Resources / AI Ops setup

Set up the AI Ops Playbook, step by step

The playbook explains why. This is how: the exact folders, settings, files and prompts I use, in the order I would build them again.

SetupAbout 3 hours
Running itA few minutes a day
You needObsidian, Claude Code
Ten steps at a glance
01

Create the vault and its folders

A home for every kind of note, numbered so it sorts in the order you use it.

15 min
  • Install Obsidian, choose Create new vault and name it. Mine is Personal Vault, on the Desktop.
  • Create only the folders you will use this week. Add the others when the first real note needs one; never create empty folders to make it look complete.
  • Give each working folder a short README.md saying what goes in and who clears it.
  • Number projects (0301, 0302 ...). When one ends, move the whole folder to 05 Archive.
Folder map
00 Inbox/        fast capture, unedited; the agent triages it
00 Templates/    daily note, tracker, processed call
00 Trackers/     running lists you add to
01 Daily/        one note per day
02 Meetings/     calls worth keeping as a note
03 Projects/     things with an end date
04 Areas/        ongoing parts of work and life
05 Archive/      finished projects, read only
Clippings/       things saved from the web
Dashboard.md     the one page you check in on
TASK HQ.md       every open task, in one list

Example README. Mine for the Inbox: "Drop anything here, unedited. Don't sort it, don't title it nicely. Claude triages this during the check-in. Nothing sits here forever."

02

Set Obsidian up so notes land in the right place

New notes go to the Inbox, daily notes to their folder, templates fill themselves in.

20 min
  • Settings, Files and links: set Default location for new notes to "In the folder specified below", then pick 00 Inbox.
  • Settings, Core plugins: turn on Daily notes, Templates, Backlinks, Properties and Bookmarks.
  • Settings, Daily notes: new file location 01 Daily. Settings, Templates: template folder 00 Templates.
  • Settings, Community plugins: turn them on and install the plugins below. Start with the first two.
  • Save the daily note template below in 00 Templates.
PluginWhy
TemplaterTemplates that fill in today's date
TasksTasks with deadlines and repeats, collected in one list
QuickAddOne shortcut to drop a thought into the Inbox
Natural Language DatesType "next Friday", get a real date
Granola syncMeeting notes into the vault, if you use Granola
00 Templates/Daily Note Template.md
---
title: <% tp.date.now("YYYY-MM-DD") %>
created: <% tp.date.now("YYYY-MM-DD HH:mm") %>
tags: []
---

## Today's three
1.
2.
3.

## On my mind

## Done
03

Agree the note conventions

The agent can tell a current rule from an old draft, and every file name explains itself.

15 min

Four statuses, two markers

ValueThe agent should
draftUse with care and say it is a draft
activeUse it
supersededNever cite it; follow superseded_by
archivedIgnore unless asked about the past
[VERIFY]Check the live system first; the date says when it was true
[OPEN]Unresolved. Ask, do not guess

Three tiers: what the agent reads first

TierFoldersUse it
A · current truthRoot files, Projects, Areas, people, decisionsFirst, for every question
B · evidenceDaily notes, Meetings, dated reportsFor proof, history and conflicts
C · intake and historyInbox, raw imports, ArchiveOnly with a reason
NeverPasswords, plugin and editor state, chat logs, TrashExcluded from search and backup

The durable rule. An important fact is not "known" until it lands in a tier A note. A daily note or meeting can hold it first, but should link to where it now lives.

Two habits on every lasting note

  • Name files YYYY-MM-DD-doctype-project-details.md, all lowercase, for example 2026-10-14-guide-personal-my-writing-voice.md.
  • End with ## Changelog and one dated line per change, so you can see what the agent changed and why.
Frontmatter for every lasting note
---
title: Lead delivery runbook
type: sop            # guide, sop, plan, analysis, tracker
status: active       # draft, active, superseded, archived
updated: 2026-10-16
tags: [ops, delivery]
sources:
  - "[[2026-10-13 ops sync]]"
superseded_by:       # only when status is superseded
---
04

Write your personal layer

One file, "How I operate", that tells any agent how you think and work. It travels with you to every job and tool.

