- Company: the facts every draft is written from, your brand and values, and the competitors you track.
- Skills: your playbooks, and the skills built into every agent.
- Guardrails: your compliance rules, how the agents write, and the words a draft may never use.
/knowledge, opens Customize → Company. The other two pills, Connectors and Agents, are described on Connectors.
What does this change?
Knowledge and Skills changes how the agent writes, not what it measures. A brand profile tells it who you are, what you sell, and how you talk. Skills give it reusable instructions to apply every time. Your Share of voice, Citation rate, and Source share are untouched by anything on this page. Without it, the agent still answers correctly — it just answers generically, and you end up pasting the same context into every chat and every automation.Company
Customize → Company holds the brand profile. The brand profile is the standing context the agent carries into every conversation in the workspace. Put in it the things you would otherwise repeat: what the product does, who buys it, the words you use for your own categories, the competitors you care about, and the claims you are not willing to make. Be specific and be short. Vague positioning statements make output vaguer. Concrete constraints — “never call it a platform”, “always distinguish crawlers from models”, “prices are in USD, monthly” — do visible work on every answer that follows.Drafted for you at setup
When a workspace is set up, Attensira reads your homepage, plus/about and /pricing when they exist. It drafts only the profile fields those pages actually state, and each draft names the page it came from. Drafts show as drafted until you confirm them. A field that already holds a value is never overwritten.
When the agent needs a field that is empty, it raises one row in the Inbox, fill in your business profile, shared by every empty field. It is a question: answer it in the row and the answer is saved to the field it names. The row closes when a profile field is saved.
Answers you give to the agent’s other questions in the Inbox are kept too, as facts you stated, and every run reads your newest ones, up to 40. There is no screen that lists them.
Brand and values
Two groups of the profile say who you are rather than what you sell:- Brand: your name as said aloud, your tagline, your colours and your logo URL. Scripts and voice answers use the spoken name, listings reuse the tagline and logo, and images and charts use your colours.
- Values: what you stand for, as a list. Drafts are written so they do not contradict it.
Competitors
The competitors you track are listed on the Company pill too, and you can add or remove them there. Each tracked competitor adds an Against rival field to the profile: why a buyer would choose you over them, used on every comparison and alternatives page about them. A competitor you add is measured from the next reading on; see Competitors.Reporting standard
The profile’s Reporting group has one free-text field, Reporting standard: the tone of your reports, who reads them, and words to avoid. Every client report follows it and stores a copy of it. It never overrides the report rules: no percentages, and no movement that a re-read did not prove.Skills
Customize → Skills lists two kinds: your playbooks, which you write, and Built in, the skills we maintain. Every agent has the built-in ones; they read your company facts and follow your guardrails. You cannot edit a built-in skill. The list in the app is always the current set. A skill is a named, reusable instruction for a recurring job: a weekly summary format, the structure of a competitor brief, the tone for a customer-facing note. Instead of restating the format each time, you point the agent at the skill. Skills compose with the brand profile. The profile decides voice and boundaries; the skill decides structure and steps for one kind of output. Keep each skill narrow — one job, one shape. A skill that tries to cover every report produces output that fits none of them.1
Start from something you have already asked twice
The second time you retype an instruction is the signal it should be a skill.
2
Write the output shape explicitly
Sections, order, length, what to do when a number is not measured.
3
Use it in a chat first
Check the shape on real workspace data before you attach it to a cadence.
4
Attach it to an automation
Once the shape holds, a recurring run produces comparable output week over week.
Guardrails
Guardrails are your rules for everything the agents write in your name. Every draft, whether an article, a page fix, a social post, an outreach email or a community reply, is checked against them before anyone can approve it. The agents do the checking and the rewriting. You write a rule once, and every draft after it is held to it. Customize → Guardrails has three sections, hardest first:
Words to avoid are matched exactly, ignoring case, and on whole words: avoiding “cheap” does not stop “cheaper”, and avoiding “HIPAA certified” stops “hipaa certified”. They are listed explicitly, never guessed from a rule’s text, so a sentence like say “practice”, not “clinic” never stops a draft for saying “practice”. Write that sentence as the rule, and put “clinic” in its avoid chips.
