Tempio
the house byline
Tempio is an AI ads command center. It connects a founder's ad accounts into one normalized workspace and hands their own AI the numbers and the tools, and every change it writes is a change request on the record before it is a change. This page is the byline on the writing: the guides, the posts and the reference pages here are written in-house, by the people building that product, and published under the company's name rather than under an invented persona.
That is a deliberate choice and it has a cost, so it is worth saying why. A byline exists to tell a reader who is accountable for a claim. A fabricated author with a stock portrait tells them nothing, and the reader who checks finds out they were handled. Until a named person's own biography and photograph sit on this site, the accountable party is the company, and the company signs its work.
What gets published here
Three kinds of page carry this byline.
Guides are procedures. A guide says what to do, in order, with the real configuration, and it is written against the product as it exists on the day of publication rather than as it is described in a plan. If a step involves a command or a file, the command and the file are the ones a reader can copy. A guide that cannot be followed end to end by somebody who has never seen the product is not finished.
Posts are arguments. A post says why something is built the way it is, or what a thing in the industry actually means, and it is allowed to have a position. It is not allowed to have a position without giving the reader what they need to disagree with it.
Reference pages — the glossary, the calculators, the specification tables — are facts about the world rather than claims about the product. They are written to be useful to somebody who will never become a customer, because a definition that bends towards a sales pitch is a definition nobody trusts, and because a reader can tell.
How a figure gets onto a page
Every number stated here as a fact about the world carries the page it came from and the date that page was read. That is not a convention; it is enforced at the boundary. A figure arrives in a content entry as a value with a source attached, the source carries a title, a URL and an access date, and the loader that reads the entry refuses it without one. A figure whose source could not be found is left off the page rather than rounded into existence.
The consequence is visible and intended: some pages here carry fewer numbers than comparable pages elsewhere. That is what it looks like when nobody is allowed to invent one.
Figures about this business — spend under management, accounts run, customers — are held to a stricter rule again. They appear only once they have been confirmed as fact by the person who would have to stand behind them. There are no placeholder statistics, no rounded-up totals and no testimonials that were not given.
How a claim about the product gets onto a page
Every statement this site makes about what the product does resolves from a single file that records, per capability and per platform, what is shipped, what is in beta, what is on the roadmap and what is out of scope. Page copy reads from that file rather than asserting on its own account. A capability that is on the roadmap gets its page built and written in full, and that page stays out of the search index and out of the sitemap until the capability's status changes. A reader who arrives at one by link is told so in a banner at the top.
This is why a page here will sometimes describe something in the future tense and say plainly that it has not shipped. The alternative — writing the marketing now and the feature later — is the practice this arrangement exists to make impossible.
The same file holds the vocabulary rules. There are phrases this site does not use, because each of them sells autonomy by implying the owner stops seeing it. Copy is checked against that list on the rendered page, in continuous integration, and a build that would have shipped one of them fails instead.
The framing you will see everywhere here
Every write to an ad entity is a change request first. It records what it changes, the state the entity was in before it, the checks it passed, and the authority that cleared it. That authority is one of two things: you, or a rule you wrote. Approval by a rule is still approval — the record and the reversibility hold either way — and nothing here will tell you that a person must press a button for the product to work, because that is not how it is built and saying so would be a promise nobody should make.
That framing is the reason this site is careful about verbs. While a capability is in beta, the pages do not describe it in the present tense as though it were general. The distinction matters more than it reads, because the gap between "this is how it works" and "this is how it will work" is exactly where advertising software usually lies.
Corrections
Pages here carry the date they were last updated, and reference pages carry the date their claims were last checked against reality. When a figure's source publishes a new edition, or a platform changes how it counts something, the affected page is rewritten rather than patched around, and the date moves with it.
If something on this site is wrong, the address in the footer reaches the people who wrote it. A correction is a change to the page, made in public, not a quiet edit with the date left where it was.
