curating-bsky-list

Builds and maintains Muninn's own Bluesky curate list ("Muninn reads", on the muninn.austegard.com account). Accounts are chosen by per-topic classifiers on Ettin-17m post embeddings (topics are Munin

SKILL.md

GitHub

Curating Muninn's Bluesky list

scripts/curate.py owns one Bluesky curate list, "Muninn reads", stored in the ATProto repository of Muninn's account (muninn.austegard.com, did:plc:tc43adcqjhjmtnncm723ztvv). It finds accounts, scores them, has Claude Haiku judge them, writes the list items, and keeps its state (the list's URI and URL, each member's kind, topic, scores and dates) in Muninn's config key muninn-bsky-list-state. Muninn's part is the reading: it saves insights from the list's posts as memories.

Members are of two kinds. Topic members write mostly about one of the topics below. Exploration members (about a third) fit no topic and are kept only while reading them produces saved insights.

Boundaries

  • The script writes nothing to Bluesky except list items on that one list, plus creating the list on the first run. scripts/bsky.py signs in only with MUNINN_BSKY_HANDLE / MUNINN_BSKY_APP_PASSWORD, refuses a session whose DID is not Muninn's, and refuses a list URI outside Muninn's repository. The environment may also hold BSKY_HANDLE / BSKY_APP_PASSWORD, which are Oskar's; the script never reads them.
  • Do not post, reply, like, repost or follow from Muninn's or Oskar's account while using this skill, and do not follow list members: the list is how Muninn reads them.
  • Candidates already on Oskar's list "Artificial (and real) Intelligence" are skipped, so the two lists do not duplicate each other. Bridgy mirrors of Mastodon accounts (*.ap.brid.gy) are skipped too.

Requirements

Python 3.10+ and pip, nothing else preinstalled: numpy, onnxruntime, tokenizers and scikit-learn are pip-installed on first use, and the Ettin-17m embedding model (13 MB) is downloaded to ~/.cache/curating-bsky-list/ettin/ on first use.

needs for
MUNINN_BSKY_HANDLE, MUNINN_BSKY_APP_PASSWORD post search (Bluesky answers it only when signed in) and the list writes
TURSO_URL, TURSO_TOKEN the state key, and the insight memories that decide whether exploration members stay
claude CLI on PATH the Haiku judge, claude -p --model haiku (without it, see "Judge fallback")
egress bsky.social, public.api.bsky.app, *.host.bsky.network (Muninn's and Oskar's account hosts; Muninn's is stropharia.us-west.host.bsky.network), plc.directory, raw.githubusercontent.com, the Turso host, pypi.org and files.pythonhosted.org

--state-file PATH keeps the state in a local JSON file instead of Turso. Insight memories are then unreadable, so the exploration drop rule is skipped rather than applied to zero insights.

Commands

Each prints one JSON object on stdout and logs progress to stderr. Global options (--state-file) go before the command name.

S=<skill dir>/scripts
python3 $S/curate.py build   [--dry-run] [--out build.json]    # first run: create the list, fill it to 120 members
python3 $S/curate.py refresh [--dry-run] [--out refresh.json]  # weekly
python3 $S/curate.py read    [--days 7] [--out digest.md]      # also writes the digest (markdown) to --out
python3 $S/curate.py status                                    # list URL, member split, recent runs
python3 $S/curate.py add-topic "river ecology" --terms "salmon run,watershed,river restoration" [--lang any]
python3 $S/curate.py set-param topic_drop 0.15                 # override a threshold; 'default' clears it
python3 $S/curate.py restore-state muninn-bsky-list-state.backup.json

--dry-run computes everything and writes nothing, to Bluesky or to the state. build refuses to run when the state already holds a list. add-topic --lang takes a language code for the search; the default is English, and any searches every language. build and refresh also take --verdicts FILE (see "Judge fallback") and --dump FILE (every scored candidate, for tuning). The first build took 15 minutes and a refresh 10 (2026-11-08, 4 CPUs), most of it searching about 170 terms and fetching and embedding up to 100 posts from each of about 1,000 candidates. read takes about 20 seconds.

The weekly routine

  1. Run refresh --out refresh.json. Exit 0 is success. Exit 3 means the judge is missing: follow "Judge fallback". Any other exit: report the error field and stop.
  2. Store the refresh's memory field verbatim with Muninn's remember tool: type experience, tags muninn-bsky-list, bsky-list-refresh and the date as YYYY-MM-DD. The script cannot write memories, so this is the only record of the week's changes outside the state key.
  3. Run read --out digest.md and read the digest. Each line carries the date, @handle, the member's DID, its topic or explore, an engagement count and the post URL, followed by the post text.
  4. Save two to five insights, each its own remember call: type world for a fact about the world, analysis for an argument or interpretation; tags bsky-list-insight, muninn-bsky-list and the date. Write each as YYYY-MM-DD bsky-list insight from @handle (did:plc:…): <what was learned, in one to three sentences>. <post URL>. Next week's refresh counts an exploration member's insights by looking for its DID (or, failing that, its current handle) in live memories tagged bsky-list-insight. Handles change; DIDs do not, so include the DID.
  5. The refresh output's themes lists words and phrases that stand out in the posts of three or more exploration members, compared with Discover posts. When one names a subject worth following, promote it with add-topic NAME --terms "…" (three or more search phrases). From the next run it is a topic like the others, with its own classifier and share of slots. A promoted topic lives in the state's extra_topics; remove it by editing the state and running restore-state.

