Thoughts·Apr 24, 2026
Synopsis: Your brand has a memory problem. Not a style ...
Synopsis:
Your brand has a memory problem. Not a style guide problem. Not a Figma file problem. A memory problem.
Right now, your brand’s creative intelligence is stored in three places: your best people’s heads, files nobody can find, and conversations that were never documented. When a creative director leaves, the first one walks out with them. When you switch tools, the second disappears. The third never really existed to begin with.
This is the problem a brand knowledge base with AI actually solves. Not faster content generation. Not better prompts. A persistent, living system that holds your brand’s beliefs, visual language, aesthetic instincts, and creative references in one place, maintained by an AI that compounds its understanding every time new material is added.
The concept comes from Andrej Karpathy, co-founder of OpenAI, who recently described a shift in how he uses LLMs: less time generating outputs, more time building structured, interlinked markdown wikis that an AI compiles and maintains from raw source documents. Farza extended it with a personal Wikipedia built from diary entries and notes, proving the pattern works for identity and creative history, not just technical research.
This guide extends it one step further. Your brand is a person too. It has opinions, aesthetic preferences, a history, a voice, and a visual language. It deserves its own Wikipedia. This is how to build it, as a living AI brand wiki that any agent, any new hire, and any creative brief can draw from immediately, without re-explaining who you are.
This guide builds on a concept pioneered by Andrej Karpathy. He described spending less time using LLMs to generate code and more time building personal knowledge bases: structured, interlinked markdown wikis that an LLM compiles and maintains from raw source documents.
Instead of building a traditional RAG system, Karpathy’s approach treats the LLM as a compiler. It reads raw source documents and produces a structured, interlinked wiki. Farza extended this with “Farzapedia,” a personal Wikipedia built from diary entries, Apple Notes, and iMessages, demonstrating that the pattern applies to personal identity, taste, and creative history.
This guide applies it to brand teams. The architecture is the same. The application is different. And for anyone trying to make AI actually understand their brand, it changes the game entirely.
Most brand documentation lives in PDFs that go unread, Figma files that go un-referenced, and Notion pages that go un-maintained. The problem is not the format. It is the architecture.
Static documents require humans to manually retrieve and synthesize knowledge on every single creative brief, every new campaign, every vendor onboarding call. Most people’s experience with LLMs and brand documents looks like RAG: you upload a collection of files, the LLM retrieves relevant chunks at query time, and generates an answer. This works. But the LLM is rediscovering knowledge from scratch on every question. There is no accumulation.
The brand knowledge base fixes this.
Karpathy’s approach treats the LLM as a compiler that reads raw source documents and produces a structured, interlinked wiki. The wiki itself becomes the knowledge base. No embeddings or vector search needed at the scale of a brand knowledge system.
For brands, the equivalent of “personal” is institutional: your founders’ original intention, your creative director’s instincts, your company’s evolving aesthetic point of view. All of that is currently locked in people’s heads, buried in Slack threads, or scattered across years of campaign decks. The brand wiki excavates all of it and makes it queryable, composable, and permanent.
This is how you stop briefing AI from scratch every time. This is how to make AI understand your brand at depth, not from a one-time upload, but from a living record that compounds.
Before building any individual file, understand the three-layer structure that makes this work.
Layer 1: Raw Sources (/raw) Your immutable source material. Founder interviews, campaign decks, moodboards, screenshot folders, Figma exports, brand guidelines PDFs, competitor teardowns, press coverage, client briefs. The LLM reads from here but never writes to it. This is your source of truth.
Layer 2: The Wiki (/wiki) Four LLM-generated and LLM-maintained markdown files. This is where the knowledge lives in compiled, synthesized, interlinked form. The LLM owns this layer. You read it. It writes it.
Layer 3: The Schema (CLAUDE.md or AGENTS.md) A configuration document that tells your LLM agent how your brand wiki is structured, what the conventions are, and what to do when new sources arrive. This is the key file. It transforms a generic LLM into a disciplined brand archivist.
