Mbox Knowledge Pack
Use this skill when the user has a local .mbox export and wants to turn it into a structured, private knowledge asset. The default outcome is an analysis pack, not a public archive or live-mail action.
Boundaries
- Work locally from files the user provides or points to.
- Never mutate the original mbox.
- Do not upload, send, or publish email content.
- Do not fetch links by default; record URL provenance unless the user asks for link checking.
- Do not create final wiki pages directly for a Claude/Karpathy wiki unless the user explicitly asks. Prefer raw compile packets.
- Treat fine-tuning as an optional advanced preparation mode, never the default.
Core Workflow
- Inspect the mbox and output location.
- Parse messages safely: MIME decoding, dates, subjects, message IDs, thread IDs when available, labels, plain text bodies, URLs, and attachment metadata.
- Build a local analysis pack:
messages.jsonlthreads.jsonllinks.jsonlattachments/manifest.jsonlanalysis.sqlitemanifest.jsonsummary.json
- Produce reports: corpus profile, timeline, theme map, source/domain signals, and import/defer/exclude recommendations.
- If requested, curate strongest threads into a smaller first batch with evidence message IDs and rationale.
- If a wiki target is provided, prepare raw source packets that match that wiki's compile conventions.
- If writing output is requested, create book/article concept notes and outlines grounded in selected threads.
- If fine-tuning prep is requested, create a separate redacted dataset candidate pack and audit report.
Helper Script
Use the bundled helper when a deterministic local pack is useful:
python3 skills/mbox-knowledge-pack/scripts/mbox_knowledge_pack.py --mbox /absolute/path/export.mbox --out /absolute/path/analysis-pack
Optional modes:
python3 skills/mbox-knowledge-pack/scripts/mbox_knowledge_pack.py --mbox /absolute/path/export.mbox --out /absolute/path/analysis-pack --fine-tune-prep
The script is intentionally conservative. It extracts and scores; it does not claim semantic understanding equivalent to a manual review.
Wiki Import
Read references/karpathy-wiki-import.md when the target is a Claude/Karpathy-style markdown wiki, especially one with CLAUDE.md, CODEX.md, or WIKI-JOURNEY.md conventions.
Important default: generate grouped raw packets under a target such as raw/memories/<collection>/; let that wiki's /compile and /lint commands create final articles and backlinks.
Fine-Tuning Prep
Read references/fine-tuning-prep.md before creating model-training files. Email archives are usually noisy, private, and copyright-mixed. Prepare fine-tuning datasets only from explicitly selected/approved material, with redaction, split separation, and an audit report.
Privacy And Verification
Read references/privacy-and-verification.md when creating shareable exports, wiki packs, or fine-tune datasets. Preserve provenance while removing private addresses, raw recipient lists, local absolute paths, and sensitive context from public-facing outputs.
Reporting Rules
- State which outputs are deterministic extraction vs heuristic analysis vs human-curated recommendation.
- Include counts and source hashes in manifests.
- Keep evidence IDs so recommendations can be traced back to source messages.
- Distinguish
import, defer, and exclude; never imply every email deserves promotion.
- For fine-tuning, distinguish RAG/search suitability from supervised-training readiness.
1---2name: mbox-knowledge-pack3description: Convert local .mbox email exports into local-first knowledge packs for analysis, curation, wiki compile staging, writing projects, and optional fine-tuning dataset preparation. Use when asked to analyze an mbox, summarize old emails, curate email threads, prepare a Karpathy/Claude-style wiki raw import, or derive supervised examples from an email corpus. Do not use for live mailbox operations.4---56# Mbox Knowledge Pack78Use this skill when the user has a local `.mbox` export and wants to turn it into a structured, private knowledge asset. The default outcome is an analysis pack, not a public archive or live-mail action.910## Boundaries1112- Work locally from files the user provides or points to.13- Never mutate the original mbox.14- Do not upload, send, or publish email content.15- Do not fetch links by default; record URL provenance unless the user asks for link checking.16- Do not create final wiki pages directly for a Claude/Karpathy wiki unless the user explicitly asks. Prefer raw compile packets.17- Treat fine-tuning as an optional advanced preparation mode, never the default.1819## Core Workflow20211. Inspect the mbox and output location.222. Parse messages safely: MIME decoding, dates, subjects, message IDs, thread IDs when available, labels, plain text bodies, URLs, and attachment metadata.233. Build a local analysis pack:24 - `messages.jsonl`25 - `threads.jsonl`26 - `links.jsonl`27 - `attachments/manifest.jsonl`28 - `analysis.sqlite`29 - `manifest.json`30 - `summary.json`314. Produce reports: corpus profile, timeline, theme map, source/domain signals, and import/defer/exclude recommendations.325. If requested, curate strongest threads into a smaller first batch with evidence message IDs and rationale.336. If a wiki target is provided, prepare raw source packets that match that wiki's compile conventions.347. If writing output is requested, create book/article concept notes and outlines grounded in selected threads.358. If fine-tuning prep is requested, create a separate redacted dataset candidate pack and audit report.3637## Helper Script3839Use the bundled helper when a deterministic local pack is useful:4041```bash42python3 skills/mbox-knowledge-pack/scripts/mbox_knowledge_pack.py --mbox /absolute/path/export.mbox --out /absolute/path/analysis-pack43```4445Optional modes:4647```bash48python3 skills/mbox-knowledge-pack/scripts/mbox_knowledge_pack.py --mbox /absolute/path/export.mbox --out /absolute/path/analysis-pack --fine-tune-prep49```5051The script is intentionally conservative. It extracts and scores; it does not claim semantic understanding equivalent to a manual review.5253## Wiki Import5455Read `references/karpathy-wiki-import.md` when the target is a Claude/Karpathy-style markdown wiki, especially one with `CLAUDE.md`, `CODEX.md`, or `WIKI-JOURNEY.md` conventions.5657Important default: generate grouped raw packets under a target such as `raw/memories/<collection>/`; let that wiki's `/compile` and `/lint` commands create final articles and backlinks.5859## Fine-Tuning Prep6061Read `references/fine-tuning-prep.md` before creating model-training files. Email archives are usually noisy, private, and copyright-mixed. Prepare fine-tuning datasets only from explicitly selected/approved material, with redaction, split separation, and an audit report.6263## Privacy And Verification6465Read `references/privacy-and-verification.md` when creating shareable exports, wiki packs, or fine-tune datasets. Preserve provenance while removing private addresses, raw recipient lists, local absolute paths, and sensitive context from public-facing outputs.6667## Reporting Rules6869- State which outputs are deterministic extraction vs heuristic analysis vs human-curated recommendation.70- Include counts and source hashes in manifests.71- Keep evidence IDs so recommendations can be traced back to source messages.72- Distinguish `import`, `defer`, and `exclude`; never imply every email deserves promotion.73- For fine-tuning, distinguish RAG/search suitability from supervised-training readiness.
Laksh-star