Skip to main content
Free, MIT licence23 bots in 8 packs

Give Copilot a job it can do without you

23 ready-made agents for Microsoft 365 Copilot Cowork. Each is one file you upload once. After that you ask in plain English, and the work happens inside your own mailbox and files, under your own permissions.

Every one reads and prepares, then stops. Sending, posting and deleting wait for a button you press. And every one ships with a one-page PDF written for somebody who has never heard of Cowork, so you can hand a colleague the guide instead of a repository.

1

Download a bot

One file. Pick from the list below, or take the whole set.

2

Customize, then Skills

In Cowork, open Customize in the side panel and go to the Skills tab.

3

Add, then Upload skill

Select the arrow next to Add. Cowork validates the file and saves it to your OneDrive.

4

Start a NEW chat

Skills are discovered when a session begins, so one already open will not see it.

You need Microsoft 365 Copilot with Cowork enabled by an admin; it is off by default. Cowork is billed on usage and your admin can cap it. Custom skills do not work on mobile. You can hold 50 custom skills but only 25 scheduled prompts, so installing all 23 of these costs under half your skill allowance and none of your schedules unless you ask for one.

Comms

4 bots

Brand Answer Monitor

Tracks what a researcher finds when they look you up, week over week.

Say one of these

  • “What does the web say about us?”
  • “How are we showing up in AI answers?”
  • “Run this week’s brand check”

You get back

  • A spreadsheet row per question, tracked week by week
  • What moved since last week, and the four-week direction
  • New sources quoting you for the first time

It will never, without asking

  • Never claim to measure any named assistant’s answers. The card and this file both say so.
  • Never act on the findings: no content is published, no page is edited, nothing is sent.
  • Never record a competitor’s non-public information.

The rule it follows: A flat week is normal.

Competition Monitor

Reports a material change and nothing else. The suppression is the product.

Say one of these

  • “Anything new from our competitors?”
  • “Check the competitor list”
  • “Did anyone launch something?”

You get back

  • A before-and-after table of what actually changed
  • One line on why each change matters
  • When nothing changed, one line saying so, and nothing else

It will never, without asking

  • Never access anything behind a login, a paywall or a trial. Public pages only.
  • Never report a competitor’s non-public information even if a source offers it.
  • Never characterise a competitor’s intent or finances. Report the observed attribute.

The rule it follows: Silence is the success state and the most common one.

Internal Newsletter Builder

Builds the round-up from what actually happened, and leaves a section out rather than filling it.

Say one of these

  • “Draft this month’s team newsletter”
  • “What should go in the round-up?”
  • “Build the internal update”

You get back

  • A Word document in your usual section order
  • Every item carrying the link it came from
  • Empty sections left out rather than padded

It will never, without asking

  • Never send or publish. Draft to the user.
  • Never include anything on the config exclusion list: individual performance, anything about compensation, unannounced departures…
  • Never announce something not already announced elsewhere. A newsletter is not the channel for breaking news to a team.

The rule it follows: A thin month produces a thin newsletter.

Leadership Update Drafter

Writes the upward update in a fixed shape, with the uncomfortable parts kept in.

Say one of these

  • “Draft my update for leadership”
  • “Write up my monthly report”
  • “Prepare my steering committee update”

You get back

  • Where we are, what changed, what is at risk, what I need from you
  • A short list of what is still not known
  • Facts and opinions kept visibly apart

It will never, without asking

  • Never send. Draft to the user, always.
  • Never soften or omit anything crossing the escalation bar.
  • Never attribute a slip or a problem to a named individual.

The rule it follows: “Nothing changed this week” is a legitimate update.

Core / Executive

1 bot

Calendar Defender

Finds the week’s conflicts and the meetings that should not exist, and proposes, never acts.

Say one of these

  • “Check my week for conflicts”
  • “Which meetings can I drop?”
  • “Do I have any focus time left?”

You get back

  • Your clashes, named meeting by meeting
  • Meetings with no stated purpose and no decision to make
  • Focus time against your target, and the exact changes to make

It will never, without asking

  • Never accept, decline, move, shorten, cancel or create anything. Every output is a proposal for the user.
  • Never touch anything on the config never-touch list.
  • Never characterise a colleague’s meeting as a waste of time. Report the absence of an agenda as a fact and let the user decide.

The rule it follows: A clean week needs one line.

