~/brand-design capability 06 of 6 · The system, automated

Consistent at a scale no team can check by hand.

Brand AI Tools are the AI-enabled tooling that keeps a brand system consistent and efficient at scale. Models tuned on your brand, guardrails that hold, and every part of it yours.

Start an AI tools brief Watch a run

Typical length: 6–10 weeks
Tooling: Flux · Firefly · ComfyUI
Ownership: You own the model

pipeline/your-brand/prod run 0412 Illustrative

studio 2 · rec

01 Training setReference curated 2,418refs Locked · v7
02 Tuned modelImage model v4.2weights Gate passed
03 GenerationVariants in flight 146jobs Running
04 Brand checkColour · space · tone 139checked 7 flagged
05 Human reviewBrand lead signs off 12waiting People decide
06 PublishFormats × markets 54live Logged

09:41:02genvariant 0412-c · Product C · Market 03 · 4:5 · done 3.8s
09:41:03checkcolour ΔE 1.2 ✓ clear space ✓ tone 0.91 ✓ rights ✓
09:41:03queue0412-c → review · owner: Brand lead
09:41:05check0412-f · clear space 62% of rule ✕ · auto-fix proposed
09:41:06review0409-a approved by Brand lead · note attached
09:41:07pub0409-a → 6 formats × 9 markets · credentials signed

Illustration: a brand content pipeline in six stages — training set, tuned model, generation, automated brand check, human review by a brand lead, and publish — with job counts and a streaming run log.

$ brandctl --help six tools · one brand model

Six tools, one brand model

Generation starts from your identity, a check catches what slips, and a person signs off wherever the stakes are high.

USAGE brandctl <command> [flags]

COMMANDS

↑↓ select a command

brandctl tune --image --language --from ./brand-system

Brand-tuned models

Image and language models fine-tuned on your identity, so generation starts on brand instead of drifting toward it.

People decideWhich references the model may learn from, and whether a version ships.

brandctl tune --image --from ./brand-system/v7
reading 2,418 references · 311 rules · 64 examples
fine-tuning image model · epoch 12/12
eval brand fidelity 0.93 ≥ gate 0.90 · awaiting sign-off

Illustrative run

$ brandctl curate ./references phase 01 · Ground · Wk 01–02

A model learns what it is shown. People choose what it sees.

Before any tuning, the brand system becomes a training set. An agent sorts, labels and rights-checks every reference and suggests a verdict; a person approves or rejects each one. Nothing enters the set without a name against it.

batch 14 · 0 left to decide A approve R reject U undo

ref-0231Product

Studio 2 · owned

Agent: approve · palette match 0.94
ref-0232Rules

Brand system v7

Agent: approve · tokens v7 · exact
ref-0233Mood

Stock · licence ends

Agent: reject · rights unclear
ref-0234Texture

Studio 2 · owned

Agent: approve · tone match 0.91
ref-0235Rules

Archive 2019

Agent: reject · superseded identity
ref-0236Product

Archive 2023 · owned

Agent: approve · palette match 0.89
ref-0237Rules

Brand system v7

Agent: approve · type scale · exact
ref-0238Product

Agency upload

Agent: reject · near-duplicate of ref-0231

$ brandctl run --watch one brief · checked · approved · live

One brief in. Checked, approved and live in every format.

Generation starts from the tuned model with the identity locked. Every variant is scored by the brand check, only passing work reaches a person, and nothing publishes until someone approves it. Try a run, or watch one.

Interactive demo. Choose a product, market, format and mood, then press Generate. Four variants are scored for colour, clear space, tone and rights; variants that pass wait for your approval; approving one publishes it to six formats in nine markets.

