ANY answer containing a business number — load this first.
7 rolesRead more
Your agents already know the job.
A skill is a working playbook — how to review a pipeline deal by deal, close the month so the numbers hold, read an ad account without double-counting. Each one was written around a specific way the work goes wrong, and the agent loads it when the task calls for it. Every one we ship is below.
All skills124Marketing22Revenue19Product & delivery15Connector literacy11Strategy10Finance9People9Execution8Grounding7Build5Legal5CEO office4
SHOWING 1–30 OF 124
Use it when answering from email, Slack, WhatsApp or Teams messages.
7 rolesRead moreUse it when answering from Notion, Drive, Docs, Sheets or a wiki.
7 rolesRead moreUse it when you are about to guess at something only a human can decide.
7 rolesRead moreUse it when a source is not connected, or cannot answer this question.
7 rolesRead moreUse it when something is stuck and you must find who unblocks it.
6 rolesRead moreUse it when deciding whether something is genuinely slipping or noise.
6 rolesRead moreUse it when tracking what somebody promised, to whom, by when.
6 rolesRead moreUse it when deciding whether to escalate, and what to put in it.
6 rolesRead moreUse it when a problem is vague or the question keeps moving.
4 rolesRead moreUse it when answering an invoice, receivables or ledger question.
4 rolesRead moreUse it when answering from a store or payments connector.
4 rolesRead moreUse it when analysing a competitor, or finding our positioning.
4 rolesRead moreUse it when setting or changing a price, a plan or packaging.
3 rolesRead moreUse it when building charts, KPIs or an analytics dashboard view.
3 rolesRead moreUse it when a number or claim is about to be repeated as fact.
3 rolesRead moreUse it when finding customers at risk of leaving.
3 rolesRead moreUse it when asked why conversion or activation is down.
3 rolesRead moreUse it when asked which channel is working, or why one dropped.
3 rolesRead moreUse it when building a tracker, task board or kanban view.
3 rolesRead moreUse it when asked for a report, a readout, or "how are we doing".
3 rolesRead moreUse it when answering from a calendar, Zoom, Meet or meeting notes.
3 rolesRead moreUse it when answering from product or web analytics.
3 rolesRead moreUse it when answering from a tracker — what is done, late, or blocked.
3 rolesRead moreUse it when writing docs, a guide, a README or an internal how-to.
3 rolesRead moreUse it when planning a campaign, an ad set, or a scheduled send.
2 rolesRead moreUse it when a writer needs a brief for a search-driven piece.
2 rolesRead moreUse it when designing or fixing a referral or advocacy program.
2 rolesRead moreUse it when we are about to publish a name, mark or borrowed asset.
2 rolesRead moreUse it when checking whether what was agreed actually happened.
2 rolesRead moreNever state a number without its source7 roles
×A plausible figure invented because the shape of the answer wanted one. A wrong number said confidently costs more trust than fifty "I don't have that"s.
Reading an inbox without believing it7 roles
×Quoting a message as company fact — "we agreed 20% discount" — when it is only evidence that one person typed it. Mailboxes are tier `memory` (lib/connectors/ingest-policy.ts), so this is the family that writes company truth.
Reading docs, drives and wikis7 roles
×Reading a strategy doc written eighteen months ago as the current strategy. Documents carry no expiry and search ranks by relevance, not recency.
Stop and ask instead of guessing7 roles
×Guessing to keep momentum. It is the single most trust-destroying agent behaviour and it is invisible in the output — a guess and a fact are written in the same voice.
What to do when a connector cannot answer7 roles
×Reporting 0 when the honest answer is "not connected" — two different facts that must never be reported the same way.
Name who unblocks it6 roles
×Reporting "blocked on legal" — a situation, not a person, so nobody is asked anything and it stays blocked for another week.
Tell slipping from noisy6 roles
×Flagging everything that moved, which trains everybody to ignore the flags. Our tracker rows and the chat about them disagree by design, so "changed" is not "slipping".
Track what was promised, by whom, by when6 roles
×Recording "we will look into the export bug" as a commitment. No owner, no date, no way to ever be late — so it is never chased and never done.
