AI Use Case Patterns: 38 Tasks and Whether Each Should Be a Prompt, Skill, MCP Server, Agent or App

Every pattern below went through the same eight questions as the decision tool, and each verdict is exactly what the tool returns for those answers. A test in the site's code checks that on every build, so the library and the tool can't drift apart. Open any card for the starting point, what would push it up or down the ladder, and a link that loads its answers into the tool so you can change one and watch the verdict move.

TeamSearch

38 patterns

PromptMarketing
Alt text for a site's image backlog
A one-time cleanup. Upload the images in batches to an assistant that reads images and ask for alt text under 125 characters. Building anything costs more than the job.
Starting point and what would change it
Start
Batch 20 images per message, give the page context for each, and ask for a two-column table: filename, alt text.
Climbs if
If every new upload needs alt text, it belongs in your CMS upload flow.
Drops if
Already at the bottom.
See the tool's reasoning →
PromptLegal
Explaining one clause you don't understand
A one-off question. Paste the clause and ask what it means for you, what's unusual and what to ask the other side.
Starting point and what would change it
Start
Paste the clause and the surrounding definitions, say which side you're on, and ask for plain-English meaning, risks and two questions to raise.
Climbs if
If you review these weekly, build the playbook skill above.
Drops if
Already at the bottom.
See the tool's reasoning →
PromptData
Cleaning up a messy spreadsheet export
A one-time cleanup. Assistants with code execution can read the file, fix it and hand back a clean one.
Starting point and what would change it
Start
Upload the file, describe the target columns and formats, and ask it to report every row it changed or dropped.
Climbs if
If the same export arrives every week, save the cleaning steps as a skill with a script.
Drops if
Already at the bottom.
See the tool's reasoning →Rated skills: Spreadsheets & Excel analysis (Skillspace, ours)
Saved promptOperations
Weekly status update to your manager
One person, once a week, same shape. The value is in not re-typing the format. That's a saved prompt, not a build.
Starting point and what would change it
Start
Save the prompt with your headings (shipped, blocked, next) as custom instructions in a project, and paste your rough notes in each Friday.
Climbs if
If the whole team writes the same update, make it a skill. If it should pull your tickets and commits, add a connector.
Drops if
If you write it twice a quarter, type it fresh.
See the tool's reasoning →Rated skills: Standups & status reporting (Skillspace, ours)
Saved promptMarketing
Social posts from a new blog article
One marketer, weekly, a short voice guide as reference. A project with the guide attached and a saved prompt is all it takes.
Starting point and what would change it
Start
Save the prompt: platform, length limits, three hook styles, a link placement rule, and your voice guide as an attached file.
Climbs if
If a team posts across many channels with approvals and scheduling, use your social tool's AI features or a skill.
Drops if
If it's an occasional post, paste the article and ask.
See the tool's reasoning →Rated skills: Social media content (Skillspace, ours)
Saved promptPeople
Interview question sets
An occasional, personal task. Save the prompt with your competencies and format; there's nothing to share or connect.
Starting point and what would change it
Start
Save a prompt that takes the role and level and returns questions per competency, with what a strong answer includes.
Climbs if
If every interviewer needs consistent question banks per role, make it a team skill.
Drops if
For one interview, just ask.
See the tool's reasoning →Rated skills: Résumé screening & interviewing (Skillspace, ours)
SkillOperations
Meeting transcript to summary and action items
A fixed procedure the whole team repeats every week: same sections, same owner-and-date format for actions. A skill gives everyone the identical output without anyone re-explaining the format.
Starting point and what would change it
Start
Write a SKILL.md with your section order, the action-item format (owner, task, due date) and one good past example. Paste the transcript in.
Climbs if
If transcripts should be pulled from the meeting tool and actions filed in your tracker automatically, add connectors, then an agent on the 'meeting ended' event.
Drops if
If only you take notes, a saved prompt does the same job. Check your meeting tool's built-in summaries first; they may already be enough.
See the tool's reasoning →Rated skills: Meetings, agendas & notes (Skillspace, ours)
SkillSales
Proposal and quote drafting
Proposals follow your structure, pricing rules and case studies, which change rarely. A skill carries the template, the price list and the do-not-promise list, and outputs a document.
Starting point and what would change it
Start
