Capture Conversation

Overview

This skill captures cognition, not just content. Conversations on the move are meandering, multi-threaded, full of dead ends — on purpose. The meandering is the raw material. A normal summarizer flattens everything into a tidy result and throws away the reasoning; this skill preserves the reasoning so future conversations can build on it and link to it.

Two artifacts per capture, always:

1. Source — the verbatim record. Sacred, never trimmed, kept forever.
2. Thinking note — the distilled layer: decisions, considered-and-rejected threads, open questions, actions. This is the linkable, wiki-facing artifact.

The summary inside the thinking note is a convenience layer. It serves the thinking, it never replaces it.

The target format is llm-wiki style ingestion (after Karpathy's llm-wiki idea): raw sources preserved forever, with a distilled, densely-wikilinked layer on top — and it's the wiki layer, not the source, that the rest of the vault links against. Where those files finally live is the user's decision, never this skill's (see Artifacts & Naming).

Assumed environment: an Obsidian-style knowledge vault — plain Markdown files, [[wikilinks]], YAML frontmatter. Obsidian and VS Code Foam both qualify, as does any similar system; features specific to one tool (e.g. Obsidian Bases) are marked as such.

Origin: this skill was specced inside the first conversation it ever ingested (2026-08-04, "obsidian voice capture and privacy"). v0.

When to Use

• User shares a claude.ai conversation link/export and wants it "in the vault"
• A voice memo or recorder file (audio + transcript) needs ingesting
• A meeting or conference talk recording needs to become notes
• User says "process this conversation", "ingest this transcript", "capture this"

Input Shapes × Purposes

Detect the shape from the input itself — don't make the user route it.

Edge case (ingestion #2): a meeting where the user is an equal co-decider, not a passive audience. A planning call where both sides jointly negotiate and land on decisions together is still tagged Purpose: external-material in frontmatter (meetings are external-material by this table's own definition, and the frontmatter drives future Base filtering) — but forcing the distillation into "What was presented" / "My takeaways" would misrepresent it as one-directional. When the content is genuinely joint, use the own-cognition template's structure (Decisions reached / Considered and rejected) under the external-material frontmatter, and say so in one line so a future reader knows the mismatch was deliberate, not an oversight.

Audio-first captures (voice recorder): the audio file is the canonical original, the transcript is derived. Preserve both, linked. Transcribe locally if needed.

Artifacts & Naming

The captures folder is per-vault — never assume it; where files finally go is always the user's call. Resolve it in this order:

1. The target vault's/folder's CLAUDE.md (or equivalent conventions doc) declares where conversation captures live — use that.
2. Not declared? Ask the user where captures should go and use the answer for this run. Propose the CLAUDE.md line in the step-9 report, and write it only after the user says yes, so future runs don't ask.

No assumptions beyond that — folder names, organizational philosophy, and note style all belong to the target vault, not this skill.

Confirmed alternative: a topic-scoped verbatim appendix, no separate pair. Not every piece of source material earns the full captures-folder pipeline. When material is narrowly about one existing topic note — not a broader, multi-topic conversation — embedding it directly inside that topic note as a # <description> (verbatim) section is a valid lighter-weight version of the same discipline: a one-line provenance note, the verbatim text (typos and all, never silently cleaned), findings distilled into the note's living sections above it, action items still copied to the TODO surface, the day's daily note still updated. Seen twice now on the same vault's DSCI 521 note — a co-instructor's proposal doc, then (independently, without this skill being invoked) a Slack thread — both done correctly by the vault owner working solo, matching every rule above without prompting. Recognize this as a legitimate pattern, not a deviation to "fix" back into a captures-folder pair — the two-artifact pipeline is for material that stands on its own or spans multiple topics; a single-topic excerpt can just live where it's used.

Within the captures folder, nest by year/month (<captures folder>/YYYY/MM/) unless the vault says otherwise — every capture is 2+ files, so a flat folder grows fast.

Naming: YYYY-MM-DD-HHmm-<slug> — ISO date, time to the minute (conversation start time), then a short human slug, lowercase, dash-separated. One session split into two topics = same timestamp, two slugs.

<captures folder>/2026/08/
  2026-08-04-1755-obsidian-voice-capture-and-privacy.md             ← thinking note
  2026-08-04-1755-obsidian-voice-capture-and-privacy-transcript.md  ← verbatim source
  2026-08-04-1755-obsidian-voice-capture-and-privacy-audio.mp3      ← only for audio-first captures

The -transcript / -audio suffixes keep the pair adjacent in file lists and give Obsidian Bases a metadata-free filter (!file.basename.endsWith("-transcript")).

Frontmatter Schemas

Match the target vault's frontmatter key style (the examples below use capitalized keys). Created is the conversation's start time, not the file-write time — the conversation is the knowledge event.

Thinking note:

---
Created: 2026-08-04 17:55
Modification date: <write time>
Tags: #conversation <topic tags, e.g. #homelab #voice_capture #privacy>
Type: dialogue            # monologue | dialogue | multi-party
Purpose: own-cognition    # own-cognition | external-material
Source: <URL or origin description>
Transcript: "[[<basename>-transcript]]"
---

Transcript (source):

---
Created: 2026-08-04 17:55
Modification date: <write time>
Tags: #conversation_transcript
Type: dialogue
Purpose: own-cognition
Source: <URL or origin description>
Thinking note: "[[<basename>]]"
Audio: "[[<basename>-audio.mp3]]"   # audio-first captures only
---

Keep keys consistent — they are the future Obsidian Base's columns.

