2026-08-25 · Reference

Bots For Every Role: The Complete Directory

Job titles are a bad way to choose automation, because a title describes what you are responsible for and a bot can only help with what you repeat. Two people with the same title on the same team have different weeks, and the useful question is never "what do marketers automate". It is "what did I do four times this month that produced the same shape of output every time".

This is the directory over fifteen role playbooks and the seventy-seven bots in the catalogue. Each section below describes what that role's week actually contains, names the two or three bots worth starting with by their real slug, and states the one risk that is specific to that role rather than generic to bots. Three of the roles carry constraints that are not advisory, and those sections say so plainly.

On this page

• Read your role as a week, not as a job title
• Start with the whole directory in one table
• Founders need a router before they need a specialist
• Sales work splits into research you can hand over and conversations you cannot
• Marketing gains the most and risks the most from a publishing grant
• Support is the role where drafting beats sending by the widest margin
• Engineering wants a reviewer that comments and never merges
• Product managers need synthesis, not another backlog groomer
• Recruiting bots handle logistics and never touch the decision
• Finance bots produce exceptions lists, never entries
• Agencies win back the reporting hours and nothing else
• Consultants should automate the deliverable scaffold, not the thinking
• Writers should automate the research and defend the draft
• Educators keep the gradebook and the student roster off the bot
• Real estate bots are bound by fair housing on wording and on targeting
• Ecommerce bots watch the queue and stay out of the warehouse
• Personal life is where the boring bots pay off fastest
• Pick your first three by category, not by enthusiasm
• Every bot in the directory holds the same shape of line
• Build from the nearest neighbour when your role has no bot yet
• Check at thirty days whether the role bot still earns its seat
• Answer the strongest objection: a role bot is a prompt with a logo
• Where the role framing stops being useful
• Frequently Asked Questions

Read your role as a week, not as a job title

Open last week's calendar and inbox and sort what you find into three piles.

The first pile is work only you can do: the pricing conversation, the decision to cut a feature, the difficult reply to a customer who is about to leave, the judgment call that has your name on it. It is smaller than it feels, usually between five and eight hours.

The second pile is work that repeats with identical steps and teaches you nothing after the fifth time. Rebuilding the same account brief from twenty browser tabs. Checking whether a competitor moved. Retyping the same three replies. Reconciling two systems that should agree and never do. This pile is where the calendar quietly went, and it is the entire target for a bot.

The third pile is the trap: work that repeats and still teaches you something every time. Reading churn notes. Sitting inside a support ticket. Writing the weekly update. From outside it looks exactly like pile two, and handing it over buys back hours while quietly removing the input that keeps your judgment current.

Every section below is a claim about which parts of a particular week land in pile two. If your week disagrees, trust your week. What is a Grok Bot is the plain explanation if you are still deciding whether any of this applies, and how to create a Grok Bot is the mechanical first setup.

Start with the whole directory in one table

Fifteen roles, the residue in each one, where to start, and the risk that belongs to that role specifically rather than to bots in general.

Read the last column as the thing to design against. Everything else on the row is upside, and upside takes care of itself.

Founders need a router before they need a specialist

A founder's week is five jobs wearing one hat, and the failure is always the same: you build the impressive specialist bot first, because it is the interesting one to build, and you still have no idea what is on fire on Monday morning.

Build the router first. A router reads across mail, calendar, tracker, and whichever channel your customers actually use, and produces one document that says what needs you today and what is drifting. It decides nothing. That is the job of Chief Of Staff, which never reassigns, closes, or comments on an issue and never posts in Slack, and of Chief of Staff Briefing, which never sends a message, replies to a thread, or moves a calendar event.

Add Inbox Triage second, because a founder inbox is where sales, support, and finance all arrive wearing the same envelope, and add Personal CFO third if your company finances and personal finances are still the same conversation.

The full argument for five and only five is in bots for solo founders, the comparison shortlist is the best AI bots for founders, the routing setup itself is in the chief of staff bot, and the wider case is building a one-person company.

Sales work splits into research you can hand over and conversations you cannot

The sales week divides cleanly, which is why it is the best-served role in the catalogue with thirty-one bots. Everything before the conversation is assemblable. The conversation is not.

Start with Lead Scout, which never contacts anyone and does research and ranking only, then Meeting Prep Brief, which never emails or messages anyone on the account. Add Call Follow-Up Drafter once you trust the first two, and Prospecting Sheet Builder or LinkedIn Signal Watch when volume justifies it. If you want drafts in your own voice waiting for you rather than sent, Outbound In Your Voice never sends a message and never sends a connection request.