Put your template, rate card and three best past proposals in a skill folder, with steps for scope, timeline, price and assumptions.
Climbs if
If line items should come from the CRM deal, add a connector.
Drops if
If you write two a year, a saved prompt with last year's proposal attached is fine.
See the tool's reasoning →Rated skills: Proposals & quotes (Skillspace, ours)
SkillMarketing
Rewriting copy in your brand voice
Several people rewrite copy daily against the same style guide and examples. This is the textbook skill: reference files plus rules, loaded whenever someone asks for a rewrite.
Starting point and what would change it
Start
Write the voice rules as do and don't pairs, add five before-and-after examples, and list banned words. Test on three real drafts.
Climbs if
If every CMS draft should be checked automatically on save, an agent in the publishing flow.
Drops if
A solo writer can keep the style guide in a project's custom instructions.
See the tool's reasoning →Rated skills: Brand voice & messaging (Skillspace, ours)
SkillMarketing
SEO content brief for a writer
A team produces briefs weekly in the same format, with web research each time. A skill sets the method (search intent, outline, questions to answer, sources) and runs in an assistant with web search.
Starting point and what would change it
Start
Write a skill with your brief template and the research steps, and test it on a keyword you already rank for.
Climbs if
If it needs your rank-tracking or keyword tool data, add that tool's connector.
Drops if
For one article, ask a research-capable assistant directly.
See the tool's reasoning →Rated skills: SEO content writing (Skillspace, ours)
SkillLegal
Contract redlines against your playbook
Your negotiation playbook (fallback positions, must-haves) is stable reference material and the output is a marked-up document. A skill carries both; a lawyer still signs off.
Starting point and what would change it
Start
Turn the playbook into a clause-by-clause table: our position, acceptable fallback, walk-away. Test on two signed contracts you know well.
Climbs if
If contracts come in through a CLM or e-signature tool, a connector saves the upload step.
Drops if
For a single contract, paste it with your three non-negotiables and ask.
See the tool's reasoning →Rated skills: Contracts & agreements (Skillspace, ours)
SkillEngineering
Code review checklist
Your standards (naming, error handling, security rules) are fixed reference material applied many times a day. A skill in the coding agent keeps reviews consistent.
Starting point and what would change it
Start
Write the checklist as a skill with examples of good and bad code from your own repo, and tell it to cite file and line for each finding.
Climbs if
Run it on every pull request automatically: an agent in CI.
Drops if
Several code hosts and IDEs ship AI review now; check yours before writing your own.
See the tool's reasoning →Rated skills: Change & PR review (Skillspace, ours)
SkillEngineering
Release notes from merged changes
Coding agents already read the repository, so no connector is needed. What's missing is the format: grouping, audience, what to leave out. That's a skill.
Starting point and what would change it
Start
A skill that reads commits since the last tag, groups them (new, improved, fixed), rewrites for customers and drops internal changes.
Climbs if
Generate them on every release tag: an agent in the release pipeline.
Drops if
For a one-off release, ask the coding agent directly.
See the tool's reasoning →Rated skills: Changelogs & release notes (Skillspace, ours)
SkillEngineering
Writing unit tests in your conventions
Coding agents write tests already; what they don't know is your framework choices, fixtures and naming. A skill teaches them once.
Starting point and what would change it
Start
Add a testing skill with one exemplary test file, the fixture helpers, and rules such as 'no mocks for the database'.
Climbs if
If tests must be generated for every change unattended, it's an agent in CI.
Drops if
A line in your project's agent instructions file may be enough.
See the tool's reasoning →Rated skills: Test authoring & frameworks (Skillspace, ours)
SkillProduct
Translating product interface strings
A glossary and tone rules per language are stable reference material, and the output is a strings file. A skill keeps terminology consistent across releases.
Starting point and what would change it
Start
A skill with the glossary, do-not-translate terms, length limits and the file format, tested on last release's strings.
Climbs if
Run it on every pull request that adds strings: an agent in CI.
Drops if
Your localisation platform may already offer machine translation with a glossary.
See the tool's reasoning →Rated skills: Translation & localisation (Skillspace, ours)
SkillPeople
Job descriptions in your format
Hiring managers across the team need the same structure, levels and inclusive-language rules. A shared skill holds them.
Starting point and what would change it
