Grok Bot skills: the six-section format xAI recommends
xAI recommends six sections for a Grok Bot skill: when to use it, inputs and access, order of work, validation, what to return, what requires approval.
A Grok Bot skill is, in xAI’s words, “a reusable set of instructions for how to do a task.” The skills page recommends six sections, in this order: when to use it, inputs and access, the order of work, how to validate the result, what to return, and what requires approval. Skills are available across your bots and are reached by typing / in the desktop composer.
Those six headings are the format. This page is about what belongs under each one, and why the order is the order.
What does xAI say a skill is?
The docs say a skill captures “steps, decision rules, expected output, and safety boundaries,” and that skills “are available across your Bots.” The FAQ draws the line between a skill and a routine: a skill describes how a task is done; a routine “assigns a workflow to one Bot and tells it when to run.” So a skill is the method, a routine is the schedule, and the bot’s instructions are the role that ties them together. A template carries all three, because all three transfer (Bots docs).
What goes under each of the six sections?
| Section | The question it answers | Common mistake |
|---|---|---|
| When to use it | Which request or routine invokes this, and which does not | Only the positive case; no "do not use it for" |
| Inputs and access | Which systems, at which scope, and what to do if one is missing | Assuming a connector is authorised |
| The order of work | The steps, in order, with the order justified where it matters | A recipe with no reason for step four |
| How to validate the result | What must be true before the output is trusted | None. The skill trusts itself |
| What to return | The shape of the deliverable, and what is attached as evidence | Prose where a table was needed |
| What requires approval | Which actions stop for a named human, and which never happen | "Use judgement" |
Scroll sideways for the third column on a narrow screen.
When to use it. Name the triggers: the routines that call it and the phrases a person would say. Then name the cases it must refuse. Waterline’s margin skill says not to use it on a period whose reconciliation has not passed, and not to use it to answer “how did we do last week” with revenue. The negative cases are where a strong model earns its keep, because they stop it doing the plausible wrong thing.
Inputs and access. Each system by name, the scope needed, and whether it is read or write. Then the missing-data rule. The docs ask for “a no-data and stale-data policy” in every automation, and this is the section that holds it for a skill.
The order of work. Steps only where the order carries weight, and the reason beside each one that does. Reconcile before you chase, because chasing an invoice that was already paid is the most expensive mistake in collections. A sequence without reasons is a recipe, and a recipe performs exactly as well as its author understood the job on the day it was written.
How to validate the result. The checks that run before the output leaves the skill. Does the bridge sum. Did the pull reach the last page of every endpoint. Do two systems agree, and if not, is the disagreement reported as a finding rather than silently resolved.
What to return. The deliverable’s shape: which fields, which evidence attached, which system and record each number came from.
What requires approval. Three lists: what the skill may do alone, what stops for a named human, and what it never does at any authority level. xAI’s security page says Require Approval wins whenever both kinds of rule match, so this section can be narrow and specific without fear of a broad allow rule overriding it. The approval rules guide covers the scoping sentence that belongs here.
Why does the order matter?
The six sections run from the decision to invoke to the decision to act, which is the order a person reading the skill needs them in. Inputs come before sequence because a skill that discovers a missing connector at step five has already done four steps of work it will throw away. Validation comes before return because a result that has not been checked is not a result. Approval comes last because it is the section most likely to be truncated if the block runs long, and the last thing a bot should lose is procedure, never permission. Botemplate’s release gate measures where authority material sits in every instruction block for exactly that reason; in the first release it sat in the first 5 to 18 percent of every block, so truncation destroys procedure and never permission.
What does a skill in this format look like?
Ledger’s /ar-reconcile is the shortest example, sketched from its public summary. When to use it: before any customer is contacted, because Ledger will not chase an invoice it has not reconciled in the last 24 hours on a pull that reached the last page of every endpoint. Inputs and access: Stripe and QuickBooks, both authorised read only, and the AR ledger tab in the register sheet, keyed off QuickBooks invoice ids. The sequence: a three-way reconciliation of the three sources. Validation: a disagreement between two systems is a finding, not a tie to break. What to return: the outcome Ledger owns, which is that every invoice past terms carries a next action, a named owner and a date. What requires approval: this skill holds no write scope on any source system and contacts nobody, so nothing here needs approval, but its output gates the skill that does.
The one-line summary of each of Ledger’s eight skills is on its page. The full six sections are in the SKILLS.md that ships with the bot.
Does a skill need to be written for a strong model?
Yes, and the format helps. Grok 4.7 is weeks away, and a skill written as a rigid recipe makes a stronger model dumber on purpose: it will follow step four even when step four is wrong, because it was told to. The six sections leave room for judgment in the sequence while pinning down the parts that must not move, which are inputs, validation and approval. Every Botemplate skill adds one clause for that: if following a listed step would produce a worse outcome than the step was meant to achieve, state the conflict, do the better thing within existing authority, and record the deviation. Never silently skip a step. Never invent authority.
Every Botemplate bot ships six to eight skills in this shape, 35 across the five, each one reachable from the composer with a slash. How it works covers what else is in the box, and the Operator Pack is the five together for $69.