Depth lives in bots for sales reps and the best AI bots for sales teams, with the two builds in researching leads overnight and following up with prospects.

Marketing gains the most and risks the most from a publishing grant

Marketing is the role where a bot most obviously pays for itself and most obviously embarrasses you, and both come from the same permission. The upside is that a marketing week contains an enormous amount of assembly: calendar upkeep, competitor checks, repurposing one asset into six, briefing the same brief again. The downside is that the output is public.

The catalogue is uniform here. Content Planner Manager never publishes and proposes every status change, Evergreen Content Flywheel leaves recycled drafts unscheduled with no slot time, X Account Crew never posts or schedules from your account, and Ad Creative Generator never launches a campaign or spends a credit. Pair one of those with Competitor Pricing Watch or Competitor Ad Watch, which read published sources and nothing else.

Read bots for marketers and the best AI bots for marketing teams for the roster, running a content calendar and monitoring competitors for the builds, how to stop your bot producing slop for the quality problem, and automating social content without losing your account for the platform risk.

Support is the role where drafting beats sending by the widest margin

Support has the highest volume of near-identical work of any role here, which makes it the most tempting place to let a bot reply and the worst place to learn that it should not have.

The split that works is simple and it holds across every helpdesk. The bot reads the queue, groups what is actually the same ticket wearing different words, writes the first draft, cites the article it drew from, and stops. A human reads and sends. That last step takes seconds and is the entire difference between a fast team and an apology.

Support Queue Pass never replies, reassigns, re-prioritises, merges, or closes a ticket. Support Reply Drafter never sends or marks a ticket answered. Support Autopilot leaves unsent drafts and internal notes only, and Help Center Updater ships each change as current text beside proposed text for a human to paste.

The role guide is bots for support leads, the shortlist is the best AI bots for customer support, the build is triaging support tickets, and the design question of when to stop entirely is in designing the handoff.

Engineering wants a reviewer that comments and never merges

Engineers automate for a living, so the bar for a bot in this role is higher and the useful set is narrower. What survives is the work that is genuinely repetitive and genuinely low-stakes: a first review pass that catches the obvious things before a human spends attention, a standup note assembled from commits and tickets, a periodic sweep for the config and dependency drift nobody has time to look for.

What does not survive is anything that can merge. A review bot with merge rights is a deployment pipeline with a language model at the front of it, and the failure mode is not a bad review, it is a bad review that shipped. PR Review Sentinel never merges, approves, pushes, or requests changes and comments only. Codebase Hardening Auditor never opens a pull request or pushes a commit, and findings land in a report you read. Engineering Agent Manager never merges, approves, or pushes to the default branch.

Standup Scribe is the cheapest win in the whole catalogue for this role, and it posts only to your own direct message rather than to a shared channel, which means a wrong standup is a note to yourself.

Read bots for engineers and the best AI bots for developers, then the builds in reviewing pull requests and writing your standup.

Product managers need synthesis, not another backlog groomer

The product week is mostly reading: tickets, calls, sales notes, support threads, analytics, the churn survey nobody reads. The bottleneck is not capture, it is synthesis, and synthesis is exactly where a bot is good and where it is dangerous in the same motion.

Good, because a bot will actually read all four hundred tickets rather than the twelve you remember. Dangerous, because a summary flattens intensity: one furious enterprise customer and forty mild mentions come out of the blender as "users say", and you cannot tell which from the output. The fix is to require counts and quotes rather than adjectives, and to keep the raw list attached to every summary so you can check the claim.

Tickets To Changelog never publishes or replies to a matched ticket, What Did We Promise never contacts the account and never answers a promise on your behalf, and Churn Watch never pings the customer. Docs Self-Serve Assistant produces a ranked gap list for a human writer and never edits a page.

The role guide is bots for product managers, with adjacent builds in triaging bugs, catching churn early, and cleaning up stale docs.

Recruiting bots handle logistics and never touch the decision

Recruiting has the clearest line of any role on this page, and it is not a matter of taste. A recruiting bot does logistics. It never rejects a candidate, never scores one, never ranks or orders a list of them, and never contacts an applicant directly. Extraction is fine: pulling stated years of experience or a named certification out of an application is reading. Producing a number, an order, or a decision is not, and a total is a score even when you call it a summary.

That leaves a large amount of genuinely useful work, because the recruiting week is mostly scheduling across four calendars, reformatting the same information, chasing confirmations, and keeping notes consistent. None of it touches the decision.