Start
A skill with your levelling guide, the template, a banned-phrases list and two strong past descriptions.
Climbs if
If they should post to your applicant tracking system directly, add its connector.
Drops if
If one recruiter writes them, a saved prompt with the template works.
See the tool's reasoning →Rated skills: Job descriptions & sourcing (Skillspace, ours)
SkillPeople
New-hire onboarding documents
Each new hire needs the same set of documents tailored to their role. The templates and policies are stable reference material; the output is files.
Starting point and what would change it
Start
A skill with the welcome letter, first-week plan and role checklists as templates, filled from the role and start date.
Climbs if
If it should create accounts and tasks in your HR system, that's connectors and probably your HR platform's own onboarding flows.
Drops if
Your HR platform may already generate onboarding checklists.
See the tool's reasoning →Rated skills: Employee onboarding & offboarding (Skillspace, ours)
SkillPeople
Answering questions from the policy handbook
Many people ask the same handbook questions. You don't need a retrieval app: a skill or project with the handbook attached answers with citations inside the assistant people already use.
Starting point and what would change it
Start
Attach the current handbook, tell it to quote the section it relied on and to say 'ask HR' when the handbook is silent.
Climbs if
If the handbook changes weekly or lives across many drive folders, connect the drive. If employees outside the assistant need it, it's an app.
Drops if
For your own occasional question, search the handbook.
See the tool's reasoning →Rated skills: HR policies, payroll & employment compliance (Skillspace, ours)
SkillResearch
Market or competitor research brief
Open-ended research, but it only needs the web. A skill sets the method (sources to prefer, how to cite, the brief's structure) and runs in an assistant with research tools.
Starting point and what would change it
Start
A skill with the brief structure, source quality rules and a requirement to link every claim.
Climbs if
If it should pull from paid databases or your CRM, add connectors.
Drops if
For a one-off question, use your assistant's deep research mode.
See the tool's reasoning →Rated skills: Research reports & synthesis (Skillspace, ours)
SkillResearch
Literature review for a paper
Repeated, open-ended and producing a document with citations. A skill fixes the search method and citation style; the assistant's research tools do the reading.
Starting point and what would change it
Start
A skill with inclusion criteria, the citation style and a rule to verify every reference exists.
Climbs if
Connecting a reference manager saves the export step.
Drops if
For one quick scan, ask a research assistant directly and check the citations.
See the tool's reasoning →Rated skills: Literature review & citations (Skillspace, ours)
SkillFinance
Monthly investor update
Monthly, one author, but it produces the same document each time. A skill with last month's update as the template keeps the structure and tone steady.
Starting point and what would change it
Start
Template: highlights, metrics table, lowlights, asks. Paste your numbers and notes each month.
Climbs if
If metrics should come straight from your finance tools, add connectors.
Drops if
If you're happy pasting last month's into a chat, a saved prompt works.
See the tool's reasoning →Rated skills: Fundraising, investors & corporate finance (Skillspace, ours)
MCP / connectorData
Weekly KPI report from your analytics
The numbers live in a warehouse or analytics tool the model can't see, and copy-pasting them each Monday is where errors creep in. A read-only connector fetches them; a skill fixes the layout and commentary rules.
Starting point and what would change it
Start
Connect a read-only database or analytics MCP server with a service account, and write a skill that names the exact metrics, comparisons (week over week) and chart types.
Climbs if
Once the format is stable, schedule it: a plain scheduled workflow or an agent posts it every Monday at 9am.
Drops if
If someone already exports a CSV, attach it to a skill and skip the connector.
See the tool's reasoning →Rated skills: Business analysis & reporting (Skillspace, ours)
MCP / connectorSales
Sales call prep brief
A good brief needs the CRM history, open deals and recent emails, which the model can't see. A CRM connector plus a skill for the brief's layout gets reps a consistent one-pager in a minute.
Starting point and what would change it
Start
Turn on your CRM's official connector read-only, then write a skill: company snapshot, people, deal history, open questions, three suggested talk tracks.
Climbs if
Generate it automatically an hour before every external meeting on the calendar: that's an agent.
Drops if
If reps paste the account notes in themselves, a shared skill is enough.
See the tool's reasoning →Rated skills: Discovery & demos (Skillspace, ours)
