Obsidian is how you give your AI a memory

If AI is becoming your agentic operating system, Obsidian can become the part that remembers what you believe, what you are building, how you work, and what the robots should absolutely stop doing.

Most people are using AI like a very smart stranger with short-term memory problems.

You open a chat. You explain your life. You explain your company. You explain your goals. You explain your voice. You explain the project. You explain the context behind the project. You explain the weird little preference you have that is not rational but is deeply yours and, obviously, correct.

Then you get an answer.

Sometimes it is good.

Sometimes it is “good,” which is different and spiritually dangerous.

Then you close the tab.

The next day you come back and do the whole ritual again, like some productivity-themed Greek punishment. Roll the boulder up the hill. Re-explain the audience. Re-explain the strategy. Re-explain that “make it punchier” does not mean “turn this into a LinkedIn post written by a regional sales director trapped in a Canva template.”

This is not an AI problem. It is a memory problem.

If the next era of work is agentic, meaning we are going to have more AI systems doing more repeatable work on our behalf, then the question is no longer just…

“How do I prompt better?”

It is: “Where does my context live?”

Where do your standards live? Where do your decisions live? Where does your strategy live? Where do your examples live? Where do your customer insights, personal taste, project notes, working theories, recurring prompts, content ideas, lessons learned, and things-you-never-want-the-AI-to-say-again live?

For a lot of people, the honest answer is:

Everywhere.

A few Google Docs. A notes app. Slack messages. Screenshots. Some cursed downloads folder. Ten chat threads. One Notion workspace you swear you are going to clean up. A voice memo named “idea???” from March. Maybe a notebook if you are hot and mysterious.

Fine.

Human.

But not agentic.

Because agents cannot reliably use context they cannot find, and you cannot build a real AI operating system on top of scattered digital confetti.

This is why I think Obsidian is about to matter a lot more.

Not because it is trendy in the knowledge-management corners of the internet where people say things like “second brain” with the overwhelming intensity of a man optimizing his grocery list.

Because Obsidian is simple in the way that matters…

It is a folder of plain text files.

That is the magic.

Not the graph view. Not the plugins. Not the theme that makes your notes look like a cyberpunk monastery. Those are fun. Go with god.

The important part is that Obsidian lets you build a local, readable, linkable, durable memory layer for yourself and your AI.

A place where the human context lives.

A place your agents can eventually read from.

A place you can understand without needing a computer science degree, a Zapier fever dream, or a consulting retainer from someone named Chase who says “knowledge graph” too much.

Wanna build one? Let’s build one.

first, what are we actually making?

We are not making a notes app.

You already have notes.

You have too many notes. I say this with love. We all do. Notes are where good thoughts go to become sediment.

We are making an **agentic memory system**. Wish I could add digital sparkles to that.

Don’t worry, it sounds more complicated than it is.

An agentic memory system is just a place where you store the context an AI needs to help you do better repeatable work.

That context usually falls into six buckets:

1. **Who you are** — your role, goals, standards, voice, preferences, constraints.

2. **What you are building** — active projects, products, campaigns, ideas, decisions.

3. **What you know** — research, customer insights, market notes, source material.

4. **How you work** — workflows, checklists, review criteria, recurring tasks.

5. **What you have decided** — decisions, tradeoffs, reasons, things you are not reopening every Tuesday like a chaos advent calendar.

6. **What good looks like** — examples, templates, past work, style references, quality bars.

That is memory.

Not “remember my favorite color.”

Not “call me bestie unless I am sad.”

Actual work memory.

The stuff that lets an AI stop acting like a clever tourist in your life and start acting like a trained collaborator inside your operating system.

why Obsidian?

There are many tools you can use for memory.

Notion. Google Docs. Apple Notes. Mem. Tana. Reflect. Roam if you still have a graph-shaped tattoo on your soul. A folder of Markdown files. A legal pad and a dream.

Use what you like.

But Obsidian has a few advantages that make it especially good for AI memory.

1. Your notes are plain Markdown files

This is the big one.

Obsidian notes are not trapped inside a weird proprietary database wearing a productivity costume. They are `.md` files sitting in folders on your computer.

This means:

- You can read them in Obsidian.

- You can open them in any text editor.

- You can back them up.

- You can sync them.

- You can search them.

- You can give an AI tool access to them later.

