, and IANA zones with date-specific DST. Reject incomplete dates and ambiguous/nonexistent local DST times. Interpret PT as America/Los_Angeles. Quarantine unknown timezone labels rather than guessing. Do not apply independent official-clock overrides or external-source conflict quarantines during publication. In particular, use EDBT’s repository deadline/timezone unchanged and include eligible MLSys repository entries through the existing explicit exception. Preserve CCF A/B plus ICLR/MLSys scope, nearest-stage selection, horizon, TBD handling, precise seconds and normal source-parsing safeguards. Do not edit conference YAML as a side effect of this workflow. Keep known outside-source discrepancies in separate data review, not as publishing blockers or silent substitutions. ## Star-count snapshot for the detail link Before building each run, fetch `https://ccfddl.com/api/bootstrap` anonymously and require HTTP 200 with the complete JSON response `{"user": null, "counts": {...}, "starred": []}`. This public endpoint’s `counts` keys are repository edition IDs, not conference-series names or GitHub stars. Fetch once for both category digests and record actual fetch time; the server may cache anonymous counts for up to 300 seconds. Save the complete public response body outside the skill, then freeze it with: ```bash python3 /scripts/countdown.py stars \ --response /bootstrap.json --fetched-at \ --http-status 200 --output /stars.json ``` Use `--http-status 200` only after actually verifying that successful response; do not fabricate response contents or refresh the timestamp on cached local data. The snapshot stores source URL, exact public response and SHA256, fetch UTC, complete-response marker, edition-ID semantics and the known server-cache bound. `build --stars` verifies schema, integrity and freshness (at most 15 minutes old and at most one minute in the future), including preview builds. The API returns all database counter rows without pagination and deletes rows when their count becomes zero. Therefore an absent edition ID means zero **only within this validated complete sparse response**. Missing/malformed/truncated response, wrong endpoint, failed fetch, missing edition ID, invalid count, stale snapshot or broken hash must stop the build. Never default those failures to zero or silently use the first row. A successful empty complete counts object is the documented all-zero case; ties retain current row order. Star counts affect only the linked conference among the selected maximum six. Do not sort the deadline rows by stars, add an unselected popular conference, change deadline facts or alter normal category/CCF/horizon selection. Record selected edition counts, featured edition/count, source provenance and tie policy in the private manifest. Keep counts out of tweet text unless requested. ## 3. Build and inspect Use a current explicit UTC as-of time. Require Python 3.10+, PyYAML, Pillow, and Liberation Sans (Arial-compatible). Use the executor's existing packages or an authorized reputable package source; do not silently install from an unknown source. ```bash python3 /scripts/countdown.py build \ --snapshot /snapshot.json --config /references/config.json \ --as-of --stars /stars.json --output /posts --preview ``` For an actual publication run, omit `--preview` after all gates are satisfied. A non-preview build requires a complete catalog, confirmed editorial configuration, and a snapshot fetched within 15 minutes. Recheck latest source SHA immediately before sending; if it changed, regenerate the affected content before posting. The following visual rules apply only to the optional generated-card fallback (`media_mode: generated_card`), which requires user authorization before replacing the default link-preview route. Inspect every fallback PNG with an image-view tool. Confirm legible text, no overlaps/clipping, correct category, approved heading and event labels, exact deadline and zone, branding, and no crossed-out footer elements. Keep as-of UTC and source SHA in the private manifest only; retain the canonical link in tweet text, not in the image. Adapt the live repository email reminder template (`worker/src/email_reminders.js`) for social cards: gray page, white bordered cards, brick-red reminder accent, Arial-compatible sans-serif typography, and countdown → event title → exact deadline hierarchy. Use exactly `NAME YEAR Abstract Deadline` for abstract events and `NAME YEAR Deadline` for paper events, retaining a meaningful round label. Do not display an `abstract closed` warning in tweet text, cards, or alt text; keep the factual flag in the private manifest only. Keep the card heading as `CCFDDL deadline reminders (AI)` or `CCFDDL deadline reminders (Data Systems)`, without the tweet-only publication date. Remove the in-image website CTA and the visible as-of/source footer; preserve top branding, REMINDER, DEADLINE NOTICE, and the good-luck/team closing. Use the final category-specific plain-text contract below for tweet wording and