/mind-clone

name: mind-clone
description: Build and talk to an AI "mind clone" of a real public author — a model of how they think, built from their public content, that reasons in the author's own frameworks and cites them with deep-links to the exact source moment. Use when the user wants to clone someone's mind/thinking from their public content (articles, YouTube, podcasts, talks, interviews), get advice "in the thinking of" a specific person, extract that person's step-by-step method or playbook, or predict what a named person would think about a topic. Triggers on "clone <person>", "mind-clone <person>", "build a clone of", "think like <person>", "what would <person> say/think about", "turn <person> into an AI", "create a digital clone of someone's mind". Self-contained: relies only on built-in tools plus optional free CLIs (yt-dlp, whisper).

mind-clone — a clone of how an author thinks, from public content

This skill builds a **cognitive model** ("mind clone") of a real author from their public footprint, persists everything as a local database in the user's project, and then answers questions, gives advice, and predicts positions **the way that author would think** — not by imitating their tone, but by reasoning through their actual mental models.

The clone is an **interpretive model built from public sources**, not the real person. It must always label predictions as predictions and never fabricate quotes. See `reference/ethics.md`.

---

Two modes — route first

On invocation, decide which mode applies:

| Signal | Mode | Go to | |---|---|---| | "Build / create a clone of X", no clone exists yet for X | **BUILD** | Pipeline below | | A clone folder already exists and the user asks a question, wants advice, or "what would X think about…" | **CHAT** | `reference/04-clone-runtime.md` | | Unclear | Ask: build a new clone, or talk to an existing one? List clones found in `./clones/`. |

Check `./clones/` (in the current working directory) for existing clones before deciding.

---

Step 0 — Environment manifest (read before anything, both modes)

Before discovery, harvesting, or chatting, **read `ENVIRONMENT.md`** in this skill's directory — the per-machine manifest of where tools live, the OS/path layout, and the proven build/chat commands for *this* user. It exists so you don't re-probe the OS or hunt for libraries every session.

• **Missing `ENVIRONMENT.md`?** (first run on a new machine) — run `python scripts/setup.py`: it creates the manifest from `ENVIRONMENT.template.md`, auto-detects tools/paths, and prints the welcome message to show the user. Fill in the prose sections as you learn the machine.
• **Stale after installing/removing a tool?** — re-run `python scripts/setup.py` (or `bash scripts/probe_env.sh --write`).
• `ENVIRONMENT.md` is git-ignored, so it stays local and survives skill updates.
• Append every environment trap you hit to its "Known quirks" section, so the next session skips the dead-end.

---

Step 1 — Language protocol (always first, both modes)

1. On the **first interaction of a session**, ask the user which language they want to converse in (offer the language their message is in + English). 2. Store the choice in the clone's `config.json` (`"chat_language"`). On later sessions, read it and skip the question. 3. **Always converse in the user's chosen language.** 4. **All persisted artifacts** (the database, cognitive model, source tables, this skill's files) are written in **English** regardless of chat language — so the clone is portable and shareable. Translate to the user's language only in the live conversation. 5. **In CHAT mode, the author's quotes display in the user's language too** — lead with a faithful translation, but **always keep the verbatim original alongside** (the deep-link verifies the original). Never present a translation alone as a verbatim quote. See `reference/language.md` → "Quoting the author in CHAT mode".

Full rules: `reference/language.md`.

---

BUILD pipeline (5 phases)

Run phases in order. Each writes to `./clones/<author-slug>/`. Load the linked reference at the start of each phase. Tell the user what you're doing between phases; don't run silently for minutes.

Phase 1 — DISCOVERY (exhaustive search)

Load **`reference/01-discovery.md`**. Enumerate name variants and affiliations, then run a multi-modal web sweep (articles, YouTube, podcasts, radio, talks, interviews, social, papers). Goal: find *everything* public, not the top 10 links. Deduplicate into a source registry. → emits `meta` rows (a transient delta) → `loop.py <slug> import <delta.jsonl>` (validated on write → log)

Phase 2 — SOURCE TABLE (review gate)

Present the user a table of every source found: **Date · Type · Source · Summary · URL** via `loop.py <slug> render sources`. This is the review checkpoint — the user can prune, add, or approve before the expensive harvest. → `render sources` (JIT view from `meta`; nothing stored)

Phase 3 — HARVEST (extract-on-the-fly)

Load **`reference/02-harvest.md`**. Default is **fast**: read each source once and extract the distillate as **EP points** (a transient delta `.jsonl`, thrown away after `import`) — dated, attributed, grounded quotes (`text`) + their group/vector/relation — **and save the raw transcript/text to `raw/` by default** (text is cheap — tens of MB even for a big channel — and it powers deep-links + CHAT-mode live-grounding fallback). **Capture timestamps** (`t_start` in seconds + a ready `deeplink`) for every audio/video quote so the clone can later cite the author verbatim and link to the exact minute; text quotes store the plain URL. Use existing text/captions/transcripts only (fast); **no slow ASR by default**. Articles via WebFetch, YouTube via yt-dlp captions.

• Sources with **no subtitles/transcript** → collect them, **show the user a table** (Type · Title · Source · Duration · URL) and **ask** whether to run whisper (slow). Note that a few missing files rarely shift t