The four core files of the brand knowledge base are:
/wiki
brand.md
design.md
taste.md
references.md
index.md
log.md
1a. Install Obsidian
Obsidian is a note-taking app that stores everything as plain Markdown files on your computer. Unlike cloud-first tools like Notion or Roam Research, your notes live locally as readable text files you own permanently.
Download Obsidian at obsidian.md. Create a new vault called brand-wiki. Inside it, create two directories: /raw and /wiki.
1b. Install the Obsidian Web Clipper
The Web Clipper browser extension converts web articles, competitor pages, and editorial references directly into markdown files. This is how your raw layer stays fed without manual copy-paste work. Clip everything: competitor landing pages, design award winners, editorial pieces about brands you admire, articles that articulate the cultural moment your brand lives in.
1c. Set Up Your Agent
Open Claude Code, Cursor, or your agent of choice and point it at your vault directory. Create a file called CLAUDE.md at the root of your vault. This will be your schema file. Leave it empty for now. You will fill it as you build.
1d. Initialize the Index and Log
Create two additional files in /wiki:
The outputs of queries get filed back into the wiki to enhance it for further queries. Explorations always add up. The log is how you track what has been added and when.
What it is: The soul document. The compiled answer to: What does this brand fundamentally believe, and why does it exist?
What goes into raw sources first:
Drop these into /raw/brand/:
What the LLM should compile into brand.md:
Tell your agent:
“Read everything in /raw/brand/ and compile brand.md. This file should capture: the founding intent and origin story in plain language; the core beliefs the brand holds about the world; the brand’s relationship with its customer; the cultural context the brand was born into; the tensions the brand holds (e.g. premium but accessible, minimal but warm); and the questions the brand is still actively working through. Cross-link to taste.md and design.md wherever a belief manifests visually or aesthetically.”
What brand.md should look like:
# brand.md
## Origin
[[Founder Name]] started [Brand] in [year] because…
→ See also: [[references.md#founding-era-inspo]]
## Core Beliefs
– The world has too much X and not enough Y
– Our customer is not buying a product; they are buying…
– We believe design is not decoration, it is…
## The Brand’s Tensions
– Premium vs. Approachable
– Minimal vs. Human
– Confident vs. Humble
## Open Questions
– Are we a lifestyle brand or a utility brand?
– How do we express [value] as we scale?
## Backlinks
[[design.md]] [[taste.md]] [[references.md]]
Every claim in brand.md must be traceable to a raw source. Separate facts, inferences, and open questions explicitly. If something is a belief held by the founder, cite the interview. If it is an inference from multiple sources, label it as such.
What it is: The system document. The compiled answer to: How does this brand look, feel, and behave across surfaces?
This is not a traditional style guide. It is a living system that captures not just the rules but the reasoning behind them, and that updates itself as the brand evolves. Think of it as a brand guidelines AI system, one that grows smarter every time a new decision is made.
What goes into raw sources first:
Drop these into /raw/design/:
What the LLM should compile into design.md:
Tell your agent:
“Read everything in /raw/design/ and compile design.md. This file should document: typography rules with the reasoning behind typeface choices; color system with emotional intent behind each color decision; spatial philosophy; image treatment standards; voice in visual language; surface-specific rules. Where design decisions express brand beliefs, backlink to brand.md. Where design reflects aesthetic taste, backlink to taste.md.”
What design.md should look like:
# design.md
## Typography
– Primary: [Typeface] — chosen because it reads as [quality]
without performing [undesired quality]
– Rule: Never set body copy below 16px. Never.
→ See also: [[references.md#typography-references]]
## Color System
– Primary: [HEX] — the brand’s resting state
– Accent: [HEX] — used only for moments of action or emphasis
– Background: Never pure white. Always [HEX].
Reasoning: [[brand.md#warmth-over-precision]]
## Spatial Philosophy
– We breathe. Density signals urgency; we are not urgent.
