---
title: "How to Use Claude at Work, and Why It Still Sounds Generic"
description: "Claude handles the tutorial part fine. The output stays generic because the context it needed was said in a meeting and never written down."
canonical: "https://hirekai.ai/blog/how-to-use-claude"
date_published: "2026-09-10"
last_updated: "2026-09-10"
article_type: "guide"
author: "lambert"
---
# How to Use Claude at Work, and Why It Still Sounds Generic

### Key takeaways

- Claude does the tutorial part fine. What breaks after the tutorial is that the model never sat in the meeting the work came from, so a correct prompt still returns generic work.
- Anthropic diagnoses this itself, in lesson 3 of its own free course. What no guide answers is where the context comes from when it was never written down.
- Connecting Gmail, Slack and Drive does not fix it. A connector searches when you ask; it does not give Claude a memory of what you already did on Tuesday.
- [Kai](/blog/ai-executive-assistant) compiles the meeting, the thread and your preferences into the prompt and hands the whole thing to your own Claude, which is a different job from writing a better prompt.

Claude will do almost anything you ask it to, and it will still hand you something flat, because the one thing it needed to know was said out loud on a call and never written down anywhere you can paste from.

Here is the version I use when someone asks me why their output is disappointing. Imagine asking someone for a pasta salad with tomatoes. They make you a pasta salad with tomatoes. You wanted olives in it, and you are allergic to lemon, and you never said either of those things because everyone who cooks for you already knows. Word for word, the answer you got was correct. It just was not the one you wanted. Now imagine that person had read a page about how you eat. Same request, different dinner, and you never had to write the page twice.

![A diagram of three kinds of context. A document, such as brand guidelines or a spec, is handled by Projects and costs one upload. A preference, such as writing in British English or who your team is, is handled by Memory and costs one statement. An event, such as what the client agreed on Thursday or the constraint raised at minute 34, is handled by nothing and costs you, every time.](/blog/images/how-to-use-claude/three-kinds-of-context.webp)

That is the whole problem, and it is not a prompting problem. This guide is for people who pay for Claude, open it every day, and do not write code. If you are here for Claude Code, [there is a paragraph for you at the end](#if-you-write-code), and the context problem below is the same one you have, with a filesystem in front of it.

One reviewer put the symptom better than any of us could. He cancelled after two months:

>

There is no code anywhere in that review. Below: what to set up in your first week, why good prompts go generic, why connecting your tools does not fix it, what people do instead and what it costs them, and one worked example that shows exactly where the manual work hides.

## What Claude actually is, in one paragraph

Claude is a family of models from Anthropic that you reach through a chat window on the web, a desktop app, a phone app and a Chrome extension. It reads what you put in front of it and writes back. It can search the web, read files you upload, make documents and spreadsheets, and query a handful of apps you connect. It has no standing view of your work: it does not sit in your calendar, it does not watch your inbox, and unless you tell it, it does not know that yesterday happened. Everything else in this guide follows from that last sentence.

## The five things worth doing in your first week

The interface rewards about an hour of setup, and most people skip four of these.

**Set your default model, then leave it alone.** The picker sits under the message box. Sonnet 5 is the default on Free and Pro and it is the right daily driver: Anthropic positions it for "coding, writing, analysis, research, and complex problem-solving". Opus 5 is the default on Max and is worth switching to for one genuinely hard task, not for your inbox. Haiku 4.5 is the fast one, available on every plan including Free.

**Find the Effort setting, which is in the same menu and matters more than the model.** It controls how long Claude thinks before answering. Raising it on a hard analysis is the single cheapest quality improvement available. One trap: Effort is not offered on Haiku 4.5, so "drop to Haiku and lower the effort" is one lever, not two.

**Put stable reference material in a Project.** A Project is a workspace with its own uploaded knowledge and its own instructions, and everything in it rides along in every chat inside that Project. Brand guidelines, a positioning doc, the way your team writes: upload once, stop pasting. This is the right home for anything that does not change week to week.

> Anthropic's own two pages disagree about whether Projects exist on the Free plan. The [help centre article](https://support.claude.com/en/articles/9517075-what-are-projects) says "Free users can create a maximum of five projects", and the [pricing page grid](https://claude.com/pricing) marks Projects as a paid feature. Both were live on 10 September 2026. If you are on Free and the button is missing, that is why, and it is not you.

**Look at what Memory already holds.** Memory is on by default on Free, Pro and Max, and it stores things like your role, your ongoing projects, the people in your work and your communication preferences. Open Settings, then Memory, and read the Topics list. You can edit any entry or delete it. Two things worth knowing: Memory changed twice in eight weeks over this summer, and it holds preferences, not events. It will remember that you write in British English. It will not remember what your client agreed to on Thursday.

**Stop counting messages and learn the window.** There is no published message limit anywhere on Anthropic's surfaces. The [pricing FAQ](https://claude.com/pricing) puts it plainly: "Every plan has usage limits that reset on a rolling five-hour session window, and paid plans add weekly limits on top." Long conversations with big attachments burn through a window much faster than short ones, which is the mechanism behind most "why am I limited already" complaints.

| Plan | Price | What it adds |
|---|---|---|
| Free | $0 | Sonnet 5 and Haiku 4.5, web search, Memory, Artifacts, connectors, five-hour window only |
| Pro | $20 a month, or $17 a month billed annually at $200 | Projects, Research, Opus access, Cowork, Claude in Chrome, weekly limits on top |
| Max 5x | $100 a month, monthly only | Five times Pro's session usage, priority access at busy times, early access |
| Max 20x | $200 a month, monthly only | Twenty times Pro's session usage |

Prices read from [claude.com/pricing](https://claude.com/pricing). *Verified 2026-09-10*. Anthropic shipped four consumer models between June and September 2026, so check the page rather than trusting a table.