Selection pipeline

Topics. Nine of Muninn's interests (copied into assets/topics.json from its profile key intellectual_interests; edit the file when the profile changes) and one topic per cluster of Oskar's experiments atlas (oaustegard/experiments experiments-atlas/data.json, 12 clusters, fetched fresh each run). A cluster's search terms are the seeds in assets/topics.json followed by words extracted from its experiments' one-line purposes and directory names: words in at least two of its experiments and rare in other clusters.

Classifiers. Each topic gets a logistic regression on Ettin-17m post embeddings, refitted every run. Positive examples are posts from the last 14 days that contain one of the topic's terms (Bluesky post search; at most 1,500, at most 25 per author; terms with fewer than 15 hits are dropped). Negative examples are about 1,200 posts from Bluesky's Discover feed plus up to 150 search hits from each other topic (minus any that contain this topic's own terms); with Discover alone, anyone who writes in sentences scored as on-topic. A topic with fewer than 80 positive posts or a held-out AUC below 0.80 sits out that run.

Scores. An account's share on a topic is the mean classifier probability over its last 100 posts. Shares are normalised per topic as (share − mean on held-out Discover posts) / (mean on held-out term posts − mean on Discover), so 0 reads like Discover and 1 like the topic's search hits. An account's topic is the one with the highest normalised score, called its norm below.

Topic members. Norm at least 0.45. Topics take turns, least represented first, and no topic takes more than ceil(1.5 × topic slots / number of topics) members, where topic slots are two thirds of the 120-member target; this keeps AI and embeddings from taking over the list.

Exploration members. Norm below 0.30 on every topic, drawn from Oskar's follows, the authors of his last 1,500 likes and 500 reposts (read from his public repository), and accounts followed by three or more current members. They are ranked by a sum: 3 for an Oskar follow, up to 5 for likes, up to 6 for reposts, up to 5 for member follows.

Screen and judge. A candidate needs at least 10 posts, at most 40 a day, a post in the last 30 days and no bot or feed words in handle, name or bio. Claude Haiku then judges batches of 25 from bio, posting stats and six sample posts. Topic candidates must be individuals writing substantively on their topic; exploration candidates must be individuals writing thoughtfully in their own voice. Organisations, feeds, marketers and AI personas are rejected either way. A batch whose reply cannot be parsed twice is left unjudged, and those accounts come up again next run. Rejected accounts are skipped for 90 days, removed members for 120.

Weekly rules. A refresh drops a member with no post in 60 days, a topic member whose norm is below 0.30 two weeks running, and an exploration member at least 28 days on the list with no insight saved in the last 28 days, at most 15 removals a week. It then adds up to 10, about a third exploration, never above 150 members.

The thresholds are DEFAULTS in curate.py: target 120, max_size 150, explore_frac 1/3, topic_add 0.45, topic_drop 0.30, explore_max 0.30, weeks_below_drop 2, silent_days 60, explore_grace_days 28, insight_window_days 28, max_add 10, max_remove 15. set-param overrides one in the state.

The classifier recipe follows List Scout (oaustegard.github.io bsky/list-scout/topic.js); the screen and judge follow oaustegard/experiments bsky-list-ai-share/vet.py.

Judge fallback

Without the claude CLI, build and refresh exit 3, write nothing, and save the candidates (handle, bio, stats, sample posts, topic or exploration) to pending-judge-<date>.json beside --out. Judge each one another way, for instance with Muninn's decide tool and one yes/no question per account, write {did: {"keep": 0 or 1, "why": "…"}} to a file, and rerun with --verdicts FILE. Accounts missing from the file are left unjudged.

Failure signals

signal meaning
topics_skipped lists most topics post search failed: check the MUNINN_BSKY_* sign-in and egress to Muninn's account host
an interest topic skipped for AUC its terms are too general; edit assets/topics.json
state_save_error Turso refused the write and the state went to muninn-bsky-list-state.backup.json beside --out. Run restore-state on it before the next refresh
a refresh removing many members this week's classifiers misfit; the cap stops it at 15. Check topics_fitted AUCs before set-param topic_drop
members_silent_this_window near the member count in read Bluesky's list feed lagged or failed; rerun read

Verification

From the skill directory: python3 -m pytest -q tests (offline). After a live run, status prints the list URL and the member split.