– Minimum margin: [X]px at all breakpoints.
## Image Treatment
– Photography: Natural light, slightly underexposed
– No stock photos. Ever.
– Illustration style: [[references.md#illustration-direction]]
## Open Design Questions
– How do we handle dark mode without losing warmth?
## Backlinks
[[brand.md]] [[taste.md]] [[references.md]]
The key shift from a static style guide: every time a new design decision is made, drop the relevant screenshot or exported notes into /raw/design/ and tell the agent to ingest it. The design doc does not go stale because the LLM is always catching it up.
What it is: The instinct document. The most underrated file in the whole system. This is where the brand’s aesthetic point of view lives, not what it has built, but what it finds beautiful, interesting, and worth paying attention to.
Farza built something like this for himself. His wiki captured his favourite films and their impact on him, his philosophical notes from creative documentaries, his aesthetic inspirations. For a brand, this is the equivalent of the creative director’s notebook.
What goes into raw sources first:
Drop these into /raw/taste/:
What the LLM should compile into taste.md:
Tell your agent:
“Read everything in /raw/taste/ and compile taste.md. This file should articulate: the aesthetic categories this brand finds beautiful and why; the cultural references that recur across the team’s inspiration; the things the brand actively dislikes and considers off-brand; the decade or design era the brand unconsciously gravitates toward; the tension between aspiration and differentiation. This file should read like a manifesto of aesthetic belief, not a list of rules.”
What taste.md should look like:
# taste.md
## What We Find Beautiful
– Restraint that reads as confidence, not timidity
– Objects that show their making
– Typography that has opinions
– Silence as a design element
→ Recurring reference: [[references.md#dieter-rams]]
→ Recurring reference: [[references.md#japanese-minimalism]]
## What We Actively Avoid
– Gradients that try too hard
– Irony without warmth
– Density mistaken for richness
– Trends mistaken for identity
## Our Aesthetic Era
Gravitates toward late 1970s to early 1990s design:
the moment before digital made everything possible
and nothing felt considered.
## Aspiration vs. Differentiation
– Looks like: [Brand X] — use of type, spatial restraint
– Must not look like: [Brand Y] — too cold, too corporate
## Backlinks
[[design.md]] [[references.md]] [[brand.md]]
Why this file matters so much for AI output quality: when you ask an LLM to generate copy, design briefs, or creative direction, the difference between generic output and genuinely on-brand output is almost entirely determined by how well the model understands taste. A taste document gives the model the aesthetic vocabulary and the reasoning behind preferences. Not just rules to follow. A point of view to channel.
This is how you stop AI from getting your brand wrong. You give it taste, not just rules.
What it is: The media library. Every image, film, album, brand, book, campaign, or artifact that has shaped the brand’s visual and creative language, catalogued, described, and cross-linked to the other three files.
Think of it as a fan wiki for the brand’s creative influences.
What goes into raw sources first:
Drop these into /raw/references/:
What the LLM should compile into references.md:
Tell your agent:
“Read and view everything in /raw/references/ and compile references.md. For each reference, create a named section with: a one-paragraph description of what it is; why it matters to this brand specifically; which qualities are being cited; and which other wiki files this reference informs. Organize sections by category: visual brands, films, photography, typography, spatial, music and sound. The goal is that any agent reading this file should understand not just what a reference is, but how to use it as a creative input.”
What references.md should look like:
# references.md
## Visual Brands
### Braun (Dieter Rams era)
Category: Product design, visual language
Why it matters to us: The idea that good design is as
little design as possible. Not minimalism for aesthetics,
but function so pure it becomes beautiful.
What we cite it for: Spatial restraint, typeface choices,
the relationship between utility and elegance.
→ Informs: [[design.md#spatial-philosophy]]
→ Informs: [[taste.md#what-we-find-beautiful]]
## Films
### [Film Title]
Category: Cinematography / Color
Why it matters: The way [director] uses underexposure
to create intimacy without sentimentality.