45 min

This is the most important file. Mine lives at 04 Areas/Ways of Working/How I operate.md. It has seven short sections:

SectionWhat goes in it
Who I amYour role in one line, and your known weakness
How I thinkFive or six principles, each with its reason
Working contractHow you want work handed to you and closed
Rules of thumbA do and don't table
Knowledge loopCapture, classify, file, write back, clean up
GuardrailsWhat the agent must never do
Start of sessionWhat it reads first, every time
How I operate.md · template with real lines from mine
---
title: How I operate, the personal layer for any agent
type: guide
status: draft
updated: YYYY-MM-DD
---

# How I operate
Paste this into any agent before the company layer.

## 1. Who I am
- [Your role in one line, and what you are known for.]
- [Your known weakness.] e.g. "I tend to over build. When in doubt:
  do less, finish it, then improve."

## 2. How I think
1. Context before instructions. Every priority says why, what happens
   next, and who does it, decides and is informed.
2. Stable or volatile. Numbers, owners and dates are read live, or carry
   [VERIFY] and the date they were true.
3. True apart from should. "How it runs today" and "what would fix it"
   are separate sections.
4. Gaps are answers. "To be mapped" and [OPEN] beat a plausible guess.
5. A correction is a rule, written down with its reason.
6. Judgement stays with me. Challenge my assumptions; I decide.

## 3. The working contract
1. Read the brief for four things: deliverable, reader, done, deadline.
2. Name the deliverable back in one line, then build.
3. Vault first, live systems second, your own knowledge last.
4. A task is done only when I say it is.
5. Close in 3 to 6 lines: what you made, where it is, what is open,
   what you wrote back.
6. Ask before anything outward facing: sending, posting, publishing.

## 4. Rules of thumb
| Do                                  | Don't                           |
|-------------------------------------|---------------------------------|
| Lead with the answer                | Open with background            |
| Number action items 1. and 1.1.     | Bullets for things to point at  |
| Link every source inline            | Name a source without a link    |
| Mark volatile facts [VERIFY] + date | State a number nobody checked   |

## 5. The knowledge loop
1. Capture. Nothing stays only in the chat.
2. Classify. Stable or volatile? Which note owns it?
3. File and link both ways.
4. Write back the rules, not only the facts.
5. Weekly: empty the inbox, merge duplicates, re-verify anything older
   than 30 days, archive what is superseded.

## 6. Guardrails
1. Personal data stays in approved systems.
2. No credentials, ever. Flag an exposed secret; never copy it.
3. Never edit raw transcripts. Correct them in the processed note.

## 7. Start of session
1. Read this file, then the company layer.
2. Read today's and the last daily note.
3. Say in two lines what you understand the state to be, then start.

## Changelog
- YYYY-MM-DD: created.
Prompt · let the agent interview you instead
Interview me so you can write my "How I operate" file.

Rules:
- Ask no more than four questions at a time, most important first.
- After each round, show a checkpoint: Confirmed, Assumptions,
  Open questions, Contradictions, Next round. Let me correct it.
- Challenge vague answers. Let me say "unknown" or "skip".
- Never invent a fact about me.

Cover, in rounds: what I am responsible for and judged on; work that
repeats; where I lose time re-explaining context; how I decide; how much
detail I want; when to challenge me; how I write; what you must never do
on my behalf.

When the rounds are done, draft the file using the seven sections
I give you, and list what is still unknown.

Worth adding later: a voice guide. I had the agent read about 130 of my own work messages and describe how I write. Linked from this file, it makes drafts come back sounding like me.

05

Write the company layer

Each employer gets its own vault and one manual the agent reads automatically.

30 min
  • Create a separate vault per employer, with the same folders. Never mix two employers in one vault.
  • At its root, create CLAUDE.md. Claude Code reads it automatically when you start in that folder.
  • Point to your personal layer by its path. Do not copy it in, so there is only ever one version.
  • Fill in what you know. Mark everything else [OPEN].
  • Add a short MEMORY.md: current focus, durable choices, open questions. Prune it weekly; it is not a history log or a second task list.
  • Register every source before the agent relies on it (table below). A connector being available does not make a source the authority.
CLAUDE.md · company vault root
# [Company] operating manual

Read first: ~/Desktop/Personal Vault/04 Areas/Ways of Working/How I operate.md

## The business in one screen
What we sell, to whom, how money comes in, what good looks like now.

## People
| Name | Role | Owns | Ask them about |

## Systems
| System | Used for | Source of truth for | Link |

## Vocabulary
| Term | Means here |

## Who wins a conflict
1. Source system data, read today
2. The owner's dated statement
3. My own dated decision
4. Client or partner in writing
5. Team members on calls or chat
6. AI call summaries (check the transcript)
7. Older notes and wikis