retain category names on the attached cards. Omit unsubscribe/settings controls and retain ccfddl.com in tweet text only. Preserve precise elapsed-time countdowns rather than copying email calendar-day arithmetic. Every countdown is a snapshot, not a live clock. Inspect the manifest and each tweet. Honor the skipped-for-review list. Preserve round labels and distinguish abstract from paper. Keep the passed-abstract fact internally, but omit its public qualifier as requested. Such paper deadlines apply to already-registered submissions; do not claim a new submission is still possible. Keep `stage_selection: nearest` as the default: select the next future deadline per timeline round deterministically. When multiple stages are explicitly selected, support separate rows with `stage_selection: all_future`; do not change the campaign default merely because the wording changed. Preserve every selected stage and display every `abstract_deadline` as `abstract` in both text and card, even when the official site calls it registration. Keep the official term only in source provenance, never as a reg/registration display override. Keep the full selected digest in the text (up to six rows per category). For the optional generated-card fallback, render only its first three rows in the card and add an ellipsis if additional eligible entries exist. Do not add an ellipsis when there are no additional entries. Preserve `events` (all text rows), `card_events` (visible image rows), and `card_has_more` separately in the manifest. Make alt text describe only the visible image rows and the ellipsis, without claiming all text rows are pictured. Keep the default digests concise. Treat the script's weighted-length count and 280-character advisory as editorial diagnostics, not a universal hard limit: the connected account may support Premium long posts. Check the actual provider/account capability. Keep the approved category-specific compact style within the verified account limit; if it does not fit, report the tradeoff instead of silently dropping CCF A entries or changing the requested format. The script counts the sole full conference URL as 23 characters, NFC-normalizes Unicode, and conservatively overcounts complex emoji. Never use a naive character count, silently truncate, or automatically split a digest into a thread. If the provider returns a text-too-long error, report the exact error and do not rewrite and retry without new authorization. Include alt text from the manifest when the selected route supports it. ### Final plain-text contract Use exactly `CCFDDL daily reminders (AI) · YYYY/MM/DD` as the AI first line and `CCFDDL daily reminders (Data Systems) · YYYY/MM/DD` as the Data Systems first line, with no trailing colon. Resolve `YYYY/MM/DD` from the explicit as-of time in `America/Los_Angeles`, including daylight-saving changes; add the date to tweet text only, keeping the card heading unchanged. Render each selected event as `NAME'YY (stage, round N) · N days`, omitting `, round N` for a single-round conference. Omit `left` from day-based tweet countdowns (use `1 day` for singular). Keep card countdown wording unchanged. Use the existing hour/minute countdown for near-term deadlines. Put one blank line after the heading and one blank line before `see details: ` followed on the same line by the most-starred selected conference’s full canonical detail URL. Follow the URL immediately with exactly `#ccfddl #conf_deadline #蓝v` on the next line. Omit the entire `see details: ccfddl.com` line and any additional homepage link. Keep hashtags out of the image. Keep exact dates and timezones on the website rather than repeating them in the compact tweet. Optional fallback cards retain their three visible rows. Use ordinary text with no Markdown bold or Unicode bold. Uniformly label `abstract_deadline` as `abstract` in the tweet and `NAME YEAR Abstract Deadline` in its matching card; do not relabel CVPR as reg or registration. Do not invent dates from examples. Omit `abstract closed` from all public output while retaining its private fact flag. Do not claim that new registrations or new submissions are still possible. The image retains the AI/Data Systems category. ### Single-link footer and publication preflight After the selected rows and one blank line, use `see details: ` followed on the same line by the most-starred selected conference’s full detail URL, then `#ccfddl #conf_deadline #蓝v` on the immediately following line. Keep the full canonical URL in the generated text; X may display it in abbreviated form. Do not substitute a t.co short link or invent a category landing page. Do not add a standalone homepage line, a second link, hidden links, independent link attachments, or PNG by default. Choose the highest-starred edition only after digest selection and ordering; preserve all selected rows and their existing deadline order. Break equal counts by the earliest selected row. Use the fresh validated public-count snapshot described below. Match the website route `https://ccfddl.com/venues/{source-subject-lower}/{slug}-{full-year}/`, with the source title lowercased, non-ASCII-alphanumeric runs replaced by hyphens and trimmed; fall back to the edition ID slug only when the title slug is empty. Use the source subject, not the campaign category Data Systems. Require the conference URL to return HTTP 200 with the correct canonical/edition and inspect its sharing metadata and referenced image. X may omit a preview even with valid metadata; keep `link_preview_guaranteed: false` and `preview_selection_guaranteed: false` in the manifest. No independent card-target capability is needed. Preserve every selected row up to the six-per-category limit. Count the one link as 23 weighted characters. If total length exceeds 280, retain `needs_publication_capability_validation: true` and verify the actual @ccfddl provider/account accepts the full single post. Never compress, drop selected rows, split posts, hide or move the link, or switch media modes without authorization. Report exact provider errors and uncertain preview results. ## 4. Publish idempotently Use a stable, persistent campaign ledger shared by all runs, or equivalent provider-native durable idempotency. Do not use a temporary directory for live state. If durable state is unavailable, stop before sending. Compute the publication day in America/Los_Angeles. Check both the ledger and provider's recent/scheduled posts for an existing @ccfddl digest for that day/category. Reconcile existing provider records before proceeding. Never create both a direct post and a separate provider-scheduled copy. ```bash python3 /scripts/countdown.py ledger check --path --day YYYY-MM-DD --category AI python3 /scripts/countdown.py ledger reserve --path --day YYYY-MM-DD --category AI --content-sha256 ``` Use the verified authenticated @ccfddl publishing route with the approved full text, the single conference URL and hashtag footer, and no PNG attachment in default link-preview mode. If the composer still contains an earlier image attachment, remove it before final inspection. Complete the single-link and length preflight above before reserving. Only use generated-card upload/public hosting when the user explicitly selects that fallback; never silently replace a missing link preview with an attached card. Reserve immediately before the single publication request; the reservation uses an exclusive file lock and atomic state replacement. Use the exact approved text and single-link footer (or explicitly authorized fallback card and alt text). If an optional upload fails before any post request, determine the exact state without duplicate publication. If sending times out or returns an uncertain result, leave the entry pending, inspect the provider and X, and **do not resend automatically**. A pending reservation blocks additional sends. Reconcile manually only after establishing whether publication occurred. After the provider confirms publication, independently read the post and verify the @ccfddl author, full text, the conference link, actual preview or its absence, and exact post ID before confirming the ledger: ```bash python3 /scripts/countdown.py ledger confirm --path --day YYYY-MM-DD --category AI --post-id --post-url https://x.com/ccfddl/status/ ``` A queued/scheduled job is not a published tweet; retain its provider job ID in durable campaign state and confirm only after read-back. Report material blockers or conflicts and link verified published posts. Follow the user's notification preferences and avoid repetitive daily success messages. ## 5. Create or maintain the schedule Only after the authenticated-route read succeeds, create a daily morning task with the user's supported scheduler. Preserve `America/Los_Angeles` in the schedule; use flexible timing around 08:00 local. The task must invoke this workflow, refresh source data, perform read-back and idempotency checks, and keep private state outside the public repository. Verify that the task is enabled and tied to the correct provider/account before claiming it is active. Do not create a second schedule if one already exists. Maintain the reusable skill in the user's personal skill directory. When separately authorized for repository maintenance, mirror only the portable skill files to `.agents/skills/ccfddl-x-posts/` through the repository's normal review workflow. Keep the personal installed copy and repository copy aligned when updating behavior. Never commit generated posts, snapshots, private ledger state, credentials, or provider tokens. ## Validation Run the deterministic test suite after changing dates, selection, rendering, or ledger logic: ```bash python3 /scripts/test_countdown.py ``` Forward-test a new representative source snapshot, inspect the single conference link and its behavior and any optional rendered cards, and exercise the uncertain-send/duplicate guard before enabling changed live behavior. Treat tests and preview creation as validation, never as proof that a post was sent or the schedule is active.