The catalogue has no recruiting-specific bot as of writing, so build from neighbours: Inbox Triage for the application inbox, Org Chart Keeper for mapping a team you are hiring into, and Persistent Bot Memory for keeping notes consistent, which never stores secrets or customer data at all.

Employment decisions are regulated differently in different places and the rules move. This is not legal advice, and the safe reading is that a bot which never scores, ranks, rejects, or contacts stays clear of the question entirely. Bots for recruiters works through the line in detail, and screening job applicants shows what extraction looks like when it stays extraction.

Finance bots produce exceptions lists, never entries

Finance is the role where the useful output is a list of things that look wrong. Not a fix, not a posted entry, not a reconciled account. A list, with the evidence attached, that a person works through in twenty minutes instead of three hours.

Bookkeeping Auditor never posts an entry, reconciles an account, or files a return, and produces the exceptions list and stops. Subscription Pruner never cancels or unsubscribes anything, and Subscription Cancellation Advisor hands you the path and the date and leaves the decision with you. Personal CFO never moves money or trades.

Read bots for finance for the role, then reconciling expenses, chasing unpaid invoices, and auditing subscriptions for the three builds that cover most of it.

Agencies win back the reporting hours and nothing else

Agency work has one enormous, obvious target: the same report, rebuilt for twelve clients, from the same four dashboards, on the same week of the month. It is pure assembly, it is high volume, and nobody enjoys it. Automate that and you have recovered a meaningful share of a month.

Be honest that this is the whole prize. The parts of agency work clients actually pay for are the recommendation, the creative judgment, and the relationship, and none of those improve when a bot writes the commentary. A report with generated commentary that nobody read before it went out is worse than a bare table, because it commits your agency to a claim in writing.

QBR Pack Builder never shares, sends, or presents the deck to the customer and leaves it a private draft for the account owner. Account Expert never messages anyone at the account. Enablement Pack Builder lands as an internal draft link for one teammate. Add Competitor Ad Watch for the client-facing intelligence that is genuinely differentiating.

The role guide is bots for agencies, the build is reporting weekly KPIs, and once you are running one bot per client account, running a team of bots without chaos is the next problem you will have.

Consultants should automate the deliverable scaffold, not the thinking

Consulting looks similar to agency work and behaves differently, because the deliverable is the thinking rather than the report. The repeatable residue sits either side of it: assembling background before an engagement, keeping a deck library current, turning three hours of interviews into a structured note, localising the same material for a different market.

The line to hold is that a bot may build the scaffold and never fill the argument. A structured note listing what each interviewee said, with quotes and attribution, is a scaffold. A paragraph beginning "the key insight is" is the thing the client is paying you for, and outsourcing it produces work that reads like everybody else's.

Deck Updater never edits a slide or touches the live deck and sends a correction list to the owner. Brand Deck Keeper never edits the master deck or the shared library. Deck Localizer produces a separate copy and a register for a human approver rather than softening a flagged claim. Podcast Summarizer never posts or shares anything, and Meeting Prep Brief keeps the brief internal.

Bots for consultants covers the research and deliverable split, and prepping for meetings is the build most consultants run first.

Writers should automate the research and defend the draft

A writer's week has two halves and only one of them should be handed over. The first half is gathering: finding sources, reading them, extracting the numbers, noticing what everyone else has already said, checking whether the claim you half-remember is true. That half is assembly, it is slow, and a bot is good at it.

The second half is the draft. Hand it over and you get output that is grammatical, on-topic, and completely without a point of view, which is the thing readers were coming for. The economics of this are worth stating plainly: generated prose is cheap and abundant, which means it has approximately no value, and the scarce thing is a person who noticed something.

Content Idea Generator ends at a ranked idea list and never writes a script or a draft. Podcast Summarizer brings episode briefs back to you alone. Evergreen Content Flywheel never publishes or schedules a recycled post.

Read bots for writers for the workflow, how to stop your bot producing slop for the quality argument in full, and digesting your newsletters for the research build that pays off fastest.

Educators keep the gradebook and the student roster off the bot

Teaching splits into contact hours and everything around them, and the everything around them is enormous: planning, resource hunting, differentiating the same material three ways, writing feedback, and the administrative layer that grows every year.

Two constraints hold here and neither is negotiable. First, no bot grades autonomously. A bot may draft feedback that a teacher reads, edits, and owns, because the editing is the reading. A bot may not produce a mark that goes into a record without a human forming the judgment. Second, student data is protected, and the practical version of that rule is that names, identifiers, and anything that could reconstruct a specific child do not get pasted into a bot at all. Strip before you paste, not after.