Guides
Ad fatigue and when to refresh
Fatigue is a supply problem before it is a creative one. The three signals that arrive in order, and the four causes a refresh will not fix.
Audit an ad account after a bad week
Establish the week was real before explaining it. A four-pass procedure: check the data, read the record, find the concentration, then decide what to change.
Brief a creative for paid social
A brief that produces testable work names one variable and one job. Seven fields that carry, and the three instructions that guarantee unusable output.
Build a weekly performance review
Thirty minutes, a fixed day, and four questions in the same order every time. What to read, what to ignore, and the one output worth having.
How to choose a bid strategy
A decision procedure rather than a ranking: what you know, what you can afford to lose while a system learns, and how much volatility the business absorbs.
Choose an attribution window
A window is a claim about how long an ad keeps working. Derive it from your own time-to-purchase, then leave it alone long enough to mean something.
Connect any MCP client to your accounts
The general form of the setup: one endpoint, one bearer key, two credential kinds, two tool surfaces, and how to tell a client problem from a key problem.
Connect Claude Code to your ad accounts
Mint a scoped key, add the MCP server to Claude Code, read the account in one prompt, and see what a drafted change looks like at the gate.
Connect Cursor to your ad accounts
Add the MCP server to Cursor with one block of configuration, confirm the tool list matches the key's scopes, and read an account without leaving the editor.
Currency and FX in ad reporting
Minor units, one currency per scope, and a conversion that labels itself. Four places a multi-currency account quietly produces a meaningless number.
Daily versus lifetime budgets
Two different kinds of money, not two ways of saying the same thing. What each one delegates, where each fails, and why a mode change is never a decrease.
First prompts for your ad accounts
Twelve questions worth asking a freshly connected agent, what each one actually returns, and how to read an answer without being misled by a blank.
Give a contractor limited account access
Three roles, one workspace boundary, and credentials that die with a membership. How to let somebody work without handing over the ability to apply.
Guardrails that stop expensive mistakes
Five limits nothing in the product raises, checked three times each. What each refuses, why the aggregate is the interesting one, and what a warning means.
How budget changes restart learning
What a delivery system reacts to, why nobody outside a platform can give you the threshold, and how to make changes you can actually read the effect of.
How to write a media plan
Write the window, the currency, the brief and the line items, check the arithmetic underneath the budgets, then turn the whole plan into drafts.
Name creatives so a report makes sense
The name is the only field that reaches every report downstream. Four things to encode, four never to, and why the decision is made once and lived with.
Pace a budget without a scramble
Month-end panic is a planning failure, not a bidding one. How to size the daily rate, read drift early, and make the corrections that actually get spent.
Plan a product launch flight
A launch is three flights with different jobs, not one campaign with a bigger budget. How to phase the window, size each phase, and keep the record readable.
Plan for a seasonal peak
The one planning problem with real evidence behind it. Read last year properly, size the ramp, decide what happens after, and set the limits in advance.
Read a blended report honestly
A blend is an average over a changing mixture. Four things to check before you believe one, and the two questions it can answer without lying to you.
Read hook rate and hold rate
Two diagnostics, read in order, that narrow where a weak result came from. What each one cannot tell you, and why the benchmark that matters is your own.
Review a batch of campaign changes
Read three fields, then the rows that carry warnings. What a batch guarantees, what stops when one row fails, and how to fix a row without rejecting it.
Run a creative test that answers something
One variable, a size decided in advance, a stopping rule written down. The four failures that make most creative tests unreadable, and how to avoid each.
Run one agent across several accounts
One workspace can hold many accounts across many connections, and one key reads all of them. Where the boundary actually falls, and how to work across it.
How to scope an MCP key
Read, draft and approve are three decisions rather than three rungs. What each unlocks, what the failure mode of each is, and how to rotate one safely.
