Discovery Vault Template

A starter Obsidian vault for AI-assisted discovery engagements. Built for the workflow where you turn a pile of recorded interviews into a structured, citation-enforced knowledge base that feeds client-facing deliverables.

What this is

A scaffold — folder structure, page templates, citation rules, and an LLM operating manual — with no engagement content. Clone it at the start of a new project, fill in the engagement context, drop in your transcripts, and let an LLM agent (Claude Code, Codex, etc.) maintain the wiki as you ingest sources.

Architecture

Three layers:

• transcripts/ — raw, immutable sources (recorded interviews and meetings). The LLM reads but never modifies them.
• wiki/ — synthesis layer. The LLM owns this entirely: stations, flows, pain points, opportunities, all cross-linked, all citing back to transcripts.
• reference/ — primary-source client artifacts (citable for structural facts) plus any prior-synthesis material (awareness only, never a starting skeleton).

The schema lives in CLAUDE.md and wiki/schema.md. The first is the operating manual; the second is the page templates.

How to use

1. Click "Use this template" on GitHub (or clone this repo) to spin up a new vault.
2. Open CLAUDE.md and fill in the Engagement context section at the top — client name, deliverables, scope guardrails. Leave the rest alone.
3. Open project-plan.md and replace the placeholders with the engagement specifics — phases, ingest order, exit criteria.
4. Drop your transcripts into transcripts/ (kebab-case YYYY-MM-DD - Session.md format) and update transcripts/index.md.
5. Drop any primary-source client artifacts into reference/ along with a markdown companion note describing what they are and how they can be cited.
6. Open the vault in Obsidian. Open Claude Code (or your LLM agent of choice) in the same folder.
7. Start ingesting transcripts one at a time. The agent will populate the wiki.

What's in the box

.
├── CLAUDE.md              ← LLM operating manual (engagement-agnostic + placeholders)
├── README.md              ← this file
├── project-plan.md        ← phase plan skeleton, fill in per engagement
├── .obsidian/             ← Obsidian config (light defaults)
├── .claude/               ← Claude Code config
├── reference/             ← primary-source artifacts + companions (empty)
├── transcripts/
│   ├── index.md           ← catalog of source transcripts
│   └── _speaker-remaps/   ← scratch files for noisy-transcript remaps
└── wiki/
    ├── schema.md          ← four page templates: station, flow, pain-point, opportunity
    ├── index.md           ← catalog of every wiki page (skeleton)
    ├── log.md             ← chronological ingest log (skeleton)
    ├── glossary.md        ← jargon glossary (skeleton)
    ├── stations/                     ← function/team pages (empty — emerge from transcripts)
    ├── operating-flows/              ← end-to-end process flows (empty; folder name disambiguates from product-UX flows)
    ├── cross-cutting-pain-points/    ← multi-station pain pages (empty; folder name self-documents the rule)
    └── opportunities/                ← AI/automation opportunity pages (empty)

Why this exists

Most LLM workflows are RAG: upload documents, retrieve chunks at query time, generate an answer. The LLM rediscovers the knowledge from scratch every time. Nothing accumulates.

The pattern this template implements is different — and the cleanest articulation of it is in Karpathy's LLM Wiki gist. The LLM incrementally builds a persistent, structured wiki between you and the raw sources. Every ingest updates entity pages, revises summaries, flags contradictions. The synthesis is compiled once and kept current — it doesn't get re-derived on every query. By the end of an engagement you have a living knowledge graph with receipts on every claim, ready to drive any client-facing deliverable.

This template is a domain-specific instantiation of that pattern, hardened for client-facing discovery work: hard citation rules (speaker + timestamp on every claim), four page templates (station / flow / pain-point / opportunity), immutable sources, a hybrid citation policy that distinguishes primary-source artifacts from prior-LLM guesses, and an operating manual (CLAUDE.md) the LLM actually follows.

Inspirations

• The core pattern — an LLM-maintained, persistent wiki sitting between you and the raw sources, where the LLM owns the synthesis layer and humans curate sources and ask questions — is directly inspired by Andrej Karpathy's LLM Wiki gist (karpathy/442a6bf555914893e9891c11519de94f). The three-layer architecture (raw / wiki / schema), the index.md + log.md convention, the "wiki is a compounding artifact" framing, and the "Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase" mental model all come from there. Read the gist before using this template — it explains the why behind the structure.
• The four-page-template shape (station / flow / pain-point / opportunity) is specific to operational discovery work — kitchens-of-functions, process flows, friction, AI opportunities. This is a domain-specific instantiation on top of the general pattern.
• The Option C citation policy (primary artifacts citable for structural facts only, transcripts required for interpretive claims) and the strict "every claim cites a speaker + timestamp" rule are hard-won discipline from real client engagements where mixing source types caused trouble.
• The transcript noisy-attribution patterns and remap workflow (room-device-name collapse, single-label merged speakers, per-utterance attribution by content cues) come from running real Fireflies transcripts through this pipeline.

License

Private template. Not for distribution.