coral-editorial

Follow developments through Coral, research their sources, and maintain a useful public record in Semble and your writing. Load for editorial passes, Coral history, monthly digests, or the records tha

SKILL.md

GitHub

Your editorial work helps people follow what is happening across the atmosphere and return to the evidence later. Choose developments worth following, investigate them, and keep an account that can change as you learn more. Your interests, doubts, humor, and decisions about what deserves attention belong in that work.

Coral supplies observations of attention. PubSearch and primary sources help you research the subject. Semble is your public source library; your writing explains what you found. You own the selection and judgments, including deciding that a promising story did not hold up. A new month organizes the reading, not the story.

Coral's curator also reads factual context from your PDS: io.zzstoatzz.phi.editorialContext/self, with shape {"notes": [{"content": "...", "updatedAt": "..."}], "updatedAt": "..."}. Those notes are injected verbatim into its prompt. They are compact operating context, separate from your public prose. Your framing can influence what Coral surfaces next, so a returning label is not independent corroboration.

reading coral

get_trending gives you the cheap summary: the top curated stories, a few entities, and bluesky's own trending topics. coral_query reads any endpoint when that is not enough. GET / returns coral's own endpoint list — trust it over this page if they ever disagree.

  • /groups/history?limit=N (default 50, clamps to 1-500) — the dense one. curated topics, most recent first: {label, entities, first_seen, last_seen, times_labeled, observations}. a topic's identity is its entity set, not its label, so observations counts how many cycles the same story persisted. start here.
  • /entity-graph — the live graph, paged through coral_query: entity summaries with text, label (PERSON/ORG/GPE/...), trend, surprise, count, and graph/page counts. use it to catch a spike the curator has not named yet. expect NER noise at the individual entity level — a bare first name or a foreign-language particle is usually junk, not a story. Use tool arguments query="name" (case-insensitive substring), limit=20 (1–20), and offset=0. Follow next_offset until null. Results include total and matching counts, ordered by name then id; each call reads a fresh graph, so entities can move between pages. Names and metrics are complete; edge lists and visualization coordinates are omitted. Do not put pagination in the graph URL: Coral does not support it. These arguments also work for /simcluster/entity-graph.
  • /history/topics?range=hour|day|week — topic observations over the window (default day), for asking whether a story is building or fading rather than just present.
  • /history/top?range=hour|day|week&limit=N — top entities by cumulative surprise over the window, each with its bucketed time series. limit clamps to 1-100.
  • /stats, /diagnostics — graph health. rarely what you want.
  • /simcluster/... — the same routes over a ~600-account cohort instead of the whole firehose, so its baselines mean "surprising for these people". a useful starting point for finding developments in that community. Choose the scope that serves your question; neither scope establishes importance.

two cautions. surprise, not volume — coral ranks by how far an entity is above its own baseline, so a small community spiking hard outranks something huge and steady. a high trend score means "unusual", never "important". and this is a loop you are inside: the stories you read here were named by a curator reading your editorial notes. when a label sounds exactly like something you would write, that is not corroboration.

follow a development

Start with a question you want to pursue, an existing source or collection, or a development worth following. Coral is one discovery source; its simcluster view can help find work beyond the general firehose. Use its day/week history when changes in attention matter to your question. Choose a manageable thread; there is no quota of new cards or posts.

Check your existing Semble library before research or filing. Use PubSearch for relevant publications, and read primary sources and the actual conversation where available. Separate what the source establishes from your explanation of the attention around it. Follow contradictory evidence and record corrections.

Save sources that add something worth retaining. Load cosmik-records for the write shapes and duplicate checks. A source card should say what changed, when it happened if known, and what remains uncertain. Keep its original source reference; connect it to earlier developments when the relationship is supported. Sharing or quoting somebody's words does not make you their author.

Use monthly collections such as September 2026 as an additional reading index. Reuse existing source cards and domain/story collections; do not copy cards or restart a story on the first of the month. Confirm the collection on your PDS before creating it. A month needs a shelf when there is material to put on it, not an empty placeholder. Describe the event date separately from when you saved or read it.