→ Informs: [[design.md#image-treatment]]
## Typography
### [Typeface or Typographic Work]
Category: Type reference
Why it matters: Proof that [quality] and [quality]
can coexist in the same letterform.
→ Informs: [[design.md#typography]]
The four files are not four separate documents. They are four nodes in a graph, and the graph is what makes the whole system intelligent.
The defining feature of this approach is bidirectional linking. Connect ideas with [[wiki-style links]] and the system shows you how concepts relate in a visual graph. In Obsidian, press Cmd+G to open the Graph View. You will see your notes as dots and links as lines between them. As you add more notes and links, this becomes a visual map of your brand’s knowledge.
Here is how the four files connect:
brand.md ─────────────────────────────────────────
│ │
│ “our belief in warmth” “our founding in X culture”
↓ ↓
design.md ←─────────────────────────── taste.md
│ “never pure white, │ “we gravitate toward
│ see brand/warmth” │ the late 70s-90s seam”
│ │
└────────────────────────────────────┘
Both point to:
↓
references.md
“Braun (Dieter Rams era)”
“Film X — underexposure”
“Typeface Y — warmth + rigor”
The schema file (CLAUDE.md) should explicitly instruct your agent to maintain this graph:
# CLAUDE.md — Brand Wiki Schema
## Purpose
You are the archivist and editor of [Brand Name]’s brand wiki.
Your job is to compile, maintain, and cross-reference four core files:
brand.md, design.md, taste.md, and references.md.
## Rules
in brand.md or a reference in references.md.
specific reference in references.md.
and append to log.md.
new page if it constitutes reusable knowledge.
## Workflow: Ingest
## Workflow: Query
Ingest (ongoing)
Every time something happens that is relevant to the brand’s creative identity, a new campaign launches, a creative director gives an interview, a competitor does something interesting, a team member clips a piece of editorial they love, drop it into the appropriate /raw/ subfolder and tell the agent: “Ingest this and update the relevant wiki pages.”
A single source might touch multiple files. A new campaign that just launched might update design.md with how new surfaces were handled, taste.md if the campaign made a new aesthetic choice, references.md if it cited new visual references, and brand.md if the campaign crystallized a belief that was previously implicit.
Query (the payoff)
Once the wiki is populated, this is where the value becomes obvious. Tell your agent:
By building a persistent, LLM-maintained record, the wiki solves the core frustration of stateless AI: the context-limit reset. You never have to re-explain the brand to an agent. You just point it at the index.
Lint (periodically)
Run LLM health checks over the wiki to find inconsistent data, flag missing connections, and identify candidates for new article creation. For a brand knowledge base, a lint pass might find:
Schedule a lint pass quarterly, or whenever a major creative shift happens.
Part 1 gave you the four files. This part makes them indestructible.
Most brand wikis die in week three.
Not because the idea is bad. Because the idea was built for one person.
One creative director. One founder. One brand strategist who had the weekend to set it up and the discipline to keep feeding it.
Then Monday happens. The agency sends seventeen files. Slack fills up. Three people are editing the same brief. The new hire needs onboarding. The campaign goes live and nobody documents what worked. The wiki sits there. Empty. Waiting.
This is the gap between the concept and the company.
Karpathy built his system for one mind. This part solves the team problem.
Your brand’s knowledge is currently stored in three places:
When your best people leave, the first one goes with them. When you switch tools, the second disappears. The third never existed to begin with.
The brand wiki does not just organize what you have. It exteriorizes what is currently locked inside people and makes it survive them.
That is the real promise. Not productivity. Institutional continuity.
brand-wiki/
│
├── CLAUDE.md ← The schema. The brain’s instruction manual.
│
├── raw/ ← Immutable. LLM reads, never writes.
