Your agents already know the job.

Summarising a contract instead of pricing it — a readable summary that never says which clause costs money, and never says that this is not legal advice.

Diagnosing a throughput drop as a motivation problem when the tracker shows the work is queued behind one reviewer.

Calling the generic list-objects tool on a toolkit with 100+ read tools, then reasoning over metadata instead of the pipeline.

Placing reach or follower growth next to revenue in a report, which invites the reader to treat them as the same kind of number. Composio's own Instagram guidance also warns that an empty insights payload is a successful call, not an error.

Repeating a platform-reported conversion count as the business result. The platform both runs the ads and decides which sales they caused; its attribution window is its own.

Answering "are they GDPR compliant?" with yes or no. Compliance is not a property a vendor has — it is a property of a specific arrangement, and the answer depends on what data, on whose instructions, where it lands and who else can see it.

A backlog ranked by enthusiasm, run without a written hypothesis, so nothing is ever concluded and the same ideas return every quarter with the same lack of evidence.

An event judged on registrations, with no follow-up plan written before the date — so the week it cost produces a list nobody works and the same event is repeated next quarter.

A recurring review that reports activity instead of exceptions, and ends with no owner and no date on anything.

Writing release notes as a list of what the team did, ordered by what was hard. The reader is not interested in the work — they are scanning for whether anything affects them, and a note that does not answer that in the first line is not read at all.

Answering a broad research request by gathering everything available. Unbounded research returns a document that is comprehensive and undecidable, arrives after the decision, and is read by nobody — while the one question that actually mattered goes unanswered.

Treating a list as an audience that is owed nothing. Deliverability is earned per send and lost permanently, and a broadcast to a list nobody has cleaned damages every future send including the one that mattered.

Writing one post and pasting it everywhere. Each surface has a different unit of attention, a different first-line budget and a different relationship to links, so the cross-posted version is subtly wrong on every platform including the one it was written for.

A launch date agreed before support, sales and docs were asked, so the announcement arrives ahead of anyone able to answer a question about it and the first day is spent apologising.

A top-down TAM lifted from an analyst headline, which cannot be checked, cannot be acted on, and is the number an investor attacks first.

Opening a public channel with no reason for anyone to be in it, which produces a visibly empty room that reads as evidence the company has no customers.

Use it when the app stores or reads real data that must persist.

Reporting what users asked for as though it were what they need, and counting mentions as evidence of importance. A feature request is a user's proposed solution to a problem they did not state, and the problem is the part that generalises.

Summarising the source. A summary of an hour of talking is the one output nobody wants — the source already exists — and it hides the two or three claims that were worth arguing with, which is the only reason to post at all.

Reporting ticket counts as a workload number, so the recurring cause that generates them is never named and the queue is staffed instead of fixed.

A feature request list ranked by how many people said it, which reliably produces the thing customers can name rather than the problem they have — and buries the one account whose problem was worth solving.

Computing lifetime value on revenue instead of gross margin, and comparing it to a blended acquisition cost. Both errors flatter, they compound, and the resulting ratio is the number a company uses to decide to spend more.

Publishing a result the customer did not confirm and cannot be attributed to us. An unverified number in a case study is quoted back by a prospect, checked by the customer, and it is the customer who is embarrassed — which ends the reference relationship and the story at once.

Writing a procedure from memory rather than from an observed run. Memory reliably omits the steps that have become automatic — the login, the filter, the one field that must be set first — and those omissions are exactly what stops the next person.

Writing the newsletter as a round-up of company news. The list did not subscribe to the company; it subscribed to something useful, and a sequence of announcements trains the most engaged readers to stop opening — which shows up as a healthy list and a dying one.

Rewriting into blandness — losing the specific, odd, true detail that was the best line on the page.

Answering "how are we doing" with whatever metric happened to be connected, so the answer is shaped by our integrations rather than by the business.

Answering yes to everything answerable. A questionnaire answer is a contractual representation that outlives the deal, and the yes that wins a bid becomes the finding in a customer audit eighteen months later, when nobody remembers who wrote it.

Arguing for debt work on the grounds that the code is bad. "Bad" is an aesthetic claim that loses every argument against a customer-visible feature, and the loss is deserved — the real case is a cost that is already being paid and can be shown.

Preparing a narrative and treating the numbers as support for it. Diligence runs the other way — the numbers are checked first and any gap between them and the story is the finding, so a story assembled before the numbers are reconciled creates the problem it was written to avoid.