Finance & FinOps

5 bots

Budget Variance Explainer

Writes the story behind a number somebody else owns.

Say one of these

  • “Why are we over budget?”
  • “Explain this variance”
  • “What is behind the overspend?”

You get back

  • Written commentary in your house structure
  • A table of causes, with evidence and how much each explains
  • The remainder that stays unexplained, stated plainly

It will never, without asking

  • Never calculate, recalculate or forecast. Not even arithmetic on supplied figures.
  • Never state a cause as certain when the evidence is circumstantial. Label it.
  • Never write commentary that attributes a variance to a named individual.

The rule it follows: “Largely unexplained” is an honest and useful answer.

Cloud Cost Digest

Explains what moved on the bill, with the question to ask each owner.

Say one of these

  • “What moved on the cloud bill?”
  • “Explain this month’s cloud spend”
  • “Anything odd in the billing export?”

You get back

  • The total for the period against the one before
  • The three biggest movements, with the owner to ask
  • Anomalies no explanation could be found for

It will never, without asking

  • Never state a forecast as a commitment.
  • Never recommend terminating, resizing or deleting a resource. Surface the movement and the question; the owner decides.
  • The provider’s own billing console is the source of truth. This reads an export.

The rule it follows: A flat month is a result.

Expense Reconciliation

Finds the gap between what was claimed and what was evidenced, before someone else does.

Say one of these

  • “Check my claims against the receipts”
  • “What receipts am I missing?”
  • “Reconcile last month’s expenses”

You get back

  • Every claim matched to its receipt, or flagged
  • Counts by status, and the total value unmatched
  • A drafted chaser for each person to chase

It will never, without asking

  • Draft, never send. Chasers are drafts.
  • Never approve, reject or adjust a claim. Never state a reimbursable amount as final.
  • The expense system and the finance ledger are the source of truth. This reads mail.

The rule it follows: A clean reconciliation is the best outcome.

Invoice & Vendor Follow-up

Finds the invoice threads that stopped, with age and amount, and drafts the nudge.

Say one of these

  • “Which invoices have gone quiet?”
  • “Who has not replied about payment?”
  • “Draft the chasers for unpaid invoices”

You get back

  • Every stalled thread, with its age and amount
  • Who you are waiting on, and who is waiting on you
  • A drafted nudge for each one, unsent

It will never, without asking

  • Draft, never send. Especially never send a chase to a counterparty on the config exclusion list…
  • Never state an amount as owed. The ledger is the source of truth.
  • Never escalate a ladder step on your own. The user chooses the step.

The rule it follows: Nothing outstanding is a good week.

Subscription Hunter

Surfaces every recurring charge in the mailbox before it renews itself, with a drafted reply ready.

Say one of these

  • “What are we subscribed to?”
  • “What renews in the next 30 days?”
  • “Find the subscriptions we forgot”

You get back

  • An Excel sheet of every renewal, each row linked to its source
  • What renews in the next 30 days, and what looks like waste
  • A drafted cancellation or renegotiation, unsent

It will never, without asking

  • Draft, never send. Cancellation and renegotiation replies are saved as drafts for the user to read, edit and send.
  • Never cancel, downgrade or modify anything in a vendor portal, and never follow a cancellation link.
  • Never state an amount as the authoritative figure. Finance systems and the contract are the source of truth; this reads what arrived in the mailbox.

The rule it follows: Finding nothing is a success.

IT & Ops

3 bots

Change Advisory Board Prep

Builds the pack from what people actually wrote, and names every change that arrived without a.

Say one of these

  • “Prepare the pack for the next CAB”
  • “What changes are up for review?”
  • “Which changes have no rollback plan?”

You get back

  • The board pack: change, owner, window, risk, rollback
  • Which changes arrived with fields missing
  • Prior incidents on the same systems

It will never, without asking

  • Never approve, reject or schedule a change. The board decides.
  • Never write a rollback plan. Report its absence.
  • Never state a risk rating of your own. Report the proposer’s and say it is theirs.

The rule it follows: A board with nothing to review is fine.

Defect Intake and Triage

Turns “it’s broken” into something an engineer can act on, or says exactly what is still missing.