## Open questions
| Question | Who can answer | What it blocks |

Source registry: one row per source

SourceOwnsAuthorityFreshnessAgent may write?
ATSCandidate stage and statusLive truthQuery liveNo
Company manualRules, people, systemsCurrent truthReviewed monthlyDraft changes only
Meeting transcriptsWhat was saidEvidencePer callNever edit the raw file
Chat channelsDated messagesEvidenceTargeted searchNo

One first source for every recurring question

QuestionLook firstThen, for proof
What matters right now?MEMORY or the task listRecent daily notes and meetings
Where is this project?The project noteTasks, meetings, decisions
What do we know about this person?Their people noteCalls and messages
What is approved company truth?The company manual or systemDated internal evidence

If two notes both look like the answer, the system is not finished. Pick one owner, point the other to it, and keep anything unique as evidence.

Shared wiki already a mess? Do not move anything yet. Ask the agent to audit it first: page counts, duplicate titles, processes with no page. Then build the new structure next to the old one.

06

Connect Claude Code

The agent runs inside your vault and starts every session by reading your layers.

15 min
  • Install Claude Code using Anthropic's instructions at code.claude.com/docs and sign in.
  • Open a terminal in the vault and run claude.
  • Send the first prompt below before anything else.
  • Keep it in ask-before-editing mode for the first week, so you see every change.

No terminal? Use a Claude Project. Upload your layer files as Project knowledge and paste the start-of-session rules into the Project instructions. Uploads are snapshots: when a file changes, replace the upload. Anything learned only in a chat is not remembered until you write it back to the file.

Terminal
cd ~/Desktop/"Personal Vault"
claude
First prompt
Read "04 Areas/Ways of Working/How I operate.md" in full. Then list the
folders in this vault and read each README. Tell me in five lines how you
understand my system, and ask me about anything marked [OPEN].
Do not create or change any file yet.
07

Run the rituals

Five small habits. Without them, the notes go stale within weeks.

Ongoing
WhenWhat happens
Start of sessionAgent reads the layers and last daily note, states the situation in two lines
After each callOne processed note: corrections, decisions, facts, actions
End of sessionWrite back what should outlive the chat, then a short close
Each morningToday's three, what is dragging, what is untouched
New documentIngest it: preserve, find the owner note, update only what it supports, report what changed
"That's wrong"Correction: replace the claim, keep the old one as history, re-ask the question
WeeklyRead-only review first, then only the fixes you approve
Prompt · after a call
Here is the transcript of my call with [who] about [topic].
Create one note in 02 Meetings named YYYY-MM-DD-meeting-[topic].md with:
1. Corrections: what the AI summary got wrong, checked against the transcript
2. Decisions: who decided what (also add each to the decision log)
3. Facts: new facts, each marked stable or volatile
4. Actions: numbered, each with an owner
Then update the notes that own these facts. Do not edit the raw transcript.
Prompt · read this and add the useful parts
Read this and add the useful parts to my notes.
1. Preserve it: save the original with title, author, link and date.
   Check first whether I already have it.
2. Find the note that should own each useful fact. Create a new note
   only if no existing one can own it.
3. Separate confirmed facts, decisions, actions, open questions and
   claims that still need checking.
4. Compare with what the owner note says now: confirms, refines,
   contradicts or replaces? Show conflicts instead of choosing.
5. Propose the smallest update. Wait for my OK.
6. Report: source saved, notes changed, conflicts, and one future
   question that should now find this.
Prompt · that is wrong
That is wrong: [the correction].
1. Find every note and rule that states the old version.
2. Mark the old claim superseded with today's date; do not delete it.
3. Update the one note that owns this, and "How I operate" if it is
   about how I work.
4. Ask the question that gave the wrong answer again, and show me
   the new answer and where it came from.
Prompt · end of session
Before we stop: write back anything from this session that should outlive it.
Corrections I gave you become rules in "How I operate", with the reason.
Closed [OPEN] items get removed where they were. Then close in 3 to 6 lines:
what you made, where it is, what is still open, what you wrote back.
Prompt · weekly clean-up
Weekly review. Read only: change nothing until I approve.
Check:
1. 00 Inbox items with lasting value
2. Duplicate notes that both look current
3. [VERIFY] facts past their expected shelf life (age alone is not proof)
4. Rules that contradict each other across my layer files
5. Daily notes or transcripts holding facts no tier A note has
6. Broken links and notes nothing links to
Rank every finding as: fix now, merge, archive, promote, keep,
or needs my decision. Then refresh Dashboard.md, marking what you
touched "edited by Claude".