- You are not praying that some startup remembers to export your personality before pivoting to enterprise compliance.

Plain text is boring.

Boring is durable.

Durable is hot now. Get used to it.

2. Links create usable context

Obsidian lets you connect notes with simple links like `[[Brand Voice]]` or `[[Q3 Content Strategy]]`.

This matters because memory is not just storage.

Memory is retrieval.

A note sitting alone in the void is a shoebox. A note connected to projects, decisions, outputs, and examples becomes context an agent can use.

When your “Customer Research” note links to “Messaging Pillars,” and that links to “Homepage Rewrite,” and that links to “Proof Points,” you are not just collecting notes.

You are building a map of how your thinking connects.

AI f*cking loves maps.

So do humans who are looking to take a top down photo of their coffee and a camera with the caption “wanderlust”.

3. You can start simple and get fancy later

You do not need plugins on day one.

You do not need automation.

You do not need MCP.

You do not need a graph view that looks like a nervous system discovered astrology.

Start with folders, a few templates, and a weekly habit.

Then, once the system is useful manually, you can connect AI to it.

This order matters.

Do not automate a mess.

That is how you create a faster mess.

A mess with APIs.

the basic setup

Here is the simplest version.

Go to Obsidian and download the app.

When it opens, create a new vault.

A vault is just Obsidian’s word for “folder where your notes live.”

Name it something plain. Not “Second Brain.” Please. We are healing.

Good names:

- `Agentic OS`

- `Personal OS`

- `Working Memory`

- `AI Memory`

- `Zach OS` if you are me and apparently naming things like a software company is a medical condition

Put it somewhere you can find easily, like:

Documents/Agentic OS

Inside that vault, create this folder structure:

Agentic OS/

00 - CAPTURE/

01 - ACTIVE/

Projects/

Decisions/

Research/

People/

02 - KNOWLEDGE/

Permanent Notes/

Source Notes/

Examples/

03 - OUTPUT/

Writing/

Briefs/

Assets/

04 - SYSTEM/

Templates/

Workflows/

AI Instructions/

That is enough.

You may feel an urge to create thirty-seven folders.

I urge you to resist that urge.

Folder enthusiasm is where knowledge systems go to develop bureaucracy.

The goal is not to organize your whole life perfectly. The goal is to make the useful stuff findable by you and eventually usable by AI.

what each folder is for

Let’s make this painfully clear.

`00 - CAPTURE`

This is the inbox.

Everything starts here if you do not know where it belongs yet.

Ideas. Links. Quotes. voice-memo transcripts. half-formed thoughts. meeting residue. That one sentence you thought of while walking and immediately believed would change civilization. It probably will not, but let it have a note.

Create one note:

00 - CAPTURE/Inbox.md

Put this at the top:

markdown

# Inbox

Use this for raw capture. Do not organize while capturing.

## Quick capture format

- Idea:

- Source/context:

- Why it matters:

- Could be useful for:

- Next action, if any:

The rule: capture fast, process later.

If you try to organize every thought the moment it arrives, you will stop capturing. Your brain is not a librarian. It is a racetrack with feelings.

01 - ACTIVE

This is for anything currently alive.

Projects you are working on. Decisions you need to make. Research you are actively using. People you are collaborating with or trying to understand.

Examples:

01 - ACTIVE/Projects/Substack Growth.md

01 - ACTIVE/Projects/Homepage Rewrite.md

01 - ACTIVE/Decisions/Should we launch a paid tier.md

01 - ACTIVE/Research/Claude Skills for creators.md

01 - ACTIVE/People/Ideal Reader - AI-curious marketer.md

Active means: this note may change your work this month.

If not, it probably belongs somewhere else.

02 - KNOWLEDGE

This is where processed understanding lives.

Not raw highlights.

Not “interesting link maybe.”

Processed understanding.

A permanent note should say one clear thing in your own words.

Example:

Memory makes AI compound

AI work improves when lessons from one task are saved into reusable context for the next task. Without memory, every AI session starts from scratch, which makes the user re-explain standards, goals, examples, and decisions.

Useful for:

- [[Agentic OS]]

- [[AI Workflows]]

- [[Content Systems]]

Related:

- [[Prompts are requests. Skills are jobs.]]

- [[Feedback loops beat one-off prompting.]]

Notice the difference.

A raw note says: “Interesting article about AI memory.”

A knowledge note says: “Memory makes AI compound.”