Brief brand-locked

Product

Product A Product C Sub-brand B

Market

Market 01 Market 03 Market 07

Format

4:5 9:16 1:1 16:9

Mood

Calm Warm Bold

Palette v7LockupClear spaceType scaleVoice

Product C · still life · Market 03 · 4:5 · calm — locked: palette v7, lockup, clear space, type, voice

Illustrative

gen4 variants · tuned model v4.2 · seed locked · 6.1s
checkA pass 0.91 · B pass 0.94
checkC clear space 62% of rule ✕ · returned with fix
checkD blocked before render · reference rights unclear
reviewB approved by Brand lead · A held as alternate
pub54 assets · 6 formats × 9 markets · credentials signed

Variant A

Pass · 0.91

Colour
Clear space
Tone
Rights

Variant B

Pass · 0.94

Colour
Clear space
Tone
Rights

Variant C

Fail · clear space

Colour
Clear space
Tone
Rights

Variant D

Blocked · rights

Colour
Clear space
Tone
Rights

Review queue Brand lead

Only variants that pass the check reach a person. Failures go back with the reason attached.

Variant APass · 0.91
Variant BPass · 0.94
C · DReturned to generation

Publish 54 of 54 live

Variant B approved by Brand lead · published to 6 formats × 9 markets · credentials signed

$ brandctl test ./guardrails runs on every model version and every output

Guardrails written as tests, not hopes.

Before anything is generated, the rules are written as tests that can fail. When one does, you see exactly what broke, what changed to fix it and who signed the fix off. The same suite runs again on every new model version.

guardrails.spec 11 tests

11 passed (1 after a fix) · 0 failing

describe palette
- uses palette v7 tokens only · ΔE ≤ 2.012ms
- keeps text contrast at AA or above9ms
describe logo
- keeps clear space ≥ 1× mark height14ms
- never recolours or stretches the lockup8ms
describe voice
- expected headline contains 0 never-say terms
  
  received “A revolutionary way to start the day” · 1 term
```
- prompt-set v3.0 · lexicon: 41 never-say terms
+ prompt-set v3.1 · lexicon: 42 never-say terms (+ revolutionary)
+ copy assistant regenerated · eval re-run 0.91
```
  Fix Term added to the lexicon, the copy assistant regenerated and the suite run again. Signed off by the Copy lead.
- reads at tone score ≥ 0.8033ms
describe rights
- uses only references with cleared rights11ms
- blocks the likeness of a real person17ms
describe review
- sends regulated claims to a person6ms
- sends pricing and apologies to a person6ms
describe log
- records who, what, when and model version4ms

The policy the tests enforce

What the tools may produce, what needs a human, and a record of every output. Built in from the first day.

# policy.yaml · your-brand
automatic:
  - variants from an approved source
  - sizes and market versions
  - copy drafts inside the lexicon
needs_a_person:
  - regulated claims
  - people, pricing, apologies
  - anything flagged twice
never:
  - a likeness without consent
  - references with unclear rights
log: every output · who · what · when · model

Illustrative Ships as: Guardrails & eval set · Tests

$ brandctl eval --gate release every version against the same eval set

A version ships when it clears the gate and a person signs it off.

Each tuned model is scored on a fixed evaluation set built from your brand. The gate is automatic; the release is not. A version that passes on numbers can still be held on judgement.

Compare version

Gate passed

v4.2 cleared all 6 gates. Released after sign-off by the Brand lead.

Illustrative Domain 0.50–1.00 · gate marked

Brand fidelity

0.93

Eval set rated against the identity · gate 0.90
Palette accuracy

0.98

Outputs within ΔE 2.0 of palette v7 · gate 0.95
Clear space held

0.99

Lockups placed inside the rule · gate 0.98
Tone score

0.86

Copy scored against the voice · gate 0.80
Review approval

0.79

Share approved at first human review · gate 0.70
Rights blocked

1.00

Unclear references stopped before render · gate 1.00

$ brandctl inspect 0409-b content credentials · signed manifest

Every output can say where it came from.

Each published asset carries a signed record: the reference it came from, the model version, every edit, every check and the person who approved it. Open any file and read its history.

Asset: 0409-b · Product C · Market 03
Signature: Valid · your key
Disclosure: AI-generated · edited by a person
Record: 8 events · 4 people

Person Automated Illustrative

05 · Edit · a person decided

Crop and headline adjusted

A person tightened the crop and the headline. The edit is recorded, not hidden.

