adaptius
Explore
Talk with us
how the brain is built

A Brand Brain is built on useful context, repeatable ways of working, and a tested way to learn from experience.

And nothing ships without your go. It starts from Andrej Karpathy's LLM Wiki pattern: raw sources nobody edits, a Brain the AI keeps current, and one set of rules that people and AI both read. Here are the twelve principles behind it in four parts, each with a source you can open and the place it lives in the Brain.

Twelve cream rule cards arranged on an aqua sheet beside working drafts and a phone on a sunlit wooden desk.

the standards we build on

No vendor certifies Brand Brains. We build ours from published guidance and research, and label every source by tier, because a source you cannot weigh is a source you cannot check. Read the headings for the principles, the small print for the mechanism.

verified 2026-09-29 · reverify by 2026-12-28

useful context The right truth, in the right amount, at the right moment.

01

Raw sources, a compiled Brain, one set of rules.

the LLM Wiki pattern

Your decks, research, and guidelines go in as they are and are never edited. The Brain is what gets built from them, and one rulebook tells people and AI how to use it.

In the Brain: an inbox of sources, the Brain in numbered chapters, one contract, a catalog of every file, and a journal of every change. Karpathy's three operations run it, ingest, query, and lint, plus one of ours: retire.

Source: Karpathy, LLM Wiki, April 2026 published pattern

02

Context is a budget, not a bucket.

context engineering

The AI gets the smallest set of truth that fully answers the job, never the whole archive, and deep material waits until a task calls for it. Past a point, more text makes a model worse, not better.

In the Brain: a resident core capped at 4,000 characters, a routing table from task to chapter, headings that name what they hold, and size limits that fail the check when crossed.

Sources: Anthropic, effective context engineering, 2025 first-party Anthropic, Agent Skills best practices first-party Liu et al., Lost in the Middle, TACL 2024 peer-reviewed Packer et al., MemGPT, 2023 paper

04

Every fact wears a date, and stale ones retire.

temporal memory and pruning

Sources and changes carry their dates, so a stale fact gets caught instead of trusted. Removal is an operation with rules, because a Brain that only grows gets dumber.

In the Brain: dated sources, an append-only journal, reverify dates the check enforces, and a retire operation that finds every place an old idea still lives.

Sources: Wu et al., LongMemEval, ICLR 2025 peer-reviewed Anthropic, context editing first-party

repeatable ways of working Your best people's methods, followed the same way by every AI you use.

05

Your methods run as skills, not as memory.

agent skills

How your best people write, review, and sell is written down once as a skill, so every AI you install the Brain on follows the same method instead of improvising.

In the Brain: skills for writing on brand, reviewing work, making media, and enabling a sale, each loading only when the task calls for it.

Sources: Anthropic, Equipping agents for the real world with Agent Skills, 2025 first-party Anthropic, Agent Skills best practices first-party

07

Every rule travels with an example.

rules paired with exemplars

"Be confident" means little to a model. An approved example of confident done right means a lot. Style is too implicit to specify by instruction alone.

In the Brain: the voice law, paired with a library of approved work that grows as signed-off pieces are kept as examples.

Sources: Anthropic, use examples effectively first-party Wang et al., Findings of EMNLP 2025 peer-reviewed

a tested way to learn from experience Every correction makes the Brain smarter, and every rule change has to prove it.

09

Rules change on evidence, not argument.

evals as the gate

A correction fixes the piece, then asks what rule made it right. One round of feedback is a preference; two in the same direction make a rule. A new rule has to pass a fixed test, not win a meeting.

In the Brain: 80 graded test cases, and a first bar being recorded now that every rule change will be checked against. The graded run is manual for now.

Sources: Husain, Your AI Product Needs Evals, 2024 practitioner Anthropic, Building effective agents, 2024 first-party

nothing ships without your go Your brand changes only when you say so.

11

A control beats a promise.

least privilege

An AI saying it will not send an email is a promise. A mail account that cannot send is a control. Where a limit can be set at the provider, it is.

In the Brain: every tool the Brain can reach has a record naming its account and its limits. No record, no capability.

Sources: Willison, The lethal trifecta for AI agents, 2025 practitioner OWASP, LLM01:2025, least privilege standard

why small

Small on purpose, and organized so nothing gets lost.

500 Give the best models 500 rules to hold at once, and they follow about two-thirds of them. A brand is that hard case: many rules at once, all binding. IFScale, arXiv 2507.11538, 2025

That figure is about instruction load, not length. Separate work shows the other half: a model’s usable context is shorter than its advertised one, accuracy falls as input grows, and models recall the middle of a long context least reliably. Small keeps every rule live.

Small is not the point. Small and organized is. A merely squeezed corpus performs worse, and where a big context genuinely wins, single-fact lookup, we happily let your AI use it. The Brain is designed for the hard case: many connected rules applied together.

Sources: RULER, COLM 2024, on effective versus advertised context Lost in the Middle, TACL 2024 Chroma, context rot, 2025

The question you are thinking

“Why is this better than a Claude Project with my brand guide and a prompt that says be my best brand manager?”

Try it. Load your guide into a Project, and day one looks similar. We mean that.

The difference is every day after. A pasted guide does not by itself check an off-brand draft against a named rule, run your methodology as skills, separate an approved claim from a hopeful one, or record who changed what. “Be my best brand manager” leaves too much implicit, which is why ours ships as rules, approved examples, and checks.

A Project is a copy of your brand. The Brain is the source the copies come from, and it only changes when you approve the change. You can load it into as many Projects as you like. The longer answer is in the FAQ.

still earning it

Three principles are not finished, and we say so.

Every fact wears a date: sources and the journal are dated, but single claims do not yet carry their own recheck date.

Every rule travels with an example: the approved library started in August 2026 and is still young.

Rules change on evidence: the 80 test cases were written in September, and the first bar is being recorded now. Until it is, the gate cannot fail. The graded run is manual today; making it automatic on every rule change is the next step.

A page that names its gaps is a page you can check.

How often

Dated, and re-verified

Best practice has a shelf life, so this page carries a date and a due date. Verified 2026-09-29, when this version reorganized the principles around context, ways of working, learning, and your go, sourced to Karpathy's 2025 talk, Anthropic's agent patterns, Hamel Husain, Simon Willison, and OWASP. Every citation is re-opened on each pass rather than assumed: on 2026-08-28, two sources did not survive that check and were replaced, which the Brain's journal records. It falls due 2026-12-28. A page past its reverify date is stale, and the Brain's own lint says so before a reader has to. Sources move monthly; ninety days is the honest shelf life.

the next step

Where next

Brand Brain

A Brand Brain is your brand's approved judgment, built into a working system your people and your AI can use: strategy, design system, skills, connectors.

Founders

Adaptius was founded by Kip Botirius and Kevin Reilly: one leader who knows how brands grow, one who knows how products get built in the AI era. Who they are, why Adaptius exists, and how they work.