│ ├── brand/
│ ├── design/
│ ├── taste/
│ ├── references/
│ ├── campaigns/
│ ├── competitors/
│ ├── voice/
│ └── assets/
│
├── wiki/ ← LLM-owned. You read; it writes.
│ ├── index.md
│ ├── log.md
│ ├── brand.md
│ ├── design.md
│ ├── taste.md
│ ├── references.md
│ ├── voice.md
│ ├── campaigns/
│ └── competitors/
│
└── outputs/
├── briefs/
├── analyses/
└── presentations/
Notice what is not here: a database, a plugin, an account, a subscription. Just folders. Just files. Just text.
Git-compatible. Obsidian-compatible. Cursor-compatible. Claude Code-compatible. Everything-compatible.
The voice.md file is an addition worth noting. It is the file that makes copy actually sound like you, and the most frequently used file for cross-team deployment.
A three-person D2C skincare brand does not have the same problem as a 400-person SaaS company. Both can use this system. Neither should use it the same way.
D2C (1-15 people): The Founder-Dependent Brand
The problem: everything lives in one person’s head.
The wiki solves this by externalizing the founder’s taste before it becomes a bottleneck. Sit the founder down. Record everything. Two hours minimum. Ask:
Transcribe it. Drop it into raw/brand/founder-interview-01.md. Drop saved images into raw/taste/founder-phone-saves/. Tell the agent: “Ingest everything in raw/. Compile brand.md, taste.md, and references.md. Every claim in brand.md must link to the interview transcript.”
Now when the new hire needs to brief the agency, they say: “Read taste.md and references.md and write a creative brief for a winter shoot.” The agent produces something on-brand. Not because it is magical. Because the founder’s taste is now a file.
Midmarket (15-150 people): The Silo Problem
The problem: sales says one thing, design does another, product does a third.
This is where voice.md earns its keep. A vocabulary table, sentence rules, and anti-patterns with real examples. Run a monthly consistency check: tell the agent to read voice.md, then read the last 30 social posts and 10 nurture emails, flag any copy using banned vocabulary, and output as a table with location, violation, and suggested fix.
This is your brand’s spell-checker. Except it checks for soul, not spelling.
Enterprise (150+ people): The Brand Drift Problem
The problem: the parent brand and sub-brands have nothing in common.
The solution is a group-level wiki with individual brand sub-wikis underneath. One shared group-brand.md that defines the non-negotiables across all brands, and what is intentionally different per brand. Six wikis. One query. Institutional intelligence that no individual brand manager has access to alone.
Regardless of company size, these are the five files that determine whether the wiki survives past week three.
The critical line most people miss:
## What To Do When You’re Unsure
If a new source contradicts an existing wiki claim:
– DO NOT resolve the contradiction automatically.
– Add a [CONFLICT] note to log.md.
– Add the contradictory claim to the relevant wiki file
under a ## Contradictions or Tensions section.
– The human decides which claim takes precedence.
The agent is the writer. You are the editor-in-chief. The schema makes this explicit.
## [2026-04-01] ingest | Founder Interview 02
Contributor: Sarah M. | Team: Brand
Updated: brand.md, taste.md
Conflicts flagged: None
## [2026-04-03] query | Photography Brief SS26
Answer filed to: outputs/briefs/ss26-photo-brief.md
Wiki pages read: taste.md, references.md, design.md
## [2026-04-05] lint | Monthly Health Check
Findings:
– 3 orphan pages in references.md
– 1 conflict: voice.md uses “innovative”;
brand.md says never use
[CONFLICT] pending human review
The log is also how you prove the wiki is working. When leadership asks how much this actually gets used, you open log.md and show them 140 queries in Q1.
The agent reads this first on every session. It is the table of contents for the entire company’s brand brain. Keep it current. The agent updates it automatically on every ingest.
Most brands cannot answer the question: “What have we actually learned about what works for us?” Not because they did not try things. Because nobody wrote down why it worked.