One is a thing you found.

The other is a thing you believe.

AI can use both, but the second is much more powerful because it has already passed through your judgment.

03 - OUTPUT

This is where notes turn into things.

Essays. briefs. memos. decks. posts. launch plans. decision docs. campaign plans. scripts. product ideas.

This folder exists to remind you that the goal of a memory system is not having a beautiful vault.

The goal is output.

If your notes never become decisions, writing, strategy, or action, you do not have a second brain. You have a very aesthetic basement.

04 - SYSTEM

This is the operating layer.

This is where the magic starts.

Create these notes:

04 - SYSTEM/AI Instructions/Profile.md

04 - SYSTEM/AI Instructions/Voice and Style.md

04 - SYSTEM/AI Instructions/Current Projects.md

04 - SYSTEM/AI Instructions/Decision Rules.md

04 - SYSTEM/AI Instructions/Do Not Do.md

04 - SYSTEM/Workflows/Daily Capture Processing.md

04 - SYSTEM/Workflows/Weekly Memory Review.md

04 - SYSTEM/Templates/Project Note.md

04 - SYSTEM/Templates/Decision Note.md

04 - SYSTEM/Templates/Source Note.md

04 - SYSTEM/Templates/Writing Brief.md

This folder is your agentic OS manual.

If an AI could only read one folder to understand how to help you, it would read this one.

the five files that matter most

If the folder structure already made your eyes start looking for a nearby exit, breathe.

Start with five files.

That is the whole day-one setup.

1. Profile.md

This tells AI who you are and what you care about.

Template:

# Profile

## Who I am

- Name:

- Role:

- Company / work context:

- Main arenas:

## What I am trying to build

-

## What I care about

-

## How I like to work

-

## Things AI should know before helping me

-

Example:

# Profile

## Who I am

- Name: Zach

- Role: CMO / Head of Marketing at Averi

- Main arenas: AI content systems, startup marketing, personal writing, brand strategy, agentic workflows

## What I am trying to build

- A personal newsletter that teaches useful AI/content systems while sounding unmistakably human.

- A company narrative around Averi as the AI content engine for startups.

## What I care about

- Taste, usefulness, clarity, human voice, practical systems, ambitious people building real things.

## How I like to work

- Start with the messy truth.

- Turn it into a clear operating principle.

- Add tactical examples.

- Keep the jokes, but do not let the jokes hide the point.

## Things AI should know before helping me

- Do not make the writing corporate.

- Do not flatten my voice into generic thought leadership.

- Help me think, not just produce.

2. Voice and Style.md

This is the antidote to AI writing like it has been trapped in a webinar since 2016.

Template:

# Voice and Style

## The voice should feel like

-

## The voice should not feel like

-

## Strong lines / examples

-

## Common mistakes to avoid

-

## Formatting preferences

-

## Words and phrases to avoid

-

Fill this with real examples.

Not adjectives.

Adjectives lie.

Examples tell the truth.

Bad:

Voice: smart, witty, bold, human

Better:

Strong line: “A prompt asks for an answer. Management creates conditions for good work.”

Avoid: “In the modern digital world.” That sentence should be escorted from the building.

AI learns more from examples than vibes.

So does every human employee who has ever been told to “make it pop” and briefly considered moving to a remote island.

3. Current Projects.md

This keeps the AI from treating every task like it exists in a vacuum.

Template:

# Current Projects

## Active projects

### Project name

- Goal:

- Status:

- Why it matters:

- Current priority:

- Important links:

- Related notes:

- What AI can help with:

- What AI should not decide:

Example:

### Don’t Feed The Algorithm

- Goal: Build a personal newsletter around AI, ambition, content systems, and human taste.

- Status: Publishing weekly essays and turning them into LinkedIn posts.

- Why it matters: Builds trust, audience, and a human narrative around the future of work/Averi.

- Current priority: Deep tactical pieces that help people actually build AI systems.

- Related notes: [[Voice and Style]], [[AI Content Engine]], [[Claude Skills]]

- What AI can help with: drafts, outlines, title options, research, repurposing, QA.

- What AI should not decide: final point of view, personal stories, what gets published.

This is how you stop AI from giving technically decent advice that ignores your actual life.

Which is most advice.

4. Do Not Do.md

This file is underrated.

It is your immune system.