Verbatim, Defined

• Every turn, in order, speaker-labeled (**Daniel:** / **Claude:** / diarized names), timestamps kept when available.
• Voice-transcription garbles stay as spoken/captured ("Obama model", "Hartsping") — decode them in the thinking note, never edit the source.
• Strip only delivery chrome: web-page UI text, duplicated message previews, "Report" buttons. The words exchanged stay 1:1.
• Nothing else is trimmed. Dead ends, cut-offs, garbled fragments — all kept.
• Voice-mode assistant turns can be scrambled in the stored source itself (ingestion #3: interleaved word fragments like "That makot of thees sense — … a l water's probably just running out"), not just in the rendered page. Same rule: keep it exactly as stored, decode it in the thinking note, and say in the provenance note that the scramble is in the source, not a rendering artifact.

Steps

1. Obtain the full text. Every run: clean mechanically with a small script, never by retyping — on long transcripts, regeneration by hand drifts. Garbles stay verbatim in the source and are decoded in the thinking note. Pick the route by input:
   • (a) The user's own claude.ai chat → the conversation JSON from the claude.ai org API (unofficial, may break). See references/claude-chat-sources.md.
   • (b) Someone else's share link → page text via a browser tool, with turn-count reconciliation. Same file.
   • (c) Audio → transcribe locally and keep the audio file. See references/audio-transcription.md.
2. Identify shape and purpose from the content. Note conversation date and start time (in-conversation timestamps beat file dates). For the user's own claude.ai chats, the first created_at in the conversation JSON is the start time (UTC — convert); for someone else's share link, use the page's timestamps. Confirm the date with the user if ambiguous.
3. Privacy gate. Scan for sensitive or sensitive-adjacent passages (private topics, other people's information, work-policy matters). Ask the user: full verbatim / flagged redactions (marked [redacted: reason]) / excluded entirely. Never write into a vault's excluded folders (e.g. do-not-publish/).
4. Resolve the captures folder (per Artifacts & Naming — vault CLAUDE.md first, ask if undeclared), draft the slug, and build the file pair from the skill's templates (template-transcript.md, template-thinking-note.md).
5. Write the source file: provenance line + verbatim dialogue per the rules above.
6. Distill the thinking note: summary (3–5 sentences) · decisions reached · considered-and-rejected (each with why it was closed) · open threads · action items · transcription decoder table. For external-material captures, reframe as faithful record + user's takeaways.
7. Linking pass. Search the vault (filenames and content) for notes touching the conversation's topics; weave wikilinks into the thinking note inline and under # References. This is what makes captures compound over time.
8. Vault integration. Follow the target vault's conventions doc for bookkeeping, and touch existing topic notes only where the conversation genuinely settled or added something (one-line links, not rewrites) — unless the topic note explicitly pre-declared this conversation as its trigger for a real rewrite (ingestion #2: a note's provisional section had said in its own text "provisional until the meeting — after it, revise this section in place"). When a note has already marked itself as waiting on exactly this conversation, honor that instruction literally: a full in-place revision of the marked section, not a one-line link — while still leaving the note's historical sections (decision logs, prior drafts) untouched below, and flagging anything the conversation didn't resolve just as clearly as what it did (new complications count as findings too, not just clean resolutions). If the vault keeps daily notes, link the capture in the daily note for the conversation's date (create it from the vault's daily template if missing) — that's what makes captures findable in calendar views. When ingesting on a later day, do both: capture link in the conversation-date note, standard new-files entry in the ingestion-date note. If the vault declares a TODO/backlog surface, copy the action items there with a backlink to the thinking note — the thinking note's ## Action items stays as the historical record; the backlog copy is the live one.
9. Maintain the system. Ask: did this run reveal a convention gap, a schema tweak, a new input shape? If yes, propose, then apply: show the user the proposed change to the vault's CLAUDE.md or to this skill, and apply it only after the user says yes. A new lesson goes into the relevant references/ file as a short general rule, not into Steps as a dated story. The skill is expected to evolve — that's why it lives in a git repo.

Sensitive Information

The privacy gate (step 3) is mandatory, not optional. Architecture discussion that mentions private things can be fine; actual private content is not. When in doubt, ask before writing — and respect any hard exclusion rules the vault declares.

Future: Obsidian Base

(Obsidian-specific.) Not built yet — but the frontmatter above is designed for it. The draft Conversations.base, candidate future keys, and how captures stay out of a vault's other Bases: references/obsidian-base.md.

Open Threads (v0)

• Quick-capture vs deep-distill modes — v0 always does the deep pass; a fast-and-rough mode is a deliberate future split (chosen by the user, not auto-detected).
• Diarization remains unsolved — status and the content-based fallback are in references/audio-transcription.md.
• There is no official API for claude.ai chat history on Pro/Max plans. The org endpoint used for the user's own chats is unofficial and may break; otherwise it's manual export/share links — revisit if Anthropic ships connectors/API for chat history.