skills/sweep-stale-seeds/SKILL.md
--- name: sweep-stale-seeds description: Identifies substacker seeds older than 30 days with status=seed and no incoming related_seeds links, flags them for writer review, and recommends keep / promote-to-draft / kill based on density score. Does NOT auto-execute any action. Emits a review list to ops/librarian/YYYY-MM-DD-stale-sweep.md. Run at session start after ingest, once per day max. Trigger keywords: stale, sweep, review, old seeds, cleanup, gardener, corpus hygiene. --- # Sweep Stale Se
npx skillsauth add lyndonkl/claude skills/sweep-stale-seedsInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
Related skills: Called by Librarian agent's pipeline step 2. Output consumed by the writer and by curator on its monthly-ish cycle.
Sweep the corpus for stale seeds:
- [ ] Step 1: If ops/librarian/{today}-stale-sweep.md exists, skip (daily idempotency)
- [ ] Step 2: Glob corpus/seeds/*.md
- [ ] Step 3: For each seed, parse frontmatter
- [ ] Step 4: Apply stale criteria: status=seed AND created < today-30d AND no related_seeds referencing this seed
- [ ] Step 5: For each stale seed, compute recommendation
- [ ] Step 6: Write the review list to ops/librarian/{today}-stale-sweep.md
A seed is stale if ALL of:
status: seed (not yet promoted to draft)created is > 30 days before todayrelated_seeds includes this seed's id (orphan test)manual_edits: false (writer-edited seeds are never in the sweep)| If... | Recommend |
|---|---|
| density >= 7 | promote-to-draft (high-quality material sitting stale is the real loss) |
| density <= 3 | kill (low-density AND stale = not going to improve) |
| else | keep (mid-density — more time to mature) |
ops/librarian/YYYY-MM-DD-stale-sweep.md:
---
agent: librarian
date: YYYY-MM-DD
total_seeds: N
stale_seeds: M
recommendations:
promote: X
kill: Y
keep: Z
---
# Stale Seed Sweep — YYYY-MM-DD
## Promote to draft (X)
- `{seed-id}` | density={N} | created={date} | topics={comma-list}
- Rationale: high density, sitting stale. Consider promoting.
## Kill (Y)
- `{seed-id}` | density={N} | created={date} | topics={comma-list}
- Rationale: low density, stale, orphan. Safe to move to corpus/dead/.
## Keep (Z)
- `{seed-id}` | density={N} | created={date} | topics={comma-list}
- Rationale: mid-density, give it more time.
Corpus today (2026-04-23) has 47 seeds. Globbing + filtering finds 6 stale:
## Promote to draft (1)
- 2026-02-18-residuals-as-a-reset-button | density=8 | created=2026-02-18 | topics=resnet, training
- Rationale: high density, sitting stale for 2 months. Consider promoting.
## Kill (2)
- 2026-01-04-maybe-writing-about-tokenizers | density=2 | created=2026-01-04 | topics=tokenizer
- Rationale: low density, stale, orphan. Safe to move to corpus/dead/.
- 2025-12-21-quick-thought-on-sparse-moe | density=3 | created=2025-12-21 | topics=moe
- Rationale: low density, stale, orphan. Safe to kill.
## Keep (3)
- 2026-02-28-rope-intuition | density=5 | created=2026-02-28 | topics=attention-mechanism, rope
- 2026-03-05-grokking | density=5 | created=2026-03-05 | topics=emergence, training
- 2026-03-15-temperature-vs-top-p | density=6 | created=2026-03-15 | topics=sampling
- Rationale: mid-density; not stale enough to act yet.
status. The status field is owned by the writer / downstream agents.manual_edits: true in the kill list regardless of density.related_seeds is never stale by this skill's criteria, even if old and low-density.kill for a seed with density >= 7; that would contradict the promote recommendation.testing
Cluster a conference's event records into a small set of coarse themes with finer sub-clusters, an explicit outlier bucket, and soft (multi-membership) affinities — using the hybrid embed-then-label pipeline (embed abstracts, reduce, density-cluster, then LLM-label the clusters) when embedding libraries are available, and an LLM-reasoned hierarchical fallback when they are not. Embeddings do the grouping; the LLM only names the groups. Conference-agnostic. Use when turning structured event records into a navigable theme map for preference elicitation and scheduling, when you need 6-8 reasonable themes rather than 20 muddy ones, or when overlapping talks must belong to more than one theme. Trigger keywords - theme clustering, cluster talks, embed then label, soft membership, outlier talks, conference themes, topic map.
development
Build a personal conference schedule as a constraint-optimization problem — hard constraints (no time overlap, room-to-room travel time, capacity/registration, the attendee's own must-attends and blackouts) plus a user-owned weighted objective trading interest against breadth, pacing (maximize contiguous free time), and serendipity. Surfaces unbreakable conflicts (two high-value overlapping talks the model cannot rank) as decisions for the human rather than silently picking, and reports what each choice traded away. Conference-agnostic. Use to turn a preference profile plus a theme map into a day-by-day plan, to resolve overlapping sessions, or to balance a packed vs paced schedule. Trigger keywords - schedule optimization, conference schedule, constraint optimization, overlapping talks, contiguous free time, conflict surfacing, packed vs paced.
development
Parse a heterogeneous conference program (markdown, HTML, PDF-derived text, or JSON) into normalized event records with per-field confidence scores and independent classification axes (topic, depth, format, prerequisites, recorded, capacity). Detects the program's format before extracting, treats every inferred field as uncertain (present vs inferred vs missing), and flags thin or missing abstracts so downstream enrichment can target them. Conference-agnostic. Use when ingesting a conference or event schedule into a structured store, normalizing a talk/session list, or extracting per-session metadata with calibrated confidence. Trigger keywords - program ingestion, parse schedule, session extraction, event records, conference program, talk metadata, per-field confidence.
development
Build a personalized preference profile from a small number of well-chosen, cluster-grounded questions instead of a long survey. Represents the person's interests as an uncertainty region over the theme map, picks the single highest-information-gain choice-based question (contrasting real talks from different clusters), balances exploiting known interests against exploring uncertain ones, deliberately injects outlier probes to fight selection bias, and stops as soon as the schedule would be stable. Also elicits the user-owned objective weights and hard constraints. Interactive — runs where it can actually ask the person. Conference-agnostic. Use to turn a theme map into a preference profile, to decide what to ask a conference attendee, or to elicit scheduling priorities. Trigger keywords - preference elicitation, ask few questions, information gain, choice-based questions, selection bias probe, objective weights, attendee preferences.