Keep the dashboard honest. The agent marks everything it changes "edited by Claude" and never touches my personal notes there, so I always know what is mine.

08

Test it with real questions

Proof that the right note answers, before you add more material.

30 min
  • Write down ten questions you will genuinely ask this month.
  • Next to each, write the note that should answer it first.
  • Ask them in a fresh session. Check which source the agent used, not only whether the answer sounds right.
  • When the wrong source wins, fix the route (title, owner note, index, registry) before adding anything new.
TestPasses when
Who am I and how should you work with me?It answers from "How I operate", not guesses
Where is the current state of this project?It names the project note first
Which source wins on this disputed point?It names the authority and shows the conflict
What changed after this document?It lists exact notes changed and what did not
What is unknown or too stale to trust?It names gaps instead of filling them
That preference is wrong; use this insteadThe new rule wins and the old one stays as history

Fix the route before the volume. Generic answers almost never mean "add more files". They mean the owner note, title or authority is unclear.

09

Back it up privately

Full history and a second-laptop copy, with private files kept off the internet.

20 min
  • Create a private, empty repo on GitHub.
  • Add the .gitignore below before the first push. It decides what never leaves your laptop.
  • Push once from your own terminal, and read the file list before you commit.
  • Ask the agent for a one-command sync script that refuses to commit private file types or anything that looks like a password, key or bank number. I run mine as bash syncpersonal.sh.
.gitignore
.obsidian/workspace.json
.obsidian/workspace-mobile.json
.obsidian/cache
.trash/
.DS_Store

# Plugin settings can hold API keys
.obsidian/plugins/*/data.json

# Agent session logs and editor state are not knowledge
.claude/
.obsidian/plugins/*/main.js

# Documents with personal data stay on this laptop
*.pdf
*.docx
*.xlsx
*.csv
*.zip
*.eml
Terminal · first push
cd ~/Desktop/"Personal Vault"
git init -b main
git add -A
git status        # read the list: nothing private should be in it
git commit -m "First backup"
git remote add origin git@github.com:YOUR-NAME/personal-vault.git
git push -u origin main
10

Turn repeated asks into skills, then automate last

Saved instructions first; automation only once the manual loop is trustworthy.

Later

The test: you have typed roughly the same request three times, and corrected the output the same way at least once. Mine include:

  • A daily update for my manager, in a fixed format
  • A list of what only my manager can decide, with time estimates
  • A file naming rule applied before anything is saved
  • A morning start that proposes today's three

Automate in this order, one level at a time

1Read-only search and answers
2Spotting duplicates
3Suggesting where something should be filed
4Drafting updates for you to approve
5Approved, low-risk write-back
6Anything that leaves your notes: sending, posting, changing other systems

Tool access is not approval. Even when the agent can send, delete or publish, it asks first. Decide what needs your OK before you connect anything.

✓

Your first week

Small and in order. Ten honest lines beat a polished page you never finish.

7 days
Day 1Steps 1 to 3: vault, settings, conventions
Day 2Step 4: a rough "How I operate"
Day 3Step 6: connect the agent. Fix what it misunderstood in the file, not the chat
Day 4Step 7: ingest one real document and process one real call
Day 5Step 5 for your main employer, then step 8: ask your ten questions
Day 6Fix the routes that failed. Add nothing new
Day 7First read-only weekly review; apply only the fixes you approve
LaterStep 9 before anyone else gets access; step 10 once requests repeat
!

Mistakes I made, so you can skip them

Every rule above came from one of these.

2 min read
MistakeWhat to do instead
Reorganised folders three times in eight daysKeep the shape above; use frontmatter and saved views
Seventeen statuses, nothing marked oldFour statuses with fixed behaviour
Trusted an AI call summaryFacts come from the transcript, not the summary
Copied daily numbers into notesLink the live dashboard; keep only the rule
Relied on a "private" flagA separate ignored folder; a flag does not stop a push
Let two jobs share one vaultSeparate vaults, and filter at the source
Starting with hundreds of imports or a prompt libraryStart with the small kernel; add material only when a real question needs it
+

Built on

My own setup at two companies, sharpened with practices from people I learn from.

Credits
My vaultsThe folders, layers, rituals and mistakes, tested in daily work
Russel Mei De JesusHer AI Workspace Starter Kit: retrieval tiers, the source registry, ingest and correction loops, read-only cleanup, acceptance tests and the automate-last order. Her portfolio

Read the why behind it

The playbook explains the thinking in fourteen spreads: the write back loop, tiers of truth, who wins a conflict and the two layers.