Build a Local AI-Ready Second Brain

2.1

The Raw Dump: Capture Without Organizing

A raw dump folder accepting meeting notes, emails, voice notes, documents, and quick ideas — The first build step is not a perfect filing system. It is one folder that makes capture easier than losing the source.

The Thursday client call now has privacy labels and an approval rule, but the material still has nowhere to go. This chapter creates the : one folder where sources land before you sort them, summarize them, or decide what they mean. You need a single place where useful material can arrive before you make any judgment about it. Every source that lands here is meant to be processed by Claude or Codex, which catalogs it, adds , indexes it, and files it into the right folder. You drop it in; the AI does the rest.

Capture should feel easier than postponing the thought. When a follow-up email lands between calls, save it with the date, the privacy label you assigned in the previous chapter, and one plain reason it might matter. The result is a raw entry you can review later. It carries no trust yet; it simply preserves the source.

Save Thursday's follow-up email before judging it

A raw source is evidence, not a conclusion. Saving a meeting does not mean every statement in it is true. Saving an email does not mean you accept the request. Saving a note does not mean the idea will survive review. The keeps the source available so you can decide what it means later, with the original right in front of you.

Comparison

Field	What to save	Why it helps later
Source	File, link, paste, recording, screenshot, note, or email thread	You can return to the original instead of trusting memory
Date	The day the source appeared or was saved	Recent is easier to review before it goes stale
People	Names involved, if they are safe to store	Links the source to meetings, commitments, and relationship history
Project or area	Best guess, or unknown	Review has a starting point without requiring a final filing decision
Reason saved	One sentence about why this might matter	Tells future-you why the source deserved attention
Privacy label	The from your privacy configuration	Keeps sensitive sources from traveling into unsafe outputs

A raw source entry opened to show the source, privacy label, and review question fields — A raw entry stays useful later when the source, privacy label, and review question are visible from the start.

Three trust states carry a source from evidence to action

The previous chapter introduced two trust states: raw source and . Now that you are capturing real material, a middle step becomes useful. A processed source has been cataloged with , indexed in the database, and placed in the right project folder by the AI. You have not reviewed or approved it yet. It carries more structure than raw evidence, but less trust than a reviewed conclusion.

One source moving from raw evidence to a processed source and a reviewed record — Trust builds when the raw source stays visible while only checked conclusions become reusable records.

Comparison

Trust state	What happens	What it produces
Raw source	You save the , email, or note into raw/ so it can be found again	Original evidence with date, people, project guess, and privacy label
Processed source	Claude catalogs , indexes the source in brain.db, and files it under organized/	A with YAML front matter, a database row, and a copy in the right project folder
	You verify the source before reusing any conclusion from it	A decision, task, prep brief, or you are willing to act on

The AI processes every source so you never organize manually

The is temporary by design. Every source that lands in raw/ is meant to move through the pipeline: cataloged, indexed, and filed by the AI. The previous chapter showed where records end up (projects, areas, resources, archives, reviews). This chapter builds the pipeline that carries them there. Three folders and one database handle the entire lifecycle.

Project structure

second-brain

9 folders, 3 files

rawzero-friction capture
- 2026-05-08-client-call-transcript.md
- 2026-05-10-followup-email.md
processedcataloged by AI
organizedfiled by project
- 01_Projects
- 02_Areas
- 03_Resources
- 04_Archives
- 05_Reviews
brain.db

raw/ is the zero-friction landing zone. Drop a file, paste a note, save a recording. No , no tagging, no decisions required. processed/ holds sources after the AI has cataloged them, with YAML front matter prepended and a row added to brain.db. organized/ contains the destination folders from the previous chapter; the AI copies each processed source into the right subfolder based on its project or area guess. brain.db is a SQLite index the AI creates and maintains. You never need to open it directly. It makes your searchable.

Process a raw source into your second brain

Paste one source. Claude catalogs, indexes, and files it.