Template:

# Do Not Do

## Writing mistakes to avoid

-

## Strategy mistakes to avoid

-

## AI behavior to reject

-

## Claims we should not make

-

## Things that require human approval

-

Examples:

## Writing mistakes to avoid

- Do not use generic AI phrases like “unlock your potential,” “tap into the power,” or “in the modern digital world.”

- Do not turn every sentence into a standalone line.

- Do not over-explain obvious points.

## AI behavior to reject

- Do not invent sources.

- Do not pretend uncertainty is confidence.

- Do not preserve secrets, passwords, API keys, or private credentials.

- Do not send, publish, delete, buy, or contact anyone without human confirmation.

## Claims we should not make

- Do not claim a system is autonomous if a human reviews the output.

- Do not imply AI replaces strategy, taste, or accountability.

A good “Do Not Do” file will save you from yourself.

And from the AI.

Mostly from the AI.

Sometimes from yourself wearing an AI hat.

5. Weekly Memory Review.md

This is what makes the system compound.

Template:

# Weekly Memory Review

Run once per week.

## 1. Process inbox

- What should be deleted?

- What should become a permanent note?

- What belongs to an active project?

- What should become an output?

## 2. Update active projects

- What changed this week?

- What decisions were made?

- What context would help AI next time?

## 3. Save lessons learned

- What did AI get wrong repeatedly?

- What instruction would prevent that next time?

- What example should be added?

- What phrase or pattern should be banned?

## 4. Create outputs

- What note should become a post, essay, brief, memo, or action?

## 5. Archive stale notes

- What is no longer active?

This is the whole system hiding inside one habit.

Capture is easy.

Compounding requires review.

Sorry.

I did not invent entropy.

how to actually use this with AI today

There are three levels.

Start with level one. Do not skip to level three because you watched a demo and now believe your laptop should have a chief of staff.

That way lies sadness.

And YAML.

Level 1: copy/paste context manually

This is the easiest version and the one most people should start with.

When you want AI help, copy the relevant notes into the chat.

Example:

You want help drafting a newsletter.

Paste:

- `Profile.md`

- `Voice and Style.md`

- the project note for `Don’t Feed The Algorithm`

- any source notes or idea notes related to the topic

- the task you want done

Then say:

Use the context above as my working memory for this task.

Task: Help me outline a deep tactical article about using Obsidian as an agentic memory system.

Before drafting, summarize what you understand about:

1. my goal

2. my audience

3. my voice

4. the output we are creating

5. any constraints you should follow

That last step matters.

Make the AI repeat the context back before doing the work.

This catches misunderstandings early, before the machine spends 900 words confidently building the wrong little house.

Level 2: use projects / custom instructions

Most major AI tools now have some version of persistent context: Projects, custom instructions, memory, saved chats, files, knowledge bases, whatever the product team named during a meeting with too many sparkling waters.

Use your Obsidian notes as the source of truth.

Do not rewrite your entire life into every tool.

Instead, keep the canonical version in Obsidian and copy the relevant pieces into the AI tool’s project instructions.

Example:

Create a Claude Project called:

Personal Writing OS

Upload or paste:

- `Profile.md`

- `Voice and Style.md`

- `Do Not Do.md`

- `Current Projects.md`

- 3 to 5 examples of strong writing

Then update those files in Obsidian when your thinking changes.

If the Claude Project drifts, refresh it from Obsidian.

Obsidian is the source.

The AI tool is the interface.

That distinction will keep your brain from turning into soup with keyboard shortcuts.

Level 3: connect AI tools to your vault

This is the more advanced version.

You can connect some AI tools to local folders, file systems, or knowledge bases. Claude Code can work with files. MCP, the Model Context Protocol, is an emerging standard for connecting AI assistants to tools and data sources. Other tools have their own connectors.

You do not need to understand all of this on day one.

Here is the plain-English version:

MCP is like a cable that lets an AI assistant talk to other tools.

One of those tools might be your file system.

If your Obsidian vault is just Markdown files in folders, then an AI assistant with appropriate file access can read and work with those notes.

That is why plain text matters.

Your memory becomes tool-readable.

Important caveat, because I enjoy not ruining anyone’s life:

Do not give an AI tool broad file access unless you understand what it can read and change.

Start read-only if possible. Keep secrets out of the vault. Back everything up. Do not store API keys, passwords, private credentials, financial details, or anything that would make you whisper “oh no” if copied into a chat window.