That leaves the planner, which never sees a name and is the single best bot in this role. Content Idea Generator produces ranked ideas and outlines with no student data anywhere near it. Podcast Summarizer turns long source material into briefs. Persistent Bot Memory never stores secrets or customer data, which is the right default when the data is a classroom.

Student records are governed by different laws in different countries and sometimes by your institution's own policy on top. This is not legal advice, and the check that matters is your institution's, before the first paste. Bots for educators works through planning and feedback support in detail.

Real estate bots are bound by fair housing on wording and on targeting

Real estate has more repetitive work than almost any role on this page. The week is drive time, dead time, lead sorting, listing preparation, and a follow-up cycle that runs for months. All of that is automatable in shape.

One constraint governs the whole role, and it applies twice. Fair housing rules reach the wording of a listing or a message, which is the part everybody expects, and they reach the targeting of who sees it, which is the part that catches people out. A bot that writes a listing describing an ideal resident rather than a property is a problem. A bot that filters or ranks a lead list on a proxy for a protected characteristic is the same problem wearing a spreadsheet, and removing the obvious field does not remove the signal, because a postcode carries it.

So keep the bot on facts the person told you and on logistics. Motivated Seller Finder never contacts an owner, agent, or occupant and never loads a list into a dialer or a mail campaign. Booking Pipeline Runner never sends a calendar invite, a reschedule, or an email. Inbox Reply Digest never replies, forwards, archives, or labels, and the digest is read only.

Housing advertising and lead handling are regulated, the rules differ by jurisdiction, and none of this is legal advice. Check the rule where the property is. Bots for real estate goes through wording, targeting, and the follow-up charter clause.

Ecommerce bots watch the queue and stay out of the warehouse

An ecommerce week is a queue with a warehouse attached, and the distinction between those two halves is the whole design. The queue half is information: which orders look wrong, which reviews are saying the same thing, which competitor moved a price, which support tickets are one bug. A bot handles all of that well.

The warehouse half is physical. A fulfilment action reaches a picker, a box, and a courier, and the undo involves a human driving somewhere. Cancellations have the same property in reverse. Keep both away from the bot permanently, and notice that "edit a live product" belongs in the same category, because a price typo on a live listing sells at the typo.

Support Autopilot never sends a reply and never changes a conversation's status, assignee, or tags. Support Ticket Fixer never posts or sends a message in the thread and never closes or merges a ticket. Competitor Pricing Watch reads published pages only and never fills a form or starts a trial. Bookkeeping Auditor closes the loop between the payment processor and the books.

The role guide is bots for ecommerce, and the two connections that matter are covered in Grok Bot and Shopify and Grok Bot and Stripe.

Personal life is where the boring bots pay off fastest

The personal bots are the least impressive and the most consistently useful, because household admin is pure repetition with no upside for doing it well. Nobody is promoted for renewing insurance on time.

Every one of those stops one step before the irreversible act, and that is not timidity. It is that the entire value of a household bot is not having to think about the chore, and a bot that pays a bill wrongly puts the chore back in your week with interest. Bots for personal life has the full set.

Pick your first three by category, not by enthusiasm

Seventy-seven bots spread unevenly across six categories, which tells you something about where repeatable work actually accumulates.

Pick one from your role, one from productivity, and one from personal. The role bot proves the concept on work you care about, the productivity bot compounds across everything else, and the personal bot is the one that keeps running when you get busy, which is the real test.

## Role and scope

You are my [role] bot. You work on exactly these recurring jobs:

1. [job one, with the source it reads and the document it writes]
2. [job two]
3. [job three]

Anything outside those three is out of scope. If something looks urgent and
is out of scope, write one line about it in the run notes and stop.

## Boundary and escalation

You never [send / post / merge / pay / grade / score / fulfil]. When a run
reaches that point it ends with a proposal in [document], with my name on
the decision.

## Output

One document per run, at [path]. Sections: what changed, what needs me
today, what I am ignoring on purpose, and sources for every claim.

The starter roster argues for which three, and the charter template is the fill-in version of the block above.

Every bot in the directory holds the same shape of line

Read down the boundary column of any role above and the same grammar keeps appearing. That is deliberate: a listing on this site has to declare one act the bot never takes without a human, and the declaration is what makes the bot safe to leave running rather than something you supervise.

Notice that the line is always the same kind of thing: the first act with an audience outside the run. That is the general rule behind all seventy-seven listings, and it is easier to apply than a list of forbidden tools. The one line every setup prompt needs makes the case, approval gates covers where to put the stop, and least privilege for bots covers the permissions that make the line enforceable rather than aspirational.