MCP / connectorSupport
Support reply drafts from your help centre
Good replies need the ticket history, the customer's plan and the current help articles. A help desk connector gives the model that context; a skill sets tone and escalation rules.
Starting point and what would change it
Start
Check your help desk's own AI features first. If they fall short, connect the help desk read-only and write a skill with tone, refund limits and when to escalate.
Climbs if
If replies should go out without an agent reading them, you need a full system with confidence thresholds and audits.
Drops if
Most help desks now ship AI reply drafting. If yours does, configure it before building.
See the tool's reasoning →Rated skills: Ticket handling & responses (Skillspace, ours)
MCP / connectorEngineering
Incident postmortem first draft
The timeline lives in chat channels, alerts and logs. Connectors to those let the model assemble the timeline; a skill sets your blameless template.
Starting point and what would change it
Start
Connect your chat and monitoring tools read-only and write a template skill: summary, impact, timeline, causes, actions with owners.
Climbs if
If postmortems must be filed and tracked automatically, wire it into incident tooling.
Drops if
If someone pastes the channel export in, a skill is enough.
See the tool's reasoning →Rated skills: Observability & incident response (Skillspace, ours)
MCP / connectorProduct
Product feedback synthesis
Feedback sits in tickets, surveys and call notes. Connectors to those sources let the model tag themes across all of it instead of the sample someone remembered to paste.
Starting point and what would change it
Start
Connect the two biggest sources read-only, write a skill with your theme taxonomy, and ask for counts with three quotes per theme.
Climbs if
Continuous tagging as feedback arrives is an agent.
Drops if
A monthly survey export attached to a skill covers most teams.
See the tool's reasoning →Rated skills: Customer feedback analysis (Skillspace, ours)
MCP / connectorPeople
Résumé screening against criteria
Candidate files live in the applicant tracking system, so a connector is needed to read them. The stakes are high, so the model summarises evidence against your criteria and a person decides.
Starting point and what would change it
Start
Write criteria as observable evidence, have the model quote the résumé for each one, and keep the decision with the hiring manager.
Climbs if
Automating rejections is where this becomes regulated in several jurisdictions. Don't let it decide.
Drops if
For a handful of candidates, paste the criteria and résumés in.
See the tool's reasoning →Rated skills: Résumé screening & interviewing (Skillspace, ours)
MCP / connectorData
Answering business questions from the database
People across the company ask data questions daily, and the answer is in the warehouse. One managed, read-only database connector plus a skill that explains your schema beats everyone exporting CSVs.
Starting point and what would change it
Start
Create a read-only role on a reporting replica, connect it, and write a skill with table descriptions, metric definitions and common joins.
Climbs if
If customers will query it, it's a product feature with access control per tenant.
Drops if
If one analyst answers the questions, they can use the connector directly without rollout.
See the tool's reasoning →Rated skills: SQL analytics & reporting (Skillspace, ours)
AgentSales
Inbound lead triage and routing
Leads arrive all day and need enriching, scoring against your ideal customer profile and routing, with nobody pressing go. That's a trigger, live data and judgment: an agent loop.
Starting point and what would change it
Start
Write the scoring rubric first and test it on last month's 50 leads. Then run the agent in suggest-only mode, posting its routing to a channel, before letting it assign owners.
Climbs if
If routing errors cost deals and you need a queue, overrides and reporting, it becomes part of your CRM system.
Drops if
If one person reviews leads each morning, a skill plus a CRM connector does the scoring on demand.
See the tool's reasoning →Rated skills: Lead qualification (Skillspace, ours)
AgentSupport
Churn-risk alerts for account managers
It should watch usage, tickets and billing on its own and flag accounts with a reason, every day. Live data, judgment and a trigger.
Starting point and what would change it
Start
Define three warning signals from past churned accounts, check the agent against last quarter's churn, then run it in alert-only mode.
Climbs if
If alerts feed playbooks, tasks and reporting, it's part of your customer success platform.
Drops if
A weekly person-run review with a connector and a skill gets most of the value.
See the tool's reasoning →Rated skills: Account health & churn prevention (Skillspace, ours)
AgentEngineering
Triage of incoming bug reports
Reports arrive constantly and need deduplicating, labelling and assigning in the tracker. It acts in another system on a trigger: an agent.
