Second Brain v2
// a field note on a personal agentic system, november 2026

Second Brain v2

Building an AI-native second brain, five months on. One plain-text vault. Scheduled agents that fill it before you wake. A handful of root logs and rules that let several agents work on it unattended without tripping over each other. This is how a portfolio-career operator runs a working life from one knowledge system, and the patterns you'd reuse to build your own.

1
vault
5
root logs agents share
11
scheduled tasks
25
skills
// what comes in Meetings Email Calendar Web clippings Phone capture Your own notes // the vault: one folder, plain markdown TASKS.md commitments NEW.md activity DECISIONS.md decisions GOALS.md open loops ANSWERS.md questions for you projects/ people/ meetings/ briefs/ emails/ notes/ _claude/ agent memory // what comes out Morning brief Meeting prep Inbox view Night shift Drafts for review Dashboard You decide, send, publish your decisions and answers go back in // agents read and write the same files Scheduled tasksovernight Desktop agentat your desk Coding agentat your desk Voice agentbriefs only The rulesAgents draft, you decide. Nothing is sent, paid for or deleted by an agent.
01 · The Pitch

Five apps, collapsed into one vault

A portfolio career (advisory, fractional leadership and angel investing across several work streams) means hundreds of stakeholders, dozens of live threads, and meetings that jump between completely different contexts. The fix wasn't a better app. It was deleting most of them and putting everything in one place an AI could operate.

The consolidation

Before: notes, tasks, and years of context were scattered across five tools: a chat assistant's project folders, a corporate note app, a personal note app, a dedicated task manager, and assorted others. Everything got exported and imported into one Obsidian vault, filed into archive folders. Then the old apps came off the dock and the phone home screen. What's left is Obsidian and Claude.

The core idea

The "app" is the vault plus a few agents that operate on it. There's no database and no backend, just plain Markdown that's portable, diffable, greppable, and future-proof. Every tool reads and writes the same files.

What a typical morning looks like

By the time the day starts, the vault already holds:

~00:00

Meeting sync

Yesterday's meeting transcripts filed into the right project folder, with every attendee wikilinked.

~04:00

Email gravity

One ranked view of inbox pressure: who's waiting, what's urgent, what was flagged, what's safe to ignore.

~05:00

Meeting prep

A brief for every meeting today: who's attending, the last notes with them, and open items still in the air.

~05:30

Morning brief

A chief-of-staff page: today's focus, replies owed, what moved since yesterday, what's waiting on someone else.

~06:00

Night shift

A divergent overnight read. It connects threads across the vault and asks a few uncomfortable questions. No task lists.

Home page

One dashboard

A single HTML file at the vault root, set as the browser home page, that shows all of the above.

Surfaces split by job, one brain

The work is split by job, not by data. An always-on machine owns the scheduled work, the backups and remote sessions. A laptop is for interactive sessions, in a desktop agent and a coding agent. A phone captures tasks and starts remote sessions on the always-on machine. All of them share the one synced vault.

02 · v1 to v2

What changed from v1 to v2

Version 1 of this page was written in June 2026. The bottom layer is the same. Most of what sits on top of it has been rebuilt after running it every day. If you read the earlier version, this is the diff.

Areav1 (June)v2 (now)
CoordinationThree root files: tasks, activity log, decisions.Five. A goals file and a questions queue were added, plus a written delegation policy.
Always-on agentRead the whole work vault; voice and chat from a phone.Sees the briefs folder only and reads briefs aloud. Remote work goes through a full agent session on the always-on machine.
Daily surfaceA daily note that transcluded the briefs.No daily notes. The dashboard is the home page.
BriefsA tasks brief and an operational night shift.A morning brief that checks claims against sent mail, and a night shift that only dreams.
Reliability"Check the date on the brief."Drift checks, a sync health check, sync jobs that run from outside the vault, a weekly hygiene run.
InstructionsOne memory file that kept growing.A short every-session file with a load-on-demand table. One source for style, one for each skill.
SkillsAbout ten.25, including a voice pass, method skills and a publishing chain.
PublishingNothing published by an agent.Still manual, with one named exception: a monthly post that ships behind a fact-check gate.

Three things learned the hard way

SILENCE

It fails without telling you

The backup sync froze for six days behind a stale lock file and logged thousands of errors. Nothing raised an alert. A scheduled system needs a heartbeat before it needs more features.