When to escalate, and what to put in it6 roles
×Escalating a status update. An escalation with no decision in it trains the recipient to stop reading escalations, which is the failure this exists to prevent.
Break a vague problem into an issue tree4 roles
×Answering "why is it happening", "what is the plan" and "what should we do" as one question. Three answers to three unasked questions read as one confident paragraph, and the fix ends up attached to a cause nobody established.
Reading a ledger — invoices, balances and what is really owed4 roles
×Answering "how many invoices are pending and for how much" by summing `total` across whatever statuses looked pending. Zoho Books has no status called pending, the amount owed is `balance`, and a partially paid invoice contributes its remainder — three judgement calls that were all the model's guess.
Reading Shopify, Stripe and commerce data4 roles
×Saying "revenue" when the number is gross orders including tax, shipping, cancellations and not-yet-settled refunds — four different numbers that all get called revenue.
Tear down a competitor4 roles
×Treating a competitor's marketing claims as their capabilities, and shipping a feature matrix that concludes we have more features.
Change a price without breaking trust3 roles
×Picking a number off cost-plus or a competitor screenshot, then discovering the value metric was the real decision. Also: repricing existing customers silently, which costs more in churn than the increase earns.
Charts & metrics dashboard3 roles
×Use it when building charts, KPIs or an analytics dashboard view.
Establish whether a claim is actually true3 roles
×Citing the first source that states a number. Widely repeated figures usually trace back to one original that was an estimate, a projection, or a vendor with an interest — and the repetition is mistaken for corroboration, which is exactly backwards.
Find customers about to leave3 roles
×Discovering churn from the cancellation. By the time the cancel arrives the decision is weeks old, and the signals that predicted it were all present and unwatched.
Find where the funnel leaks3 roles
×Reporting the biggest drop-off as the problem. The biggest drop is structural everywhere; the CHANGED drop is the leak.
Find which channel is actually working3 roles
×Comparing channels on volume, which always crowns the cheapest top-of-funnel source and starves the one producing customers.
Monday-style board / project tracker3 roles
×Use it when building a tracker, task board or kanban view.
Produce a metrics readout3 roles
×A readout that smooths over a missing source so the report looks complete. A gap the reader does not know about is a gap they assume you covered.
Reading calendars and meeting records3 roles
×Reporting that a decision was made because a meeting called "Pricing decision" appeared on the calendar. A meeting existing is evidence it was scheduled, nothing more.
Reading GA4, Mixpanel, Amplitude, PostHog3 roles
×Reporting sessions as users, or users as customers. Three de-duplication rules produce three different denominators and every one of them is defensible on its own.
Reading Jira, Linear and issue trackers3 roles
×Reporting "8 of 10 done" from status fields as delivery progress. The status field is a claim by its owner, updated when they remember, and "In Review" is where work goes to be forgotten.
Write documentation3 roles
×Documenting the intent rather than the behaviour. If the code does something else, the doc is a bug report wearing a doc's formatting.
Brief a campaign before anything ships2 roles
×Assuming an expired approval means yes. A mission approval holds for 72 hours and then expires, and an expired approval is a DECLINE.
Brief a piece so it can rank and still be useful2 roles
×Briefing around a keyword and a word count. That produces a piece that matches the words in the query and answers a different question from the one the searcher asked, which is the standard way a technically correct article fails completely.
Build a referral that is not a discount in disguise2 roles
×Bolting an incentive onto a product nobody recommends, which buys a burst of low-intent signups at a real cost and reads as proof that referrals do not work here.
Check a name, mark or asset is ours to use2 roles
×Clearing a name by checking whether the domain is free. Availability of a domain, a handle or a company-registry entry says nothing about whether a mark is already registered in the class we would trade in — and the expensive discovery happens after launch.
Check whether it actually happened2 roles
×Accepting "done" as evidence of done. The status flip and the outcome are different events, and only one of them was checked.
Show us the work you want to delegate.
Book a tailored demo for AI Crew, Voice, Build or Missions. Track is available free for 30 days.