Role scoped
Staff, counsel, agencies, partners and licensees see only the brands, markets and notes they are permitted to see.
The prelight differencea record, not another prompt
Prompting your brand works once. A record works every time.
Most teams give AI their brand by hand: a brand book pasted into a chat, a few files attached, a prompt someone wrote last month. Prelight Atlas turns your assets, guidelines and licenses into one governed record that your team, agencies, licensees and every AI tool they use can read from.
The same brand, read two wayswithout a record · with Atlas
A prompt carries a task. A record carries the brand.
QuestionWithout a recordWith Prelight Atlas
Where the brand lives
A brand book, a DAM, shared drives, legal’s inbox and whatever was pasted into a chat.
One record. Every asset described and grouped, with the guideline and license that apply.
Who works from it
Whoever has the latest prompt or the right download.
Your team, agencies, partners, licensees and their AI tools, each scoped to what they may see.
How AI finds an asset
Files attached one at a time, until the context window or budget runs out.
It searches descriptions, previews a thumbnail and fetches only the original it will use.
Whether you may use it
Someone emails legal and waits.
Territory, window, channel and version are checked, with the clause cited.
What happens tomorrow
Someone rewrites the prompt, slightly differently, and pays again.
The record keeps every correction and decision, and answers better with use.
If you change tools
Start again.
Connect the new tool to the same record.
One record for everyonepeople and agents read the same source
One record for everyone
Your team, your partners and every AI tool they use.
A brand is used by more people than work for it. They should all be reading the same thing.
Your brand is read by in-house teams, agencies, partners, licensees and every AI assistant any of them opens. Today each keeps a private copy. Copies drift, and nobody can tell which one an output came from.
Atlas brings one record into the tools they already use. An agency in Cursor, a licensee on a free login and your team in Claude ask the same record and get the same answer, with the source it came from. Only their access changes.
Rule · mascot pose · brand-guide.pdf p.13The mascot always faces left. A mirrored pose needs prior approval.
Your team · Atlas
What can we use for EU plush this season?
Only cleared assets appear. Anything lapsed is flagged.
AI agent · MCP
Make the hero image for the German launch.
Uses the approved left-facing pose. EU window valid. Cites p.13.
Licensee · Spain
¿Puedo usar la pose reflejada?
No. Una pose reflejada necesita aprobación previa.
Agency · Cursor
Send me everything cleared for this launch.
Returns the cleared set for the agency’s markets, and nothing outside them.
Illustrative · one rule · four readers · one source
Built for media librariesread once · retrieve precisely
Built for media libraries
AI cannot read ten thousand media files every time it needs one.
Pointing an agent at a drive is slow, unreliable and expensive enough to end the project.
Models pay in tokens for everything they inspect. A library of high-resolution images, videos, PDFs and fonts fills the context window before the useful work begins. A query that returns 10,000 loose files does not work.
Atlas does the expensive reading once. It describes, groups and indexes each asset. An agent searches those descriptions, checks the details, previews a thumbnail and fetches the full-resolution original only when it needs that file.
Without a record
Open, look, discard. Repeat.
10,000 originals → context full
With Atlas
Search, preview, fetch one.
descriptions → one original
A record that growsevery decision stays useful
A record that grows
A prompt is forgotten by tomorrow. The record remembers.
Every correction, clearance and question makes the next answer better, for every reader.
A prompt lives inside one task. The next person starts again, pays again and may reach a different answer. Atlas keeps the work: the human correction, counsel’s clearance, the agent’s question and the exception that was raised rather than guessed.
That decision trail is shared by people and agents. The record becomes more precise with use without turning a chat history into company policy.
Decision trailhuman or agent · the entry looks the same
CorrectionSocial team: “hero_v3 is the German cut, not the French.” Applied to four grouped assets.
ClearanceCounsel confirmed the UK window under LIC-0007. Eleven assets cleared.
QueryAgent asked for “assets for EU plush”. Answered with sources cited.
RaisedThe French window for summer-hero.mp4 has lapsed. Raised for review, not guessed.
Rights checked before it shipscitations included
Rights checked before it ships
If it isn’t cleared, it doesn’t ship.
Your licenses sit next to your assets, so every answer already knows what you are allowed to do.
Brand assets come with conditions: territory, time window, channel and version. Those conditions usually live in contracts and legal’s inbox, away from the moment content is made.
Atlas reads licenses alongside assets. Every asset is checked on its way out. Rights answers come from your documents and cite the clause; a conflict is raised to a person rather than guessed.
Outbound check · summer-hero-v2.mp4paid social · Germany · November
Requested asset checked against the record
TerritoryDE is in the license
WindowValid to 31 Dec 2027
ChannelPaid social is permitted
Versionv2 is the approved cut
Cleared.Cites LIC-0007 §4.2 · brand-guide.pdf p.13
Scoped, traceable and openno lock-in
Scoped access, full provenance, and nothing used to train models.
The same record can safely serve people and agents because access is enforced at the record, not entrusted to a prompt.
Same record · four viewsenforced in the app and over MCP
Your teamEvery brand and asset they work on, with licenses.
CounselEverything, plus private notes no other view shows.
An agencyOnly the brands and markets it is engaged on.
A licenseeSix of ten assets, scoped by LIC-0007, United Kingdom.
2026-10-01 09:14 · agent (claude) · get_file · scoped by LIC-0007 · allowed
Every action traceable
Reads, writes and clearances carry the reader, rule and date, whether a person or an agent did the work.
Never training data
Not by Prelight, and not by the models called on your behalf. It is a contract term, not a switch.
Open MCP endpoint
Use Claude, Cursor, Lovable, n8n and the next compatible tool your team adopts through OAuth 2.0.
Nothing migrates
Atlas reads from the DAM and drives you already have. Nothing gets renamed, and the source stays yours.
The record leaves with you
Export assets, descriptions, rights and the decision trail in an open format, followed by hard deletion on request.
The standards underneathnone of them ours
MCPHow AI tools connect: Claude, Cursor, Lovable, n8n and the next tool your team adopts.
OAuth 2.0How access is granted, scoped and revoked.
ODRL · W3CHow permissions and prohibitions are written: who may do what, where and until when.
PLUSHow image-licensing terms travel with the asset.
Open exportHow you leave: the full record, readable without us.
Start here
Send us your least organized folder.
Not a demo set. Not a deck. We read it back to you on your own files, in about a week.