I saved a new source into my second brain's raw/ folder. Source file: [file name or path in raw/, e.g. raw/2026-05-08-client-call-transcript.md] Second brain folder: [your second brain folder path, e.g. ~/Documents/second-brain] Read the source and do these five steps: 1. Catalog: identify source type, date, people involved, likely project or area, and privacy label. 2. Metadata: write YAML front matter with record_type, source_id, source_type, captured_at, people, project, privacy_label, trust_state (set to "processed"), review_question, and stale_condition. 3. Index: add a row to brain.db (create the sources table if it does not exist) with source_id, source_type, captured_at, project, privacy_label, trust_state, and file_path. 4. Copy to processed/: copy the file with metadata prepended into processed/. 5. Organize: copy the processed file into the right subfolder under organized/ (01_Projects, 02_Areas, 03_Resources, or 04_Archives). If unsure, leave it in processed/ and flag the uncertainty. Show me what you did at each step. Do not delete the raw/ original. Do not mark anything as trusted or create tasks until I approve.

After the prompt runs, the source exists in three places. The original stays in raw/ as untouched evidence. The processed/ copy carries YAML front matter with every field the AI identified: record type, source ID, date, people, project, privacy label, , review question, and stale condition. The organized/ copy sits in the right project subfolder, ready to surface when you search by project, person, or date. Meanwhile, brain.db holds a new row that makes the source findable without opening any folder.

Project structure

second-brain

10 folders, 5 files

rawzero-friction capture
- 2026-05-08-client-call-transcript.md
- 2026-05-10-followup-email.md
processedcataloged by AI
- 2026-05-08-client-call-transcript.md
organizedfiled by project
- 01_Projects
  - client-onboarding-pilot
    - 2026-05-08-client-call-transcript.md
- 02_Areas
- 03_Resources
- 04_Archives
- 05_Reviews
brain.db

brain.db has a record for this sourceYou do not need to read SQL. Ask Claude or Codex what is in brain.db when you want to inspect the record.

sql

SELECT source_id, source_type, project, trust_state, file_path
FROM sources
WHERE captured_at = '2026-05-08';

-- source_id:   2026-05-08-client-call-transcript
-- source_type: meeting_transcript
-- project:     client-onboarding-pilot
-- trust_state: processed
-- file_path:   organized/01_Projects/client-onboarding-pilot/2026-05-08-client-call-transcript.md

When easy capture creates an unmanageable backlog

A common early failure: the system is so easy to capture into that the fills faster than you review it. Five to ten unreviewed items per day can pile into an unusable backlog within two or three weeks. Once the dump grows beyond what you can check in a single sitting, you start ignoring it. The system quietly stops earning trust.

Treat the as a bounded inbox. During your , promote what deserves attention, park what might matter later, and discard what no longer needs your time. If the dump consistently overflows, narrow what you capture. Save only sources tied to an active project or an open question and let the rest pass.

A raw dump overflow being narrowed by weekly triage into a bounded inbox — Capture stays safe when the has a weekly triage step that returns it to a bounded inbox.

Audio, screenshots, and transcripts take real storage

Meeting transcripts, voice memos, screenshots, and scanned pages consume far more storage than text notes. A one-hour meeting can run 5,000 to 15,000 words. A week of daily meetings adds up fast. Before you begin capturing audio and images, estimate how much storage your tools allow and how long raw sources should stay before review or archive. A system that runs out of space or becomes too slow to search is a system you will abandon.

Audio needs one extra step before normal processing: it has to become text. You do not need to decide how Whisper is installed or where every should go. Give Claude or Codex the folder path, tell it whether timestamps matter, and let it handle the transcription step. The same raw/ to processed/ to organized/ pipeline runs afterward.

Turn audio files into searchable transcripts

Paste this when recordings start landing in raw/.

I want my second brain to process audio files automatically. Second brain folder: [your second brain folder path, e.g. ~/Documents/second-brain] Audio location: [where audio files arrive, e.g. raw/audio/] Timestamp need: [plain text transcript is enough / create VTT because timing matters] Download OpenAI Whisper and set this project up so audio files become transcript files before the normal processing pipeline runs. Do these steps: 1. Inspect the second brain folder and choose the simplest local Whisper setup that works for this project. 2. Add or update the project files needed to transcribe new audio files from raw/ into text transcripts. 3. When timestamps matter, create a .vtt transcript as well as the plain text transcript. When timing does not matter, plain text is enough. 4. Save each transcript next to the audio source or in the right raw/ transcript subfolder, keeping a clear link back to the original audio file. 5. Process the transcript normally: catalog it, add YAML metadata, index it in brain.db, copy it into processed/, and file it under organized/. Show me the setup you added, the command or workflow I should use, and one example of an audio file becoming a transcript and then a processed source. Do not delete the original audio file.