Forget the sci-fi version. An agent, in the way that is useful to you this week, is a saved set of instructions that does one job the same way every time, with the inputs it needs already attached, without you re-explaining anything. That is it. The reason eight small ones beat one giant assistant is context: an agent that only reviews drafts is not distracted by your inbox rules, and an agent with fresh eyes catches things the agent that wrote the draft is blind to.
Every agent below is built from the same three pieces, so once you have made one, the other seven take about fifteen minutes each.
Gathers real sources on a topic and hands back a clean brief with links, so you never start from a blank page again. This is the one to build first because six of the other seven get better when they are fed a brief instead of a vague request.
Build me a reusable Claude skill called "researcher" that turns a topic into a decision-ready brief.
The description should say to use it whenever I ask for research on a topic, a market, a person or a competitor. The instructions should tell you to search the web, gather at least five credible sources, prefer primary sources and official documentation over blog summaries, and note the publication date of everything so I can see how current it is. Then produce a one-page brief in this order: what I need to know in three bullets, the key facts with a source link next to each one, the disagreements between sources, and what is still unknown.
Add a hard rule that you never state a number, quote or claim without a link next to it, and that you say "I could not verify this" rather than filling a gap. Before writing the SKILL.md, ask me what field I work in and what I usually use research for, so the brief format matches how I actually work.
Takes the researcher's brief and writes the first draft in your voice, not in default AI voice. The trick is feeding it three or four things you have genuinely written so it has something to copy other than the internet's average tone.
Build me a reusable Claude skill called "drafter" that turns a research brief into a first draft in my voice.
First, ask me to paste three or four things I have actually written, then analyse them and describe my voice back to me concretely: average sentence length, how formal I am, the words and constructions I use often, the ones I never use, how I open and how I close. Put that description directly into the SKILL.md so it travels with the skill.
The instructions should tell you to work only from the facts in the brief you are given, never to invent statistics or examples, to mark anything you are unsure about with a bracketed note rather than smoothing over it, and to hand back one draft rather than three options. Add my hard style rules: no em dashes, sentence case, no filler openers like "in today's fast-paced world". Finish the SKILL.md with a line telling you to list, at the end of every draft, the three weakest sentences and why.
Reviews the draft with no memory of writing it. This is the one people skip and it is the one that changes the output most, because the model that wrote something is the worst possible judge of it. In the Claude app, open a brand new chat and paste the draft in cold. In Claude Code, a subagent gets a genuinely fresh context window by design.
Build me a reusable Claude skill called "critic" that reviews a draft as if it has never seen it before.
The instructions must open by telling you that you did not write this and you have no loyalty to it, and that your job is to find what is wrong rather than to encourage me. Review in four passes and report each separately: factual claims that lack a source or look wrong, logical gaps where a step is missing between two points, anything that reads as generic and could have been written about any business, and the specific sentences a busy reader would skim past.
For each issue, quote the exact line and give the fix, not general advice. End with one honest verdict line: ship it, fix these three things first, or start again. Add a rule that you never open with praise and never end with a compliment sandwich.
Sorts your inbox into what needs you and what does not, and drafts the replies for the ones that do. Connect Gmail under Customize, then Connectors. Be aware of the real limit: the Gmail connector can search, summarise and create drafts, but it cannot send, which is a feature rather than a flaw for an agent handling your email.
Build me a reusable Claude skill called "inbox-triage" that processes my email each morning using my connected Gmail.
The instructions should tell you to read the last 24 hours of unread mail and sort every message into four buckets: needs a decision from me, needs a reply I can draft for you, needs no action but I should know, and noise. Present it as a short list, decisions first, with one line per email saying who it is from and what they want. Never quote long passages back to me.
Then draft replies for the second bucket in my voice, keeping them under 120 words, plain and direct, and save them as drafts rather than sending anything. Ask me first for the three or four senders I always want flagged as urgent and the topics I never want auto-drafted, such as anything about money, contracts or bad news, and write those into the skill as hard exceptions I handle myself.
Checks the same handful of rivals every week and tells you only what changed. The value is in the discipline of a fixed list and a fixed checklist, because "keep an eye on competitors" is a task nobody ever actually does.
Build me a reusable Claude skill called "competitor-watch" that runs a weekly check on a fixed list of competitors.
Ask me for the list first, up to six, with their websites and social handles, and store them inside the SKILL.md so I never re-type them. The instructions should tell you to check each one for the same things every week: pricing changes, new or removed offers, changes to their homepage message, new content themes, and anything they are running ads on that is publicly visible.
Report only what changed since last time, in one line per competitor, with a link to the evidence. If nothing changed for a competitor, say "no change" rather than padding. Finish with a section called "what this means for me" with at most three observations and one suggested action, and a rule that you never recommend copying a competitor directly, only what the change signals about the market.
Builds tomorrow out of your actual priorities and your actual calendar, instead of a list you write at 11pm and ignore. Connect Google Calendar so it plans around the meetings you really have. Calendar access can create and update events, so this one can hold time for you as well as suggest it.
Build me a reusable Claude skill called "planner" that builds tomorrow's plan from my priorities and my connected calendar.
The instructions should tell you to read tomorrow's calendar first, work out how many free hours actually exist after meetings and travel, and then fit my priority list into that real number rather than into a fantasy empty day. Put deep work in the longest unbroken block, batch small admin into one slot, and leave at least one buffer hour untouched.
Hand back a simple time-blocked list with the single most important task marked, plus a short "not today" section for the things I am consciously dropping so they do not sit in my head. Ask me what time I start, when my energy is best, and my hard commitments, and write those into the skill. Add a rule that you never plan more than five real tasks in a day and that you say so plainly when my list does not fit.
Categorises your spending, spots the subscriptions you forgot, and flags anything that looks odd. It organises records, it does not do your taxes, and that line matters. Give it a folder of statements or receipts in a Project and let it work from real files.
Build me a reusable Claude skill called "bookkeeper" that organises spending from statements or receipts I give you.
The instructions should tell you to read every transaction, assign each one a category from a fixed list we agree up front, and produce a monthly summary table with totals per category, the biggest three line items, and the change versus the previous month. Then list every recurring charge you can detect with its frequency and annual cost, because forgotten subscriptions are the point of this.
Flag anything unusual: duplicate charges, a subscription that has gone up in price, a category that has moved more than 30 percent, and anything you cannot categorise confidently. Never guess a category silently, put it in "unclear" and ask. Add a rule that you always state that this is record keeping and not tax or financial advice, and that I check the numbers before they go anywhere official.
Answers questions from your own notes, documents and saved research instead of from the internet. This is the one that quietly gets more valuable every month, because its usefulness scales with how much you put into it. A Project with your files uploaded is the natural home.
Build me a reusable Claude skill called "librarian" that answers questions strictly from my own uploaded notes and documents.
The core rule, written at the top of the instructions: answer only from the files provided. If the answer is not in my documents, say "that is not in your notes" and stop, rather than filling the gap from general knowledge. If you do add outside context because I explicitly ask for it, label that part clearly as outside the documents.
Every answer should cite which document and which section it came from so I can go and check. When my notes contradict each other, show both versions with their sources and their dates instead of picking one. Finish each answer with a short "related things in your notes" list of up to three items, because half the value is remembering what I already wrote down and forgot.
Agent Skills, official overview
Creating custom skills (Claude Help Center)
What are Projects (Claude Help Center)
Use Google Workspace connectors (Claude Help Center)
Claude Code subagents (official docs)