REVIEW

The click is the bottleneck

Agents drafted faster than their output could be reviewed. A dozen drafted replies sat unsent for two weeks. More drafts made it worse; batching decisions made it better.

ROT

Instructions drift

The same rule ended up in several files and the copies disagreed. Instruction files need one owner each and a script that checks the copies.

03 · The Store

The vault as second brain

Everything (notes, meetings, people, tasks, project context, even the AI's own memory) lives in one Obsidian vault of plain Markdown. This is the most shareable, least sensitive layer, and the foundation everything else assumes.

The folder skeleton

VAULT/
  TASKS.md            commitments: one file, sectioned by ### area
  NEW.md              activity: shared session log, every agent writes here
  DECISIONS.md        decisions and current priorities; Active entries are current truth
  GOALS.md            open loops agents advance one increment at a time
  ANSWERS.md          the one queue for agent questions, each with a default
  AGENTS.md           pointer file for any agent that isn't the main one
  dashboard.html      single-file home page over the vault
  briefs/            generated outputs from scheduled tasks
  agent-drafts/      agent output waiting for review
  projects/          one folder per work stream (each with meetings/, emails/, _staging/)
  meetings/          only uncategorised and external calls
  emails/            labelled emails filed as notes, with attachments
  people/            one page per person: your CRM
  notes/             everything thematic (tech, ideas, learning, talks)
  clippings/         web saves  ·  attachments/
  shared/            skills and scheduled-task prompts: the single source
  scripts/           sync, health and drift checks
  _archive/          dated archives and monthly log rollups
  _review-for-deletion/  where agents put things instead of deleting them
  _claude/           agent context and memory (not for human use)

The design decisions that make it work

People as pages

Every person gets a page in people/ with YAML frontmatter (org, role, aliases). Meeting notes wikilink attendees like [[Jane Smith]]. Obsidian's backlinks then give you a free CRM: open anyone's page and see every meeting and note they appear in, with zero upkeep. This is still the highest-value convention in the system.

Dataview for live queries

Project hubs don't maintain meeting lists by hand. They run queries like LIST FROM "projects/x/meetings" SORT file.name DESC LIMIT 15 that update themselves. People pages use WHERE contains(file.outlinks, this.file.link) to list every backlink.

Briefs are files, the dashboard is the view

Scheduled tasks write standalone files to briefs/. Nothing transcludes them any more. The dashboard reads them straight off disk, so a task only has to write one file and the view can change without touching the tasks.

One task file

No monthly pages, no nested system. A single TASKS.md at the root, sectioned by ### area, plus ## Waiting On and ## Recently Completed. The section list follows current work and is never padded to match an old list. Flat and thumb-typeable, so it can be edited from a phone.

Meetings and emails live with their project

Each project folder has its own meetings/ and emails/ subfolders, so a work stream's hub, documents and correspondence stay together. Only uncategorised or external calls go in the root meetings/.

Agent memory is separate

The _claude/ folder holds structured context the AI reads at session start: people profiles, project context, a glossary. It's deliberately separate from the human-facing people/ and projects/ pages. Don't merge them.

Two-tier AI memory, kept on a diet

Agents don't remember across sessions, so memory is files. A short file read every session, and the detail it points to:

_claude/
  CLAUDE.md            every-session rules only, with a load-on-demand table at the bottom
  memory/
    context/           operating setup, vault structure, archive safety, voice
    people/            deeper profiles by domain (beyond the CRM pages)
    projects/          per-project context
    glossary.md        acronyms, nicknames, codenames the agent must decode

The every-session file grew to nearly 5,000 words and was cut to about 1,400. The test for a line staying in it: does every session need this? Everything else moves to a topic file and gets one row in the table ("need X, read Y"). The glossary is what lets the AI understand shorthand like a colleague would.

Plugins doing the heavy lifting

Dataview · live queries Web Clipper · article capture Obsidian Sync · cross-device Mobile app · task capture on the go

Daily notes and their template plugin were dropped when the dashboard became the home page.

04 · The Engine

Automations that run while you sleep

Claude Cowork is a desktop agent that can read and write the vault folder and run tasks on a schedule. Each scheduled task is a self-contained prompt that writes one file. Daily briefs overwrite a rolling file in briefs/; the dashboard reads them. Most of the system's value still lands here.