Set a monthly budget across channels
Start from what the business can afford, not from last month's split. A procedure for sizing a month, dividing it, and writing the review date into the plan.
Set a target CPA that holds
Derive the number from your margin rather than from last month's average, give it room to be an average, and know what happens when it is set too tight.
Set up UTM conventions that last
Five parameters, one decision each, made once. Why lowercase and no spaces are not pedantry, and what a convention has to survive to be worth having.
Size a test before you run it
Three numbers decide whether a test can answer anything: your current rate, the difference worth acting on, and the volume you can afford. Work them first.
Spend caps and kill switches
What actually stops money leaving, in order of how fast it acts. The one control that stops everything, and the three limits nobody can raise.
Split prospecting and retargeting
The split that looks like a budget decision and is really a measurement one. Why retargeting flatters itself, and how to size the two honestly.
The tradeoffs of campaign budget optimisation
Putting the budget on the campaign buys efficiency and costs control. What you actually give up, when the trade is right, and how to keep the record readable.
Turn a media plan into campaign drafts
Materialising a plan, step by step: what is validated before anything is written, what comes out, the four refusals, and why the plan freezes afterwards.
What an agent can reach in your account
An honest inventory: what a connected client reads, what it can write, what it can never touch, and the four limits that hold whatever you ask it to do.
When to use a cost cap
A cap buys predictability with reach. The three situations where that trade is right, the two where it is a mistake, and how to tell which one you are in.
Who approved what, and when
The record names an actor rather than a user, which is what lets it hold a rule as readily as a person. How to read it, and the one field most worth knowing.
Why platform numbers disagree
Six causes, three of them benign. How to tell a restatement from a structural difference, and which disagreements are worth an afternoon of your time.
Write a brief that survives handover
A brief is read by the person who inherits the campaign, not by the person who wrote it. The seven fields that carry, and the four that never survive.
Write an approval policy for ad changes
A rule that never fires is worse than no rule. What one may say, the five refusals it cannot express, and how to write one you would defend later.
Posts
A week of ad changes on the record
What a week looks like when every change carries its own before state and authority: five questions that become lookups instead of investigations.
An agent with your ad account
The useful question is not what an agent can do with your accounts. It is what it cannot do, and whether that limit is a policy or a property of the system.
MCP is a contract, not a feature
Supporting a protocol is a line on a comparison table. What makes it useful is what the server refuses, what it derives, and what a caller cannot name.
Reading a change request like a diff
The habit that makes review fast is borrowed from code: read what it changes from, not what it changes to, and read the group before any single row.
The case against the autopilot toggle
A switch with two positions asks the wrong question. Autonomy is not a level to set once; it is a property of each change, with a direction and a distance.
The cost of three dashboards
The cost is not the tab-switching. It is that no number in one is comparable with a number in another, and the reconciliation lives in somebody's head.
The numbers your dashboard rounds off
Five roundings that happen before you see a figure: a blank turned into a zero, reach summed, currencies added, a partial day, and a restated week.
What a media plan is for
Not for predicting the quarter. For being wrong about it cheaply, in a document where a line costs a keystroke instead of a day of spend.
What an approval policy should say
Most approval policies name people and hope. A useful one names a direction, a ceiling and a distance, and refuses to be written any wider than that.
What we refuse to claim, and why
A list of things this site will not say, enforced by the build rather than by good intentions, and the argument for making refusals mechanical.
Why every change is a draft
A change request is the smallest unit that carries an authority, a before state and an inverse. That is the argument for making it the only unit.
Why ratios belong at read time
A stored rate can disagree with the numbers underneath it, and nothing will tell you which one is wrong. Computing at read time removes the failure entirely.
Put frontier AI to work on every ad account you run.
Tempio connects your ad accounts into one workspace and hands your own AI the numbers and the tools. Every change it writes is a draft until you approve it, or until a rule you wrote does.
Request early access- Read and draft by default. Approve is a separate permission you grant.