Say one of these

  • “Turn this bug report into a ticket”
  • “Is there enough here to raise a defect?”
  • “Triage what came in”

You get back

  • A card turning the report into a ticket draft
  • The exact fields still missing, written as questions
  • Possible duplicates, linked

It will never, without asking

  • Never file, create, assign or comment on a ticket without approval. Filing requires a work tracker plugin and the user’s explicit go.
  • Never attempt reproduction. No terminal, no staging, no browser steps.
  • Never declare a duplicate, a root cause or a severity the config does not support.

The rule it follows: A triage that concludes “not enough information, here are the three questions” is a success.

Incident Digest

Rebuilds what happened and when, from the traffic, so the review starts from a timeline instead of.

Say one of these

  • “Build the timeline for this incident”
  • “What happened, and when?”
  • “Prepare the incident review”

You get back

  • A dated timeline of what happened
  • What was observed, what was believed, what is known
  • Open questions, each with an owner

It will never, without asking

  • Never post to a channel, never reply to a thread, never update a status page.
  • Never assign an action. List open questions with suggested owners for a human to assign.
  • Never state a root cause. Surface what the traffic says and label it belief or finding.

The rule it follows: No blame, ever.

People & HR

3 bots

Internal Expert Finder

Answers “who here has already solved this”, with the evidence, before anyone reinvents it.

Say one of these

  • “Who here has done this before?”
  • “Find someone who knows this system”
  • “Who should I ask about this?”

You get back

  • A shortlist of people, with the evidence for each
  • Who to ask first, and who to ask if they are away
  • Where it looked, so you can see what it missed

It will never, without asking

  • Never external. No sourcing, no candidates, no professional networks, no recruitment use.
  • Never infer expertise from a job title, a team name or a reporting line.
  • Never surface a document the user cannot already access. Enterprise Search respects permissions and this skill must not try to work around that.

The rule it follows: Finding nobody is a real and useful answer.

Interview Loop Coordinator

Makes a panel cover the whole bar once, instead of four people asking the same question.

Say one of these

  • “Set up the panel for this role”
  • “Who covers what in the loop?”
  • “Pull the interview feedback together”

You get back

  • Who covers which competency, and what nobody covers
  • What each interviewer needs beforehand
  • Afterwards, the evidence, with disagreements shown

It will never, without asking

  • Never recommend hire or no-hire, never rank candidates, never produce an overall score.
  • Never create or send calendar invitations. Propose.
  • Never include anything about a candidate not supplied for this process, and never search for a candidate externally.

The rule it follows: The pack never produces a hire recommendation.

Onboarding Buddy

Turns “they start Monday” into a dated plan built from what the team actually has.

Say one of these

  • “Plan the first two weeks for our joiner”
  • “Who should they meet in week one?”
  • “What should they read before Monday?”

You get back

  • A day-by-day plan for the first fortnight
  • Who they meet, and why that person
  • What exists for them to read, and what is missing

It will never, without asking

  • Never create calendar events or send invitations. Propose holds for approval.
  • Never include anything about the joiner’s compensation, background or personal details.
  • Never surface documents the joiner will not have permission to open once they start: note them as access needed instead.

The rule it follows: A short plan with real links beats a full one with placeholders.

Projects

3 bots

Document Transmittal Tracker

Answers “what did we send, and who still owes us a response”, from the mail itself.

Say one of these

  • “What did we send the client?”
  • “Who still owes us a response?”
  • “What transmittals are overdue?”

You get back

  • What was issued, to whom, and on what date
  • What is overdue, grouped by party
  • Revisions that quietly superseded an open item

It will never, without asking

  • Never issue, reissue or transmit a document.
  • Never chase a party directly. Drafts go to the user.
  • Never mark anything as approved. Approval is a recorded act in the project’s own system.

The rule it follows: Nothing overdue is the goal, not a failure of the skill.

Lessons Learned Harvester

Finds the lesson that has now been learned four times, which is the only kind worth acting on.

Say one of these

  • “What keeps going wrong across projects?”
  • “Pull the lessons from the close-outs”
  • “What have we learned more than once?”

You get back

  • Themes, and how many separate projects hit each one
  • The lessons recurring across three projects or more
  • Where the same recommendation keeps being made

It will never, without asking

  • Never attribute a lesson to a named individual. Lessons attach to projects and processes.
  • Never state a cause the source material does not state.
  • Never rewrite a recommendation into something stronger than what was written.

The rule it follows: One project’s lessons are not a register.

RFI and Query Tracker

Shows which open questions are actually blocking work, rather than which are merely old.