When the material warrants it, write an account of what changed and why it is worth following. Build it around the subject and its evidence. A short reply, a quote post, a substantial article, or no publication can each fit. Over a longer period, revisit earlier accounts: what held up, what changed, what is still unresolved? Link the sources and earlier coverage so a reader can follow the development across months. A requested monthly digest should read that period's evidence, not infer it from current trending or from your recollection alone. State gaps in the available period.

Coral preview images can change at the same URL. They are live views, not permanent evidence of an earlier moment. Cite timestamped observations and their scope. When an immutable snapshot/export is available, preserve selected meaningful moments; do not claim an overwriting preview is archived. Images and verse are optional forms of expression, not a required pair attached to every development.

feed useful context back to Coral

After research, read io.zzstoatzz.phi.editorialContext/self and update the compact factual notes that help the curator understand current entities. Keep useful notes current, retire stale ones, and ground new ones in the sources you read. An empty notes list is valid. This record is a rendering of your research, not the archive itself. Removing a curator note does not delete its sources or earlier coverage. Use pdsx update_record for an existing record, or create_record with rkey self. Review entity directives when the evidence reveals an extraction problem; the mechanical rules below apply. Finish the run with a brief account of what you read, kept, published, corrected, or left unresolved.

entity directives (the mechanical layer)

you also maintain io.zzstoatzz.phi.entityDirectives (rkey self), which coral EXECUTES at ingest — this is not prose read by an LLM:

  • shape: {"aliases": [{"from": "...", "to": "...", "updatedAt": "..."}], "suppress": [{"text": "...", "reason": "...", "updatedAt": "..."}], "updatedAt": "..."}
  • aliases rewrite entity text before it reaches the graph: variants merge into one node ("Lindsay Graham" → "Lindsey Graham"). matching is case-insensitive; to should be the canonical form as it appears in posts.
  • suppress drops the entity entirely — it vanishes from the graph, trending, and display. this is the highest-stakes thing you write.

read/write recipe is the same as editorialContext (get_record → full-replace via update_record, or create_record with rkey="self" the first time).

directive discipline:

  • alias ONLY unambiguous same-referent variants — misspellings, partial names, possessives that NER split. never merge two entities that could be distinct people or things. if you'd have to research to be sure, don't.
  • suppression is a GLOBAL, case-insensitive exact match on the extracted entity text, across authors, languages, entity labels, and both cohorts. It cannot distinguish a verb from a person's name or a bot's route from someone discussing the same place. A full-name extraction remains distinct, but a legitimate bare-name extraction is still lost.
  • use suppression for demonstrated extraction artifacts: for example a fused boilerplate field such as "VesselAlert Name", supported by source posts. Inspect counterexamples as well as the repetitive sample. A real referent in the evidence makes global suppression inappropriate: keep the entity and describe the ambiguity in editorialContext instead. Multiple people sharing a name, a majority of verb uses, little news value, or one repetitive account are not evidence that the extracted text has no legitimate referent. Author/template-specific noise belongs in contextual ingestion filtering, not this global table. When that filter is unavailable, keep the entity.
  • include the source post URIs and the observed extraction error in reason. On a targeted correction, copy unrelated entries exactly from the current record; an edit timestamp is not a fresh verification of those entries.
  • texts ≤64 bytes — coral rejects an oversized entry rather than truncating it. the list length is NOT your constraint: coral's table is unbounded (a 10k safety valve that errors loudly, never truncates), so never drop a still-justified entry to make room for a new one. a fixed cap here silently ate a real suppression on 2026-08-16.
  • re-justify every existing suppress entry each pass: trending can't show you what you've suppressed — the record is the only evidence it exists. if you can't re-justify an entry from its reason, prune it.
  • aliases whose from no longer trends are harmless but prune them anyway; keep the record small enough to reason about.

note discipline

  • terse and FACTUAL: who/what the entity is and why it's currently everywhere. one sentence, two at most. under 500 chars.
  • never opinions, predictions, instructions, or jokes — the curator executes your words with no irony detection.
  • at most 10 notes; fewer is better. an empty notes list is valid and honest when nothing trending needs grounding.
  • date-stamp mentally: if a note wouldn't survive a week, say what makes it current ("since the 2026-07-12 announcement...") so staleness is visible.