The goal is useful memory, not a security incident with a cute folder structure.

the starter templates

If you want to actually build this today, steal these.

Not spiritually.

Literally.

Project note template

# {{Project Name}}

## Goal

What are we trying to accomplish?

## Why it matters

Why does this matter now?

## Current status

- Stage:

- Owner:

- Deadline / cadence:

## Context

What would someone need to know before helping?

## Important decisions

-

## Open questions

-

## Useful source notes

- [[ ]]

## What AI can help with

-

## What AI should not decide

-

## Next actions

-

Decision note template

# Decision: {{Decision}}

## Date

## Decision needed

What are we deciding?

## Options

1.

2.

3.

## Criteria

What matters most?

## Context

What facts, constraints, or prior decisions matter?

## Recommendation

What should we do and why?

## Decision

What did we choose?

## Why

Why did we choose it?

## Revisit trigger

When should we reconsider?

This is incredibly useful with AI because it prevents you from re-litigating old decisions every time your mood changes.

Which, as a creative operator, is approximately every fourteen minutes.

Source note template

# Source: {{Title}}

## Link

## What this is

Short summary in plain English.

## Key ideas

-

## Useful quotes / data

-

## My take

What do I think about this?

## Could use for

- [[ ]]

## Follow-up questions

-

The “My take” section is the important part.

Do not just save what the source said.

Save what you think it means.

That is where your taste enters the system.

Writing brief template

# Writing Brief: {{Title}}

## Working title

## Core thesis

One sentence.

## Reader

Who is this for?

## Promise

What will the reader understand or be able to do by the end?

## Why now

Why does this matter now?

## Key points

1.

2.

3.

## Examples / stories

-

## Sources / rabbit holes

-

## Voice notes

How should this feel?

## Do not do

-

## CTA / ending

This is how you make AI writing better before the draft exists.

The brief becomes the guardrail.

The draft becomes less random.

You become less likely to receive a piece that opens with “In an age where…” and forces you to stare at the wall.

what can you do with this?

Once this system exists, you can do a lot.

Not theoretical “unlock productivity” things.

Actual things.

1. Ask AI to prep you for your own work

Prompt:

Read these notes and prepare me for the project.

Tell me:

1. what matters most

2. what decisions are unresolved

3. what risks I am ignoring

4. what useful next action I should take today

5. what context should be added to memory after this session

This turns AI into a project re-entry assistant.

Extremely useful if you have more than one active thing and your brain occasionally feels like a browser with 44 tabs and one of them is playing music.

2. Turn messy capture into usable notes

Prompt:

Turn the raw notes below into:

1. one permanent note with a clear title and thesis

2. one project update if relevant

3. any decision notes that should be created

4. suggested links to existing notes

5. a list of anything that should be deleted or ignored

Do not preserve filler. Preserve only what could be useful later.

This is how you stop your inbox from becoming a landfill with Markdown syntax.

3. Build an AI briefing packet before every important task

Before a writing session, strategy session, campaign plan, or hard decision, ask AI:

Using these notes, create a briefing packet for this task.

Include:

- relevant background

- known decisions

- audience/customer context

- source notes worth using

- risks or contradictions

- recommended structure for the work

- questions I should answer before proceeding

This is one of the most underrated uses of AI.

Not doing the work.

Getting you into the work faster.

4. Keep a decision trail

Every meaningful decision should become a note.

Why?

Because future you is a liar.

Future you will say, “Why did we choose this again?” with the innocent tone of someone who was not present for the five meetings, four tradeoffs, two customer calls, and one extremely cursed spreadsheet that created the decision.

Write it down.

Then AI can remind you later.

Prompt:

Based on this conversation, create a decision note.

Include:

- what we decided

- why

- alternatives considered

- assumptions

- revisit trigger

- related notes to link

Memory is not just recall.

It is protection against your own revisionist vibes.

5. Make your AI better every week

At the end of the week, ask:

Review these recent outputs and notes.

Tell me:

1. what the AI repeatedly misunderstood

2. what instructions should be added to my system notes

3. what examples should become training material

4. what phrases or patterns should be banned

5. what workflow could be turned into a repeatable AI skill

This is the compounding loop. Every week your system gets slightly more useful.

Not because you installed fourteen plugins. Because you captured the lesson.

the minimum viable agentic OS

If you only do one version, do this.