TaskWhenWhat it does
meeting-sync~midnightFiles yesterday's meeting transcripts into the right project folder, wikilinking attendees and creating stub pages for new people. Dedupes via a state file.
email-gravity~04:00One ranked view of inbox pressure across every linked account. Reads only the priority split of each inbox. Sweeps incrementally via a state file.
periodic-digest~04:30An append-only, de-duplicated inventory of a recurring inbound stream (a running index, not a rolling file).
meeting-prep~05:00A prep brief for each of today's meetings: attendee context, last notes, open items.
morning-brief~05:30The chief-of-staff page: focus, meetings, replies owed, what moved since yesterday (checked against sent mail and chat), work in progress, waiting on, parked. Anything it can't verify is named once, not nagged.
night-shift~06:00Dream only. A random deep sample of the vault plus your own recent words as input, and a ledger that bans it from repeating a move, a theme or a note.
ai-brief~06:30A daily read of the field, written to a rolling file with a dated archive copy. A weekly synthesis is built on demand from the last seven archives.
capture-and-filehourly, daytimeStarred emails become tasks. Emails given a "vault" label are filed as Markdown notes, with attachments, into the right project's emails/ folder.
vault-hygieneweeklyApplies safe, reversible clean-up and proposes the rest in a report. Never deletes; audits inbound links before moving anything.
monthly-post1st of monthRuns three research skills in parallel, fact-checks the result, writes a public post. The one task allowed to publish, and only if every failed check has been fixed.
monthly-ledger1st of monthRebuilds a month of work records from calendars, sent mail, meeting notes and session logs, with an evidence line per entry.

The lessons baked into this model

CONSOLIDATE

Fewer, self-contained tasks

An earlier version had separate triage, stale-reply, action-extractor, Monday and Friday jobs. They were collapsed into one inbox sweep and one brief. A few self-contained tasks beat many brittle chained ones.

SILENT FAILURE

Rolling files fail quietly

A rolling output overwrites in place, so a dead pipeline just leaves yesterday's file with no error. Every brief prints its own date. Check the date, not the file's presence.

INCREMENTAL

State files for sweeps

Tasks that sweep an inbox keep a small JSON state file so each run only processes what's new, with per-item dedupe as the safety net.

REVIEW

Read your own briefs back

After a few months, 66 past briefs were reviewed in one pass. The overnight read had used one template for a month. A "did not move" list was built on absence of evidence, and on one day two of its five items were wrong. Both were redesigned.

NARROW

Cut the input, not just the output

The inbox sweep's needs-reply list grew past 500 items and stopped being read. Pointing the sweep at the priority split only took it from 567 to 71 in one run.

MODELS

Pins get retired

Tasks pointed at a specific model failed on every run once that model was withdrawn, with no visible error. Inherit the default unless a task has a reason to pin, and check the pins when a model is retired.

Model choice

One more concrete lesson: a different model family was trialled for "cognitive diversity" but its max output-token limit was far lower and it errored on real, tool-heavy runs. Check a model's max output tokens against your actual workloads before committing.

05 · The Rules

How several agents share one vault

This section didn't exist in June. Once a desktop agent, a coding agent, scheduled tasks and a voice agent all write to the same folder, across two machines, coordination becomes most of the design. The answer is five plain files at the vault root and one written policy.

Five root logs

FileHoldsThe rule that makes it work
TASKS.mdCommitmentsOne file. Completed items move to a dated "recently completed" block. Agents add tasks; they don't restructure the file.
NEW.mdActivityOne terse entry per session that changes the vault, newest first, tagged with the surface that wrote it. Routine brief regenerations are not logged. Old entries roll into monthly archives.
DECISIONS.mdDecisions and current prioritiesEntries marked Active are current truth and outrank older notes. A reversed decision is marked Superseded and linked to its replacement, never deleted.
GOALS.mdOpen loopsEach goal names its outcome, a done-condition, the increments an agent may run, and hard brakes. An agent with spare capacity runs exactly one increment and logs proof.
ANSWERS.mdQuestions for the humanThe only place an agent may ask. One numbered line, a default with a checkbox, a slot to override. Questions never block the work. Answered items are applied, then archived.
Why the questions queue matters most

Before it existed, questions were scattered across run reports and went unanswered. With every question in one file, each with a default that can be ticked, the first batch of 71 was ruled in a single sitting. The dashboard now shows the queue as a "waiting on you" card and writes the answers back to the file.