A campaigns file for each launch, linked back to brand.md, design.md, taste.md, and voice.md. Performance results filed alongside the creative rationale. What worked and why. What failed and what it contradicted in the wiki.
Now when the new creative director asks what has worked before, the agent does not say “check the drive.” It reads across every campaign file and synthesizes an answer.
Most competitive analysis disappears into a deck, gets presented once, and is never looked at again. A structured competitor file for each relevant brand, annotated with what the team thinks works, what does not, and where the opportunity lies.
Most people use the wiki for briefs. That is correct. But the highest-value queries are the ones nobody thinks to ask.
For the creative team:
“Read taste.md and our last six campaign post-mortems. What visual choices have we made repeatedly that contradict our stated taste? Where are we drifting from ourselves?”
For the brand strategy team:
“Read brand.md and all competitor profiles. Where are we the only brand making a specific claim? Where are we making the same claim as everyone else?”
For the executive team:
“Read all campaign post-mortems from the last 18 months. Extract every instance where performance data contradicted a brand belief stated in brand.md. List them chronologically. What pattern emerges?”
For onboarding:
“A new senior designer is joining next Monday. Read all five core wiki files and write a 1,200-word brand onboarding document in plain language. Assume they know design but do not know us. Include the three things that most surprise people about how we approach brand.”
That last one is worth sitting with.
Right now your onboarding process is either a 200-slide deck or a two-hour call with whoever has time. Neither transmits the institutional knowledge that makes someone good at your brand fast.
The wiki writes the onboarding document. Every time. From the current state of everything you know. Automatically updated every time the wiki is updated.
Signs: every raw source gets a summary page, no synthesis, no backlinks, references.md has 200 entries with no connections. The schema must require synthesis, not just summary. Every ingest must update at least two files.
Signs: voice.md says no adjectives. The last six campaign briefs are full of adjectives. Nobody noticed. The lint operation must include a cross-file contradiction check. The [CONFLICT] tag in log.md must be reviewed by a human.
Signs: the team is debating Obsidian vs. Notion. Someone built a beautiful Dataview dashboard. The wiki has twelve files. None of them are current. Karpathy was clear. The system is a directory of markdown files and a schema. Choose your tool in five minutes and spend the rest of the time putting things in it.
Signs: brand.md states a belief that sounds right but nobody can point to where it came from. Every claim in brand.md and voice.md must link to a raw source. The agent must separate facts from inferences and label both explicitly. The human is always the editor-in-chief.
This is the most important thing to understand about the system, and the thing that makes it categorically different from a static brand guide.
One campaign gives you results. Ten campaigns give you patterns. Fifty campaigns give you operating intelligence.
Year 0: brand knowledge trapped in three people’s heads. Onboarding time: three to six months. Creative review cycles: three to four per project.
Year 1 with the wiki running: brand knowledge externalized, searchable, auditable. Onboarding time: three to four weeks, the agent writes the onboarding document. Creative review cycles: one to two per project because the brief was better.
Year 2 with the wiki compounding: every campaign post-mortem feeds the next brief. Every successful execution becomes a reference. Every failed test becomes a named anti-pattern. Every new hire contributes to raw/ immediately.
Year 3: the wiki knows more about what works for your brand than any individual employee. And it never leaves.
This is what people mean when they talk about a brand second brain. Not a note-taking tool. A system that holds institutional taste and never loses it.
The knowledge does not walk out the door when the creative director takes another job. It does not get lost when you switch agencies. It does not sit in a PDF that nobody reads.
It lives in a folder of text files on your server. Maintained by an agent. Readable by anyone. Queryable in plain English.
The entire system is just files. No proprietary schema, no vector embeddings, no subscriptions. Just text a human can open and read in any editor.
The mistake everyone makes: starting with the system instead of the content.
Do not set up Obsidian first. Do not write the perfect schema first. Do not design the folder structure for two hours.
