菜单
Load context for pipeline, cron, Lambda, OCR, and translation work. Use when starting any pipeline monitoring, debugging, or processing task.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Run OCR and translation quality evaluations across scripts and languages. Produces research-grade reports with MCR, cross-model agreement, embedding-space hallucination detection, and corpus readiness scores.
Run a Glass Bead Game session — a parlor game played against the Source Library corpus. Players connect verified passages from different books, times, and traditions into a growing string of 'beads,' with Claude as librarian, verifier, and scribe. Trigger when the user says 'glass bead game,' 'let's play beads,' or asks to start/resume/publish a game session. Claude runs the search-verify-display loop; humans do the noticing, choosing, and judging.
Stage-2 verification of first-translation flips using INDEPENDENT Claude subagents (subscription, not API key) with full evidence capture. Use to verify proposed badge changes (demotes/promotes) before writing any public claim — the binding step that catches AI-fabricated priors. Each subagent reports its actual search queries + sources, persisted to the first_translation_attempts provenance log.
Triage and respond to the user-feedback queue — pull open feedback, detect items already fixed but never marked, route translation-requests and metadata-corrections, mark items resolved, and (on approval) email submitters. Use when asked to check/triage/answer/clear feedback, "what are people saying?", or "respond to feedback."
Sweep Claude's private per-machine memory for team-relevant operational knowledge and propose moving it into the shared repo (or the private ops repo) as draft PRs. Use to promote lessons that accreted privately, or run on a schedule. Never auto-merges; never leaks.
One-shot audience & usage report — signups, DAU/MAU, time on-site, traffic, top content, search behavior, referrers, countries, and user journeys. Use when asked "what's our usage?", "how many signups?", "time on site?", "traffic?", or any audience/engagement metrics question.
| name | pipeline-context |
| description | Load context for pipeline, cron, Lambda, OCR, and translation work. Use when starting any pipeline monitoring, debugging, or processing task. |
Read these files before proceeding with pipeline work:
memory/pipeline-ops.md — Emergency controls, worker architecture, operational detailsmemory/lessons-learned.md — Operational postmortems and patterns.claude/docs/pipeline.md — Full processing pipeline (states, crons, prompts, costs).claude/docs/worker-architecture.md — Lambda worker details.claude/docs/page-lifecycle.md — Page processing statesgetModelForBook(book) from src/lib/types/ai-models.ts over hardcoding. It routes BPH books and non-Latin-script languages to gemini-3-flash-preview (full quality) and everything else to gemini-3.1-flash-lite-preview (50% cheaper, comparable quality on Latin-script). Enrich-worker uses gemini-3.1-flash-lite-preview for all phases (summary+index, chapters, quality scoring, collection assignment). Never use anything below Gemini v3 — gemini-2.x is deprecated.ocr.data or translation.data MUST call createRevision() firstsystem_config._id: 'processing_control', set paused: trueAll AI calls logged to gemini_usage collection via logGeminiCall() in src/lib/gemini-logger.ts.
GET /api/books/[id]/history (assembles from 6 collections)GET /api/admin/processing-dashboard?provider=iasrc/lib/errors.ts → classifyError(error)cost_tracking collection is DEPRECATED — use gemini_usage for all cost queriesAfter reading the memory files above, flag anything that contradicts what you observe in the codebase:
.claude/docs/batch-processing.md.claude/docs/observability.md.claude/docs/first-translation-system.md