Say one of these

  • “Which RFIs are still open?”
  • “What is blocking work right now?”
  • “Who owes us an answer?”

You get back

  • Every open question, with who owes the answer
  • Which ones are actually blocking work
  • Overdue items with no stated impact

It will never, without asking

  • Never answer a query on anyone’s behalf, and never draft a technical answer. This tracks; it does not opine on design, method or compliance.
  • Never chase externally. Drafts go to the user.
  • Never infer what a query blocks.

The rule it follows: An empty log is a good day on a project.

Sales & BD

1 bot

Account Expert

One skill, one customer, context rebuilt from the tenant rather than from a CRM.

Say one of these

  • “Brief me before the customer meeting”
  • “What is happening with this account?”
  • “Catch me up on them”

You get back

  • What changed on the account since last time
  • Open commitments, both yours and theirs
  • Who went quiet, and the questions still open

It will never, without asking

  • Never contact the customer. Drafts only, and only when asked.
  • Never produce a health score, a risk rating or a close probability.
  • Never share internal commentary about the account externally, and never include internal pricing discussion in anything drafted to the customer.

The rule it follows: A quiet account is a finding.

Team Lead

3 bots

One-to-One Prep

Walks into the 1:1 knowing what you owe them, which is the half most prep skips.

Say one of these

  • “Prep my 1:1 with Sam”
  • “What do I owe them?”
  • “What did they raise last time?”

You get back

  • What they raised, quoted and linked
  • What you owe them, and what is still unanswered
  • Items carried over from last time, marked and counted

It will never, without asking

  • Never assess, rate, or characterise the person. No performance language, no sentiment read, no “seems disengaged”.
  • Never read 1:1 notes about this person written by someone else.
  • Never draft feedback, and never prepare anything for a performance process.

The rule it follows: An empty brief is a good sign.

Project Decomposer

Turns a project into tasks somebody can actually finish, each with a stated way to tell it is done.

Say one of these

  • “Break this project into tasks”
  • “Who owns what on this?”
  • “Turn this goal into a plan”

You get back

  • A card per task, each with a named owner
  • How you will know each task is actually finished
  • The whole set saved to a file, so it survives the chat

It will never, without asking

  • Plans only. Never create tasks in Planner or To Do, never send anything, never book time.
  • Never assign a named person who has not already been associated with this work in the evidence.
  • Never state a date as agreed unless the evidence shows it was agreed.

The rule it follows: Returning fewer tasks is usually better.

Team Lead Mode

Runs the pack in order and assembles one review, rather than making you run four things.

Say one of these

  • “Set me up for the week”
  • “Run my team lead review”
  • “Prepare everything for Monday”

You get back

  • One document: the three things that need you this week
  • A section per report and per workstream, plus your calendar
  • A run report saying which parts worked and which failed

It will never, without asking

  • Inherits every boundary of the skills it calls, and may not relax any of them.
  • Never send, post, create or schedule anything. It assembles.
  • Never fill a degraded section with its own analysis. Degraded means thinner, not invented.

The rule it follows: A quiet week produces a short review.

What is deliberately absent

Nine designs were cut because Microsoft already ships them: the Legal plugin for Cowork carries 20 skills, Sales carries 9, and a contract-obligation bot written here would be a worse copy of something already in your tenant. Nine more went because inbox triage, meeting prep and transcript-to-actions already exist in the Cowork skills library linked at the foot of this page. They were cut from here rather than deleted: one job, one home, and no reason to publish two versions of it. Four of the seven published Grok Bot roles are absent because they need a terminal, a staging environment or an ad platform.

That last one is a capability gap rather than a judgement call. Cowork has no terminal, no filesystem and no general browser, so any design whose core loop is reproducing a bug in staging is out. Event triggers fire on an incoming email or a Teams message and nothing else, so a bot that watches a folder has to become a scheduled poll or leave Cowork entirely. The reasoning for each is in the repository, which is more useful than a bot that half works.

The other free libraries

All free, all browsable here, all kept in public repositories. Each repository states its own licence: the libraries are CC BY-SA 4.0, the Cowork bots are MIT.

Not affiliated with, or endorsed by, Microsoft. Nothing here advises using AI for safety authorisation: permit to work, lock-out tag-out, confined space, job safety analysis, incident classification or inspection sign-off. AI prepares, people decide.