seed-huntlisted

End-of-project seed hunt for Kyle's reproduce-and-measure research projects (the forge-gap / decay-pin lineage). Verifies the just-finished repo is truly closed, harvests its hard-won lessons into the living selection bar (references/selection-bar.md), sweeps arXiv for candidate papers, scores a 3–5 item shortlist against the bar, presents a decision brief for Kyle's pick, and closes with a /handoff prompt routed toward /kickoff. Builds NOTHING. Use whenever Kyle types /seed-hunt, says "seed hunt", "find my next paper", "what should I reproduce next", "ready for the next project", or wraps up a research-reproduction project and asks what to build next — even if he doesn't name the skill.
ksdisch/claude-config · ★ 0 · AI & Automation · score 70

Install: claude install-skill ksdisch/claude-config

# Seed Hunt — find the paper the next project reproduces You are running the between-projects step in Kyle's research-reproduction pipeline. A project in this lineage takes one published paper (or a 2–3-paper set forming one thesis), re-implements a narrow primitive from it at hobby scale, and *measures* the paper's claim honestly under pre-committed statistical gates. The recipe's track record lives in `references/selection-bar.md`, alongside the selection bar itself. **You BUILD NOTHING in this skill.** The deliverables are: (1) an updated selection bar, (2) a vetted shortlist, (3) Kyle's sign-off on one pick, (4) a closing `/handoff` prompt for the planning session. Design, thesis, and stage-ladder belong to the NEXT session, routed through `/kickoff`. Kyle's global CLAUDE.md binds throughout: plain English, define every jargon term on first use, decision-brief format for real choices (2–3 options, trade-offs, one "(Recommended)"). Kyle decides; you recommend. ## Parse `$ARGUMENTS` - `--audio [short|long]` → when the Phase 3 decision brief is ready, also narrate it as an MP3 so Kyle can weigh the shortlist by ear before picking (see "Audio brief" in Phase 3). `short` (default) = each candidate in a sentence or two plus the recommendation and its why (~2 min); `long` = the full brief — per-candidate merits and trade-offs against the bar, plus the single-vs-set and reuse-vs-range calls (~4–5 min). Without `--audio`, ignore all audio steps — the skill behaves ex