Who: Designer
What: Crop 4:5 safe area · kerning
When: 09-03 11:48 UTC
Hash: 5e2b…40f9

Sign-off stays with people

$ brandctl deploy --log how the programme runs · 6–10 weeks

Shipped in stages, every stage logged.

The brand system becomes the training set. Guardrails are written as tests before anything is generated.

deploy #01 · a41c9e2 queuedrunningsucceeded Wk 01–02

Ground

The brand system as training data. Assets, rules and examples curated, labelled and rights-checked.
1. agentsorted and labelled 3,112 brand assets into 14 categories
2. agentflagged 188 references with unclear rights
3. personBrand lead approved training set v7 and the tool scope
artefactsTraining setRights checkTool scope
deploy #02 · 7f03bd1 queuedrunningsucceeded Wk 03–05

Tune

Models fine-tuned and evaluated against the brand. Guardrails written as tests, not hopes.
1. agentfine-tuned the image model · 4 versions evaluated
2. agentguardrails written as 11 tests · 640 eval cases
3. personBrand lead signed off v4.2 at the release gate
artefactsTuned modelsEval setGuardrails
deploy #03 · c58e21a queuedrunningsucceeded Wk 06–08

Build

Brand check, generation and templates built into the tools your team uses. Every output logged.
1. agentbrand check shipped as a design-tool plugin and an API
2. agentgeneration pipeline wired into your asset library
3. personDesign lead approved templates and locked slots
artefactsBrand checkGeneration pipelineTemplate engine
deploy #04 · 0e9d6f4 queuedrunningsucceeded Wk 09–10

Run

A pilot team live, the model re-tuned on what they make, and the whole thing handed over with your name on it.
1. personpilot team live in Market 03
2. agentre-tuned on 1,240 approved pilot outputs
3. personhandover: weights, datasets, prompts and logs moved to your accounts
artefactsPilotRe-tuneHandover

$ brandctl registry ls deliverables · 7 artefacts

Every artefact versioned, registered and handed over.

What the programme builds lands in a registry you control, each with a format, a version and a named owner. Pull any of it without asking us.

7 of 7 match Illustrative versions

U01 released v4.2.0

Brand-tuned image model

Image model fine-tuned on your identity, so generation starts on brand.

Formats:Weightsyours

Owner
Your brand

Size
6.4 GB

brandctl pull image-model@4.2.0
U02 released v3.1.0

Brand-tuned language model or prompts

Adapter and prompt set that write in your voice and lexicon.

Formats:Weightsprompt set

Owner
Your brand

Size
380 MB

brandctl pull language-kit@3.1.0
U03 released v2.3.1

Brand check

Scores any asset against the rules and returns fixes with reasons.

Formats:PluginAPI

Owner
Your brand

Size
42 MB

brandctl pull brand-check@2.3.1
U04 released v1.8.0

Generation pipeline

Generates variants from approved sources in every format you ship.

Formats:ComfyUIAPI

Owner
Your brand

Size
120 MB

brandctl pull gen-pipeline@1.8.0
U05 released v1.4.2

Template engine

Fills templates with content and keeps layout inside the rules.

Formats:WebAPI

Owner
Your brand

Size
64 MB

brandctl pull template-engine@1.4.2
U06 released v1.1.0

Guardrails & eval set

Tests and eval set every model version must pass before release.

Formats:Tests

Owner
Your brand

Size
8 MB

brandctl pull guardrails@1.1.0
U07 streaming live

Output log & governance

Who made what, with which model, and who approved it.

Formats:Dashboard

Owner
Your brand

Size
—

Events today
18,204

brandctl pull output-log@live

$ brandctl report --outcomes what changes once it runs

Three jobs that keep running after we hand over.