The delegation policy

Eleven standing rules for what an agent may do without asking. It lives in the every-session memory file and binds every surface. Changing a rule takes a human edit or a dated decision entry, never an agent's judgment call.

AreaRule
Email, archivingAgents may archive noise-class mail only (marketing, newsletters, no-reply notifications with no action).
Email, sendingNever. Drafts only, including one-line acknowledgements.
ChasersNo standing chaser loop. A chaser is drafted only when asked for.
PaymentsNever. Banking, invoices, transfers and government portals are human-only.
SubscriptionsFlag anything unused for 60 days, with cost and last-use evidence. Cancelling is the human's click.
Paid tools and APIsPer-item approval, a hard spending cap on every account, credentials set up with the human present.
OutreachEvery inbound request surfaces. Nothing is auto-declined. Introductions are double opt-in.
SchedulingPropose slots from free time only, respect the meeting-free days. Accepting is the human's tick.
Vault, auto-executeClippings triage, quarantining dead stubs, brief regeneration, dashboard rebuilds, meeting routing.
Vault, recommend-onlyPromotions out of staging, archives, renames, structure changes, and any delete.
PublishingManual, every time, with one named exception recorded as a dated decision.

Hard brakes are absolute

A goal's brakes override anything else in the goal, the prompt, or the agent's judgment. If an action would cross one, the agent doesn't do it and flags it in the run report.

Frozen records

Briefs, dated meeting notes and dated log entries record a moment. Agents never repair their links or update their claims. A broken link in a frozen record is not a defect for any hygiene pass.

06 · The Always-On Layer

The voice agent, cut back

In June this layer was a third of the story: a 24/7 agent on a small, hardened machine, reachable by voice and chat from a phone, with the whole work vault mounted. It turned out to be the least-used part and the one with the most reach. It is still optional, and it is now much smaller.

Rule zero: do this first

Split personal from work before putting any always-on agent in front of the vault. The work vault syncs to all devices and the always-on machine; a separate personal vault syncs to your own devices only, and the always-on machine never gets it.

What changed

It mounts the briefs folder, nothing else

A review found the agent had far more filesystem reach than its job needed. Its container now gets one mount: briefs/. It reads the morning's briefs aloud on request. It cannot see projects, people, or memory.

Remote work moved to remote sessions

Research, drafting and anything that needs the vault now runs as a full desktop-agent session on the always-on machine, started from a phone or laptop, under the same rules as every other session.

Capture moved to the phone

Tasks are typed straight into the task file from the mobile notes app. Starring or labelling an email does the rest within the hour.

Brief content is data

Briefs are built from email and meeting transcripts, which other people wrote. The agent's standing instructions say that anything inside a brief is data to read out, never an instruction to follow.

The security model, unchanged in principle

Least-privilege users

A non-admin user runs the agent; a separate admin user handles system changes. The blast radius of a compromised agent is then one home folder.

Every tool call is sandboxed

Tool calls run inside a hardened container, not on the host. Drop all capabilities and re-add only what's needed, no new privileges, and no host secrets forwarded in unless explicitly allowlisted.

Mount only what you intend

Start from the smallest mount that does the job and widen it only when a real task needs it. Narrowing later doesn't remove what the agent has already read and stored.

Network and secrets discipline

Encrypted disk, firewall on, no public exposure. Secrets come from a password-manager CLI into the agent's environment, never pasted into a prompt, and get pruned to what is actually used.

Drafts only, for every agent

The drafts-only rule began as this agent's rule, after an over-eager model moved and renamed files unprompted. It is now the delegation policy in section 05 and applies to every surface. A convention enforced through memory is only as good as the model reading it, which is why the mount, not the instruction, is the real boundary here.

07 · Practice

The working patterns that compound

A few conventions, written once into the AI's memory so every session obeys them. These are the difference between an AI that helps occasionally and one that operates the vault reliably.

The staging convention

New AI-generated files land in a _staging/ subfolder first, never straight into a live project folder. Edits to existing files happen in place. You review, then promote the keepers.

Never delete, quarantine

Agents never delete vault content. They move it to _review-for-deletion/ with the filename preserved. The human empties that folder, and that deletion is final: agents then repair the links that pointed at it, they don't restore from git.