Build from the nearest neighbour when your role has no bot yet

The catalogue leans heavily towards go-to-market work, so plenty of roles have no exact match. That is a naming problem more often than a capability one. Find the bot whose shape matches yours and change the nouns.

Three things must survive the rename: the boundary, the output shape, and the source list. If you change all three you are not adapting a bot, you are writing a new one, and the four layers of a bot system is the better starting point for that. Twenty-five real setups people run daily is worth skimming for a shape close to yours.

Check at thirty days whether the role bot still earns its seat

A role bot is a hire, and hires get reviewed. Put the review in the calendar when you build the bot, because you will not think to schedule it later.

That last row retires more bots than any other, and retiring one is a good outcome rather than a failure. Testing a bot setup before you trust it covers the checks, the seven ways bot setups fail maps symptoms to causes, logs, audits, and receipts covers the evidence, and keeping bot costs predictable plus what Grok Bot costs cover the money side.

Answer the strongest objection: a role bot is a prompt with a logo

The fair objection is that nothing on this page is a product. A role bot is a prompt, a schedule, and two or three connections, and calling it a hire is marketing. That is largely correct, and it is still the right framing for three reasons.

Naming forces scope. "My sales bot" gets asked to do anything sales-adjacent and degrades within a fortnight. "Lead Scout, which researches and ranks and never contacts anyone" refuses the wrong request by construction, because the refusal is in its description rather than in your memory.

Naming forces a boundary. Every listing in this directory declares one act it never takes, and that declaration is what lets you leave it running. An unnamed prompt has no such declaration and drifts to whatever the last conversation implied.

Naming forces review. You can ask whether a named bot still earns its seat. You cannot ask that of a prompt you paste sometimes.

Where the objection genuinely wins: if you only have one recurring job, do not build a roster. Build one bot, run it for a month, and read the starter roster only when the second job appears.

Where the role framing stops being useful

Three cases break this page, and they are worth naming rather than discovering.

The first is a role that is genuinely all judgment. Some jobs are eight hours of decisions with almost no residue: an early-stage designer, a founding salesperson at week two, a therapist. There is nothing here to hand over, and a bot will manufacture busywork to look useful. That is not a bot problem, it is a correct answer.

The second is shared work. Every bot in this directory reports to one person, and in Grok Bot a routine belongs to a single bot rather than to a team, so nothing here is a team-level asset as of writing. If four people need the same output, four people build the same bot, and it drifts. Running a team of bots without chaos is the current best answer, and it is a workaround rather than a feature.

The third is that role boundaries are not tool boundaries. A marketing bot and a sales bot on the same account share one computer and one set of connections, so separating work by role organises your attention without isolating anything. Connecting bots to your tools without handing over everything covers what a connection actually grants, and what one shared computer isolates covers why the separation is not a security boundary.

Keep reading: The starter roster, Bot boundaries, Introducing botskills.sh.

Frequently Asked Questions

Which AI bot should you start with for your role?

Start with whichever bot removes the job you did four times last month with an identical output each time, not the most impressive one available. For most roles that is a research or triage bot rather than a drafting bot, because research is assembly and assembly is what a bot does reliably. Then add one productivity bot that compounds across every role, usually inbox triage or a routing brief, and one personal bot, because household admin is the work you abandon first when you get busy and the easiest place to see whether a setup actually holds up.

Are there AI bots that are safe for recruiting and hiring?

Only for logistics. A recruiting bot can schedule interviews across calendars, reformat information, chase confirmations, and extract facts a candidate stated themselves. It must not score candidates, rank or order them, reject anyone, or contact an applicant, and a total presented as a summary is still a score. Employment decisions are regulated differently across jurisdictions and the rules move, so this is not legal advice. The safe construction is a bot whose description makes the decision impossible rather than one instructed not to make it.

Can a bot handle work for more than one role at once?

It can, and it degrades quickly when it does. A bot asked to cover sales and support drifts because the two jobs want different sources, different output shapes, and different boundaries, and the broadest boundary wins by default. Two narrow bots with separate charters hold their shape far longer than one wide one. Be aware that splitting by role organises your attention rather than isolating anything technically, since bots on the same account share a computer and a set of connections.

How many bots does one person actually need?

Three to five for most people, and the fifth is usually a mistake. Each bot costs attention to review, and a roster you stop reading is worse than no roster because it produces confident output nobody checks. The practical test at thirty days is whether you would notice tomorrow if a given bot stopped running. If the honest answer is no, retire it. Adding the sixth bot almost always means the first three were scoped too narrowly rather than that you found new work.