When and where a rule applies
Every rule shows When: … · Where: … under its text.- When is the rule’s one-line description: the situation it is about.
- Where is the kinds of draft it holds: pages, articles, emails, community replies, social posts or ads. A rule applies to every kind unless you narrow it. A rule about email sign-offs, set to emails only, never holds back a page fix.
The verdict on each draft
Each draft gets one verdict, for the exact text it was checked against:
The verdict also counts the rules the draft was checked against: your rules that apply to it plus our own house safety rules, which every draft is checked against on every plan. A review or stopped verdict names each rule involved, and for a forbidden word, the word.
When the agent writes a draft, it is told the verdict straight away and revises once if it can keep the rule without losing the point. An article that is stopped is not filed for approval at all; if the revision is stopped too, the agent asks you in your Inbox what the article should say instead.
A draft nobody could check is never passed
If the check cannot finish, because your rules could not be read or the check did not answer for every rule, the draft is not checked, never passed. It cannot be approved, published or sent until a check finishes. One row in your Inbox, We could not check N drafts against your rules, says so for the whole workspace, with a Retry button that opens the draft. A draft whose text changed after it was checked is checked again before it can be approved or sent. One that changed after it was approved is not published: it has to be checked and approved again. A stopped or unchecked draft never goes ahead on its own, whatever the workspace’s approval mode. A draft marked for review does not go out on the workspace’s own say-so either: it waits for your approval instead.Unblocking a draft
There is no override button. A stopped draft moves only when it no longer breaks a rule:- Revise the draft. Edit it yourself, or ask the agent to rewrite it. An edited draft is checked again as soon as it is saved, and a clean one can be approved.
- Change the rule. If the rule is wrong, edit it, narrow where it applies, remove the word from its avoid chips, or switch it off. Every draft checked under the old rules is checked again when you approve it.
- Retry a draft that was not checked. Retry on the Inbox row opens the draft. Approve it and the check runs once more before the approval goes through.
409 Conflict and a message naming the rules it breaks. Over the REST API the error’s details.reason is guardrail_stopped, or guardrail_unchecked when the draft could not be checked, and details.guardrails is the whole verdict. A work item read over the API carries the same verdict as its guardrails field: status is passed, review, stopped or unchecked, and checked is how many rules it was checked against.
On every plan
Writing rules is part of Knowledge and Skills. Checking drafts against them is not: every rule you have switched on, and our house rules, are checked on every plan. A workspace that moves to a plan without custom skills keeps the rules it wrote, and every draft is still held to them.What it will not do
Knowledge and Skills cannot make the agent assert something the data does not support. A brand profile that says “we lead the category” does not turn an unmeasured cell into a number, and a skill that asks for a share-of-voice ranking cannot produce one — Attensira measures mention and citation rates against a shared denominator, so brand rates do not sum to 1 and there is no brand rank metric to report. It also does not change sampling. Depth stays at n=1 on Starter and n=3 on Growth, Business and Agency, and each Prompt × Model × country is still measured at most once per workspace-local calendar day.Knowledge you write vs. what the agent already knows
The brand profile and skills on this page are yours to write. Separately, and on every plan, the agent keeps its own nightly notes about your workspace — see What the agent remembers.On Starter
On Starter you have the business profile, and the agent reads and follows it and the rules in every chat and automation. Every draft is still checked against your guardrails and our house rules. Writing custom skills shows an upgrade state — a screen that names the plan requirement and does nothing else. Nothing you configure elsewhere is lost by being on Starter. Plans are Starter 349, Business 2,000 per month, with yearly billed at ten months’ price. Crawler logs are also Growth-and-above. See Plans.House skills
Besides the skills you write, the agent follows Attensira’s own procedures for recurring jobs, such asorchestrate/daily-win-plan for the Daily win plan and report/client-weekly, report/client-monthly and report/client-quarterly for client reports. See The agent’s tools.