Starting point and what would change it
Start
Start label-only (no assigning), log every decision, and review a week of them before allowing it to assign owners.
Climbs if
If triage decisions feed SLAs and on-call paging, it belongs inside your incident tooling.
Drops if
A person-run daily skill with a tracker connector does the same work with a human pressing go.
See the tool's reasoning →Rated skills: Bug triage & fixing (Skillspace, ours)
AgentProduct
Competitor pricing watch
It should check competitors' pricing pages every week by itself and report changes. That's a scheduled loop over live web data.
Starting point and what would change it
Start
List the URLs, save a baseline, and schedule a weekly diff with a one-paragraph summary of what changed.
Climbs if
If it feeds pricing decisions with history and dashboards, store results in a real system.
Drops if
The steps are fixed, so a scheduled workflow (fetch pages, diff, one model call to summarise) is cheaper than a full agent.
See the tool's reasoning →Rated skills: Competitive intelligence (Skillspace, ours)
App / systemFinance
Three-way invoice matching (PO, receipt, invoice)
High volume, real money and an audit trail. Matching needs records, exception queues, approvals and logs; the model is one step (reading messy invoices) inside a system with rules.
Starting point and what would change it
Start
List your exception types from last quarter. Price the AP automation features in your ERP first. If you build, put the model behind an approval queue with every match logged.
Climbs if
You're at the top; the question is buy versus build.
Drops if
Most ERPs and AP tools now include matching. If yours does, configure it and only build the exception handling it lacks.
See the tool's reasoning →Rated skills: Invoicing & accounts receivable (Skillspace, ours)
App / systemSales
Personalised cold outreach at scale
Sending to prospects automatically risks your domain reputation, anti-spam rules and your brand. Anything that sends unreviewed needs suppression lists, rate limits, approvals and logs, which is a system.
Starting point and what would change it
Start
Run it draft-only for a month: a skill with your voice and research steps, a person approving every send. Measure reply rates before automating anything.
Climbs if
Already at the top. Most sales engagement tools have this built in; buy before you build.
Drops if
Have the model draft and a person send, and it drops to a skill plus a CRM connector.
See the tool's reasoning →Rated skills: Cold outreach & sequences (Skillspace, ours)
App / systemSupport
Customer-facing support assistant
Customers use it, it answers unreviewed, and it runs constantly. That needs grounding in your docs, guardrails, handoff to humans, analytics and abuse limits. It's a product.
Starting point and what would change it
Start
Compare your help desk vendor's AI agent product with building. If you build, start from a test set of 200 real past tickets and measure deflection and wrong-answer rates.
Climbs if
At the top already.
Drops if
Start with an internal agent-assist version your support team uses; it'll show where the docs fail before customers do.
See the tool's reasoning →
Don't buildFinance
Categorising expenses and bank transactions
Accounting software already learns categorisation rules from your past choices. A model adds little to a rules engine you already pay for.
Starting point and what would change it
Start
Open your accounting tool's bank rules, set up rules from the last three months of transactions, and review only the uncategorised ones.
Climbs if
If your categories depend on reading receipts or contracts the software can't parse, a connector plus a skill can draft the hard cases.
Drops if
Already at the bottom: switch on the bank rules and review the exceptions weekly.
See the tool's reasoning →
Don't buildOperations
Drafting replies in your own inbox
Major email clients now ship AI reply drafting built in. Turning it on beats building anything.
Starting point and what would change it
Start
Turn on your email client's AI drafting, and add a few lines of personal tone instructions where it supports them.
Climbs if
If you need replies that use your CRM or calendar, an assistant with those connectors does it.
Drops if
Already at the bottom.
See the tool's reasoning →Rated skills: Email & inbox management (Skillspace, ours)
Don't buildOperations
Scheduling meetings across people
Scheduling links and calendar assistants already solve this. An agent that negotiates times over email is more failure-prone than a link.
Starting point and what would change it
Start
Set up a scheduling link with buffers and working hours; use your calendar's find-a-time feature for internal meetings.
Climbs if
If scheduling depends on rules your tools can't express (rooms, equipment, travel), connectors to the calendars plus a skill.
Drops if
Already at the bottom.
See the tool's reasoning →Rated skills: Scheduling, calendar & logistics (Skillspace, ours)