Archive safety

Last-modified date is not a staleness signal. Reference notes get linked, not edited, so they look old while being load-bearing. Audit inbound links before archiving anything, and repoint every link whose target moves. Links into the archive are fine; the only defect is a link that doesn't resolve.

Context lives in the logs

Sessions don't share memory. What carries between them is the decisions log, the goals file and the session log, read at the start of every session. Per-project handoff notes are no longer kept.

One source per instruction

Style rules live in one file that the others copy verbatim. Each skill lives in one folder, and every other location is a link to it. Scripts check for drift between copies, and between what's in the vault and what's deployed.

Sync from outside the vault

The vault backs up three ways: live sync, a versioned git push every few minutes, and a passive cloud mirror. The scripts that do the git push run from copies installed outside the vault, because agents can write inside it. A health check watches for three things: a failed push, commits piling up unpushed, and the job not running at all. It prints nothing when healthy.

Protect the review step

Drafting is cheap and review isn't. Any new automation that produces more prose to read makes the system worse unless it also removes a decision. The patterns that helped were all batching: one questions queue, one "waiting on you" card, one sitting.

Custom skills as reusable playbooks

Skills are instruction sets the AI loads when a trigger phrase fires: brand guidelines, a document format, an analysis playbook. Each is a folder with a SKILL.md and an optional references/ folder of worked examples. Write one once you've done the task by hand a couple of times and know the pattern you want. There are 25 now. The ones that changed how the vault is used:

PROMOTE

/ideate

Scans recent meetings, notes and staging folders for ideas worth their own page, and graduates each into a standalone note with an evidence trail.

FILTER

/crit

The counterweight. Kills the weak ideas, challenges assumptions, names the few threads worth pulling. Run the two as a pair.

RECORD

/decision

Captures a firm decision or a change of priority into the decisions log, so every other agent reads the same current truth.

VOICE

Voice pass

Rewrites an AI-drafted document into the owner's own voice, working from verbatim examples of their writing, with a hard gate that no fact, number or name changes.

EXPLAIN

Visual explainer

Turns a memo into a picture-led HTML page. Chained after the voice pass and before a brand skill, it's the standard route from a draft to something shareable.

REVIEW

/coach-me

Reads a week of vault activity and writes a short coaching brief on how the operator is working, not on the ideas.

08 · The Surface

One HTML file that is the home page

A single self-contained dashboard.html at the vault root that reads, and writes, vault files directly in the browser via the File System Access API. No server, no build step. It replaced the daily note and is set as the browser home page.

TAB

Home

Morning brief first, then a "waiting on you" card of open questions that can be answered in place, meeting prep, today's focus from the task file, the top of the inbox view, and the night shift folded away.

TAB

Tasks

Parses the task file by ### section. Board view with drag-drop between sections, list view, checkboxes that toggle completion.

TAB

Goals

One card per goal with its last logged increment, plus current priorities and waiting-on items from the decisions log.

TAB

People

A searchable CRM over the people/ folder. Grid and table views, click-to-edit fields read from frontmatter.

TAB

Briefs

Every rolling brief, driven by one list of sources. Older briefs load on open.

TAB

AI

The daily read of the field, the weekly synthesis, and the latest monthly radars.

TAB

Ideas

The ideas index as cards grouped by theme: the output of /ideate, surfaced where it gets seen.

TAB

Memory

Browse and edit the _claude/ memory files directly, to see what the AI "knows" and correct it.

Why a page and not a note

The dashboard keys off file paths and Markdown conventions, so it works on any vault that follows the same structure. It can write back (tick a task, accept a default answer), which a transcluding note can't. The vault handle is persisted in the browser, every edit autosaves, and it re-reads the briefs when the tab regains focus.

09 · Build Your Own

Replicate this in three phases

Phases 1 and 2 deliver most of the value. You can stop after Phase 2 and have a working AI-native second brain. Phase 3 is optional. Build phase by phase; don't begin one until the previous one verifies.