The programme ends; the outcomes do not. Each one is a job that runs every day inside your tools, and each one reports back.

job/on-brand-by-defaultexit 0

On brand by default

Generation starts from your identity. The check catches what slips.
1. generation starts from tuned model v4.2
2. brand check runs on every output
3. flagged work fixed or returned with a reason
96%pass the check first time

owner · Your brandschedule · continuousIllustrative
job/volume-without-driftexit 0

Volume without drift

Nine markets, six formats, one source. The system does the repetition; people do the judgement.
1. 1 approved source in
2. 6 formats × 9 markets out
3. people review the judgement calls only
54renditions per approved source

owner · Your brandschedule · continuousIllustrative
job/yours-all-of-itexit 0

Yours, all of it

Model weights, datasets, prompts and logs belong to you. Nothing is retained, resold or trained on elsewhere.
1. weights, datasets, prompts, logs → your accounts
2. copies retained by us: 0
3. trained on elsewhere: never
0copies kept by us

owner · Your brandschedule · continuousIllustrative

$ man brandctl questions enterprise teams ask first

Read the manual before you ask.

The questions that come up in the first meeting, answered plainly. Anything else goes in the brief.

Ask something else

NAME

brand-ai-tools — AI-enabled tooling that keeps a brand system consistent and efficient at scale

SYNOPSIS

Flux, Adobe Firefly and ComfyUI pipelines for image work; Gemini, OpenAI or Anthropic models for language, chosen per task on evidence. We hold no partner badges and swap tools when the evidence changes.

You do. Weights, datasets, prompts and logs ship into your accounts. We do not retain them or train anything else on them.

You need a defined identity. If the rules do not exist yet, we build Brand Systems first, because a model can only learn what has been decided.

Guardrails written as tests, a brand check on every output, and a human review step wherever the stakes are high. Every output is logged.

The model learns from the rest. One directory up.

Brand AI Tools automate a brand that has already been decided. These are the capabilities that do the deciding, with the two this one leans on most marked.

Brand Design Back to the discipline: all six capabilities and how a programme runs cd ..
Growth Strategy Finding the next best customers and the market openings worth chasing.
Brand Identity The visual, verbal, and behavioral code that carries a brand across every touchpoint. pair · what it is tuned on
Brand Foundation The core beliefs, positioning, and principles that guide every decision that follows.
Brand Systems Rules for how a brand flexes and adapts without losing what makes it recognizable. pair · the rules it learns
Brand Architecture Structuring a brand's offerings and sub-brands so growth doesn't get messy.
Brand AI Tools AI-enabled tooling that keeps a brand system consistent and efficient at scale. you are here

Let’s build what happens next.

Tell us what you’re building. We’ll answer straight.

Book a discovery call

See the work Explore our services Read our approach Industries we serve Join the team Email us

Consistent at a scale no team can check by hand.

Six tools, one brand model

Brand-tuned models

Brand check

Asset generation

Copy assistant

Template engine

Guardrails & governance

A model learns what it is shown. People choose what it sees.

One brief in. Checked, approved and live in every format.

Brief brand-locked

Variant A

Variant B

Variant C

Variant D

Review queue Brand lead

Publish 54 of 54 live

Guardrails written as tests, not hopes.

A version ships when it clears the gate and a person signs it off.

Brand fidelity

Palette accuracy

Clear space held

Tone score

Review approval

Rights blocked

Every output can say where it came from.

Reference photographed

Approved into the training set

Model version trained

Variant generated

Crop and headline adjusted

Brand check passed

Approved for release

Published with credentials

Weights, prompts, data and logs. Yours, all of it.

Image model weights

Language adapter and prompt set

Training set

Eval set and guardrail tests

Output log

Brand check rules

Shipped in stages, every stage logged.

Ground

Tune

Build

Run

Every artefact versioned, registered and handed over.

Brand-tuned image model

Brand-tuned language model or prompts

Brand check

Generation pipeline

Template engine

Guardrails & eval set

Output log & governance

Three jobs that keep running after we hand over.

On brand by default

Volume without drift

Yours, all of it

Read the manual before you ask.

MODELS Which models do you use?

OWNERSHIP Who owns the trained model?

PREREQUISITES Do we need a brand system first?

GUARDRAILS How do you keep it from going off brand?

The model learns from the rest. One directory up.

Let’s build what happens next.