Loading the tool…

How to read a verdict

A verdict is the cheapest rung that works for the typical version of the task. Your version may differ, and the "climbs if" and "drops if" lines say how. Meeting notes are a skill for a team that wants identical action-item formats, and a saved prompt for one person. Lead triage is an agent because leads arrive all day and nobody should have to press go, but if one person reviews them every morning it's a skill plus a CRM connector.

Three patterns land on "don't build it", and they're the ones people most often build anyway. Expense categorisation, inbox reply drafts and meeting scheduling are all solved by tools most teams already pay for. The work is switching the feature on and configuring it, which is less fun than building something and much cheaper.

What the 38 patterns show

The split is lopsided. Fifteen tasks are skills, seven need a connector or MCP server, four are agents, three each are plain prompts, saved prompts and apps, and three are "don't build". Two things drive most of the climbs. Live data is the big one: as soon as the answer depends on what's in the CRM, the warehouse or the help desk, copying it into a chat stops being reliable and the floor rises to a connector. The other is autonomy. Only the tasks that must run with nobody pressing go (lead triage, churn alerts, bug triage, competitor pricing) become agents, and even then two of them would do fine as a simpler scheduled workflow.

Apps are rare. The three that need one are invoice matching, a customer-facing support assistant and cold outreach that sends without review, and each has the same profile: high volume, a high cost of error, and either customers or money on the other side. If a task doesn't have that profile and someone is proposing an app, ask which answer forced it there.

About the skill links

Where a card links to "rated skills", it goes to a category on Skillspace, a directory of public agent skills that we run. Skillspace indexes skills from GitHub and grades them with a published, static scoring method; its own methodology page says the grade can't tell you whether a skill works, only how well it's built and whether its files trip a safety scan. So treat a linked skill as a starting point to read and adapt, never as something to install blind. Anthropic's own guidance says the same for any skill you didn't write. Other places to look, and how they compare, are on the registry comparison.

Questions people ask

Which AI use cases need a custom app?

In our library, only three of 38 do, and they share a trait. Customers use them directly, they run hundreds of times a day, or a wrong output moves money with nobody reviewing it. Invoice matching, a customer-facing support assistant and automated cold outreach all land there.

What is the most common verdict?

A skill. Fifteen of the 38 tasks are a stable procedure with reference files that a team repeats, which is exactly what a SKILL.md folder captures. Most of them take an afternoon to write.

Are the Skillspace links sponsored?

No, they're ours. We run Skillspace, a directory that rates public agent skills, and we link to its category pages so you can find rated skills for a pattern. We don't link to it where it has nothing relevant.

Alt text for a site's image backlog

Explaining one clause you don't understand

Cleaning up a messy spreadsheet export

Weekly status update to your manager

Social posts from a new blog article

Interview question sets

Meeting transcript to summary and action items

Proposal and quote drafting

Rewriting copy in your brand voice

SEO content brief for a writer

Contract redlines against your playbook

Code review checklist

Release notes from merged changes

Writing unit tests in your conventions

Translating product interface strings

Job descriptions in your format

New-hire onboarding documents

Answering questions from the policy handbook

Market or competitor research brief

Literature review for a paper

Monthly investor update

Weekly KPI report from your analytics

Sales call prep brief

Support reply drafts from your help centre

Incident postmortem first draft

Product feedback synthesis

Résumé screening against criteria

Answering business questions from the database

Inbound lead triage and routing

Churn-risk alerts for account managers

Triage of incoming bug reports

Competitor pricing watch

Three-way invoice matching (PO, receipt, invoice)

Personalised cold outreach at scale

Customer-facing support assistant

Categorising expenses and bank transactions

Drafting replies in your own inbox

Scheduling meetings across people