Do this instead:
Day 1: Create three folders. (2 minutes)
Day 1: Drop everything you have into raw/. (30 minutes)
Whatever you have. Messy. Unorganized.
Do not curate. Just drop.
Day 2: Write a one-page CLAUDE.md. (20 minutes)
What is this wiki for?
What are the four core files?
What is the one rule you care most about?
Day 2: Tell the agent to compile the wiki.
(15 minutes active, then step away)
Day 3: Ask it one real question.
File the answer back into outputs/.
Day 4: Notice something the wiki got wrong or missed.
Add a raw source. Reingest. Watch it update.
Week 3: Run a lint pass.
Month 2: Your new hire uses the wiki to brief the agency.
Without asking you.
That is the whole plan.
The wiki does not need to be complete to be useful. It needs to be started.
The difference between the people who bookmarked Karpathy’s post and the ones actually using a wiki is not intelligence or resources. It is the decision to create the first folder instead of thinking about creating the first folder.
Your brand’s brain exists already. It is just scattered.
The wiki collects it. The agent organizes it. Time compounds it.
Start today. The folder takes thirty seconds to create.
A brand knowledge base with AI is a structured, LLM-maintained collection of markdown files that holds your brand’s beliefs, design system, aesthetic instincts, and creative references in one place. Unlike a static style guide or a PDF brand document, it is a living system. Every time new material is added, the AI updates, cross-links, and deepens the existing knowledge. The result is a system that gets smarter about your brand over time, without anyone having to re-explain who you are.
When you upload a brand guide to ChatGPT, the session ends and the context disappears. The next conversation starts from scratch. A brand knowledge base with AI works differently. The knowledge is compiled, structured, and stored in files that persist between sessions. Any agent you point at the wiki has immediate access to everything your brand has learned, without re-uploading, re-explaining, or re-prompting. The accumulation is the point.
An AI brand wiki is the wiki layer of this system: the four core markdown files (brand.md, design.md, taste.md, references.md) that the LLM compiles and maintains from your raw source material. Any brand team dealing with high creative volume, frequent agency handoffs, new hire onboarding friction, or AI tools that keep producing off-brand outputs needs one. The problem it solves is not speed. It is institutional memory.
Yes, with the right schema. The LLM does not make autonomous decisions about what the brand believes. It compiles, synthesizes, and cross-links what your raw sources already say. The schema (CLAUDE.md) defines the rules: every claim must link to a source, every design decision must backlink to a belief, every contradiction must be flagged rather than resolved automatically. The human remains the editor-in-chief. The AI handles the filing, linking, and synthesis.
“Second brain” is the personal productivity framing, popularized by Tiago Forte and extended by Karpathy and Farza for individual knowledge workers. A brand knowledge base applies the same architecture to institutional knowledge: the collective memory of a brand team rather than one person. The mechanics are identical. The scope is organizational. And because multiple people contribute to and draw from it, the governance layer (the schema, the log, the lint cadence) becomes more important.
Most LLM interactions with brand material fail because the model has no context between sessions and no structured way to access institutional knowledge. The LLM brand knowledge base solves this by giving every agent a persistent, structured entry point into the brand’s full history. The index.md file acts as the table of contents. The agent reads it first, identifies which files are relevant to the query, reads those files in full, and synthesizes an answer with citations back to raw sources. The quality of the output is directly proportional to the quality of the wiki.
A functional first version can be built in two days. Day one: drop everything you have into the raw/ directory and write a basic CLAUDE.md schema. Day two: run the first compilation and ask the agent one real question. The wiki will be incomplete. It will also be immediately useful. From that point, every ingest makes it better. The system is designed to compound, not to be completed.
Nothing. That is the point. The knowledge is in the files, not in the people. A creative director’s taste, a founder’s instincts, a brand strategist’s competitive analysis: all of it is compiled into the wiki before it walks out the door. This is what makes institutional continuity possible at the level of brand identity, which is traditionally the most fragile kind of knowledge to preserve.