Create these folders:

00 - CAPTURE

01 - ACTIVE

02 - KNOWLEDGE

03 - OUTPUT

04 - SYSTEM

Create these five files:

00 - CAPTURE/Inbox.md

04 - SYSTEM/AI Instructions/Profile.md

04 - SYSTEM/AI Instructions/Voice and Style.md

04 - SYSTEM/AI Instructions/Current Projects.md

04 - SYSTEM/AI Instructions/Do Not Do.md

Create one weekly recurring calendar reminder:

Friday, 30 minutes: Process AI Memory

During that review:

1. Delete junk.

2. Move useful notes to active projects or knowledge.

3. Add one lesson to `Do Not Do.md`.

4. Add one example to `Voice and Style.md` or a project note.

5. Turn one note into an output, decision, or next action.

That is enough to start.

Do not wait until the system is perfect. Perfect systems are procrastination in a nice outfit.

common mistakes

Mistake 1: capturing everything

You do not need everything. You need useful memory.

If a note will never help you make a decision, create an output, brief an AI, understand a person, or improve a workflow, delete it.

This will feel violent.

Good.

Mistake 2: making folders too clever

A simple folder you use beats a perfect taxonomy you avoid.

If you need a map to understand your map, the map has unionized against you.

Mistake 3: storing secrets

Do not put passwords, API keys, private credentials, or sensitive personal information in your AI memory vault.

Especially if you plan to connect tools to it later.

This is not paranoia.

This is adulthood with file permissions.

Mistake 4: letting AI write the memory without you

AI can help process notes.

But you need to approve what becomes memory.

Because memory shapes future output.

If AI saves the wrong lesson, it will confidently repeat the wrong lesson with the posture of a substitute teacher who found a whistle.

Human review matters.

Mistake 5: never turning notes into outputs

The point is not to have a vault.

The point is to do better work.

Every note should eventually feed one of these:

- a decision

- a project

- a piece of writing

- a workflow

- a conversation

- a strategy

- a saved instruction

- a deletion

Deletion counts. Deletion is taste wearing work boots.

where this is all going

Today, this might feel like a better notes system. But I think it is more than that.

If agents become part of how we work, then memory becomes infrastructure.

Your AI tools will get better. Models will get faster. Context windows will get bigger. Connectors will get easier. MCP and similar protocols will make it more normal for assistants to access tools, files, databases, and workflows. The interface will keep changing.

But the underlying question will stay painfully simple:

What should the AI know about you, your work, and your standards?

If that context lives nowhere, you will keep starting over.

If it lives everywhere, you will keep hunting.

If it lives in one clear memory layer, you can start building on it.

This is why I like Obsidian for this job.

Not because it is perfect.

Because it is plain.

Because it is local.

Because it is readable.

Because it lets you create a memory system that belongs to you before it belongs to a tool.

The future agentic OS probably will not be one app.

It will be a stack:

- tools that do work

- agents that coordinate work

- interfaces where you direct work

- memory that preserves context

- humans who set taste, standards, goals, and boundaries

Obsidian can be the memory layer.

Not glamorous.

Extremely useful.

The part that remembers.

And in a world where every AI can generate, summarize, draft, and pretend to be a strategist after reading three bullets, the part that remembers might become one of the most valuable parts of the whole system.

-zc

P.S. If this made you want to build your own agentic memory system and also slightly reorganize your life, I support the first impulse and recommend caution on the second.

Subscribe to Don’t Feed The Algorithm for more practical notes on AI, taste, content systems, and the deeply weird work of becoming the human operating system for a bunch of machines that never sleep.

Rabbit hole for the overcaffeinated

- Obsidian — the local-first Markdown notes app I’d use as the base layer for this setup.

- Obsidian Help: Manage vaults — plain-English docs on creating and managing vaults.

- Obsidian Help: Internal links — useful if you want your notes to connect with `[[wikilinks]]` instead of sitting alone in little note caves.

- Obsidian Help: Properties — optional metadata for people who want notes to become more structured over time.

- Model Context Protocol: Introduction — the emerging standard for connecting AI assistants to tools and data sources.

- Claude Code MCP docs — practical Claude-side docs for connecting tools/context through MCP.

- Anthropic: Building effective agents — a good reminder that simple, well-scoped systems usually beat magical overbuilt agent spaghetti.