The vault Plain markdown in one folder: projects, people, meetings, notes. Every tool reads and writes the same files. Five root logs Tasks, activity, decisions, goals, questions. How each agent sees what the others did. The rules A delegation policy: agents draft, you decide. Staging first, never delete, hard brakes on every goal. The agents Scheduled tasks overnight, a desktop agent and a coding agent by day, remote sessions, a voice agent for the briefs. What you see One dashboard as your home page, and the briefs the agents wrote before you woke up. You // build it in this order

Decide these first

  • Your areas. The 4 to 10 top-level work streams that become your task sections and project folders.
  • Personal vs. work. Anything you wouldn't want an agent to read goes in a separate personal vault, from day one.
  • Connectors. Which inboxes, calendars and meeting tools the AI may read for the briefs.
  • What agents may do alone. Write your own version of the delegation policy in section 05 before anything runs unattended.
  • Backups. Live sync is a given; pick a git host and an optional cloud mirror.

Phase 1: the vault

1

Create the folder skeleton

The structure from section 03. Keep the folder names; the later automation keys off them.

2

Install the plugins

Dataview and the Web Clipper, plus a sync method. Skip daily notes.

3

Seed the CRM

Make a couple of people pages with frontmatter and a Dataview backlinks block; write a meeting note that wikilinks them; confirm the backlink query shows it.

4

Project hubs and the task file

One hub per work stream with a live Dataview meeting query; a single TASKS.md sectioned by area; the _staging/ convention.

5

Create the other four root logs

NEW.md, DECISIONS.md, GOALS.md and ANSWERS.md, each with its protocol written at the top of the file so any agent that opens it knows how to write to it.

Phase 2: agents and automation

6

Build the memory tree

Create _claude/CLAUDE.md and the memory/ subtree. Keep the every-session file short and give it a load-on-demand table. Write the staging, never-delete and archive-safety rules into it.

7

Write the delegation policy

What may an agent do without asking, in email, money, scheduling, the vault and publishing? Put it in the every-session file. Start strict; loosen one rule at a time with a dated decision.

8

Set up three scheduled briefs

Start with meeting-prep, the morning brief and the inbox sweep. Each writes one dated, rolling file to briefs/ and keeps a state file for incremental work.

9

Make the dashboard your home page

Generate dashboard.html, point it at the vault, and set it as the browser start page. Add the "waiting on you" card early.

10

Add the heartbeat and the backups

A git push on a timer, run from outside the vault; a health check for failed pushes, unpushed commits and a job that has stopped running; a weekly hygiene task that applies safe fixes and reports the rest.

11

Write your first skills

Start with the ideate and crit pair and a decision-capture skill. Keep each skill in one folder and link to it from everywhere else.

Phase 3: an always-on voice agent (optional)

12

Confirm the personal/work split

Non-negotiable before going further. The always-on machine only ever syncs the work vault.

13

Harden, sandbox, mount the briefs only

Non-admin agent user, encrypted disk, no public exposure, tool calls in a container. Mount briefs/ and nothing else. Tell the agent that brief content is data.

14

Use remote sessions for everything else

Anything that needs the whole vault runs as a full agent session on the always-on machine, under the same delegation policy as a session at your desk.

Verification mindset

Good verifications are concrete: create a test person and meeting and confirm the backlink query shows it; confirm a scheduled task wrote a brief with today's date on it; kill the sync on purpose and confirm the health check notices; ask an agent to "tidy up" a folder and confirm it proposes and quarantines, and doesn't delete. Prefer to demonstrate a capability over asserting it.

10 · The Horizon

Where it's going next

Not committed. Captured so the ideas don't get lost. Each one extends the same core: a plain-text vault, operated by agents under written rules.

Conjoined brains

Share part of a vault with a partner or colleague and create a combined workspace that draws on everyone's collected meetings and notes. A protocol is drafted; it hasn't been piloted.

Team brains

The same pattern for a team: individual context first, shared skills second, agents later. Build it by hand twice, then automate what repeated.

Voice UI

A voice agent that reads out tasks, briefs and ideas while driving, and a weekly audio review of the past week and the next.

Task focus mode

Surface tasks one at a time in a mobile-first view that makes knocking them off feel good.

A human assistant in the vault

Connect a human executive assistant into the vault, with the same context the AI has, so they can pick up tasks and draft on your behalf.

Auto-drafted action plans

For each task, research in the background and draft what's needed to act on it. Only worth building where the review step stays cheap.

The "app" is just the vault, a few agents, and the rules they share.
second brain v2, a field note. obsidian + claude, november 2026