![Screenshot of the Claude pricing page taken on 10 September 2026 from a European connection, showing three plan cards. Free at 0 euros lists chat on web, iOS, Android and desktop, memory across conversations, connecting Slack and Google Workspace, and remote MCP connectors. Pro at 15 euros a month on annual billing, or 18 euros monthly, lists unlimited projects, Research, more Claude models and Claude for Microsoft 365. Max from 90 euros a month lists a choice of 5x or 20x more usage than Pro, higher output limits, early access and priority at high traffic times.](/blog/images/how-to-use-claude/claude-pricing-page.webp)

Two things that screenshot settles. Anthropic prices in your local currency, so a European reader sees €18 a month where the dollar page says $20, and no guide quoting one figure is right for both. And the Pro column is where "unlimited projects" sits, which is the same page contradicting the help centre in the callout above. When Anthropic's own two pages disagree, the pricing page is the one that decides what you can buy.

## Why a good prompt still comes back generic

Here is the part I got wrong for months. I assumed that when the output was weak, I had written a weak prompt, so I wrote longer prompts. Longer prompts helped a little and then stopped helping, and it took me an embarrassing amount of time to work out why.

Anthropic has the diagnosis in its own free course. [Claude 101, lesson 3](https://academy.claude.com/courses/claude-101/getting-better-results), opens its troubleshooting table with the row "Claude's response is too generic", and the cause it gives is that "your prompt didn't include enough context about your specific situation". So the symptom is not a secret. It is in the official curriculum, and every good guide on the internet says a version of the same thing.

Read what all of them assume, though. They assume the context lives in your head or in a file on your disk, and that typing it out is free. Anthropic's worked example asks the reader to supply the client's identity, the two-week slip, the fact that this is the second delay, and the tone they want, by hand, each time. One of the better independent guides says the quiet part directly: "The more context you give it, the more useful it becomes."

That advice is correct and it has a hole in it. Some context is a document, and Projects handle that. Some context is a preference, and Memory handles that. And some context is an event: what your client actually agreed to on Thursday, which constraint someone raised at minute 34, why the deadline moved. That kind never became a document. Nobody wrote it down, because at the time it was just a thing everyone in the room heard.

> More context is not automatically better either. Anthropic's own documentation says so: "As token count grows, accuracy and recall degrade, a phenomenon known as context rot. This makes curating what's in context just as important as how much space is available." ([Context windows](https://platform.claude.com/docs/en/build-with-claude/context-windows), checked 10 September 2026.) Dumping everything in is not the fix. Bringing the right thing in is.

## Connected is not the same as knowing

The obvious answer is to connect your tools. Claude has connectors for Gmail, Google Calendar, Google Drive, Slack, Notion and a few dozen more, and setting them up takes minutes.

They help, and they do not do what people expect. A connector searches when you ask it to. Anthropic's own wording, in the [Google Workspace connectors article](https://support.claude.com/en/articles/10166901-use-google-workspace-connectors): "Claude only accesses your data when you explicitly ask a question or request an action requiring this information, and retrieves the minimum information needed." Nothing runs on its own, and nothing carries between sessions.

An operations manager for a consumer products brand described the result better than any documentation. He opens his post by noting that most people in that forum are developers and that ops people hit a completely different wall:

>

Access is not the same as awareness. A search tool answers the question you asked. It does not build a picture of your week.

## The three ways people move context by hand

Everyone who works this way seriously ends up doing one of three things, and all three are the same thing wearing different clothes.

**Re-paste the brief every few messages.** A copywriter working across client accounts described the drift: the first two or three iterations hold his tone and his banned phrases, then around iteration five it slides back to generic, and pointing it out buys two more messages. His conclusion is the cost: re-pasting the requirements every three or four messages "takes more time than just writing the thing myself".

**Ask Claude to write a handoff document and carry it to the next chat.** Eoghan Henn, a search and AI consultant who flags his own posts as "NOT about coding", does this on purpose: when he wants to move an idea from a phone conversation into a real work session, he [asks Claude for a downloadable handoff doc](https://www.reddit.com/r/ClaudeAI/comments/1rbiv89/my_current_cowork_setup_workarounds_heavy/) and re-uploads it on the other side. One round trip per switch, forever.

**Write your own session notes so the next session can catch up.** The ops manager above landed here, and named it exactly: "I am doing manual memory management for the AI. That's backwards. That's the thing it was supposed to do for me."

Nobody has measured what this costs in minutes, and we are not going to invent a number. What the evidence supports is the shape: the work is real, it repeats, and the people with the most elaborate systems are the least satisfied with them.

## A worked example: the follow-up after a call

Take the most ordinary task there is. You had a 40-minute call this morning. You owe someone a follow-up email that reflects what was actually agreed.

The cold version is one line: "write a follow-up email about this morning's call with Northwind". You get a competent, empty email. It has to be, because nothing in that sentence tells Claude what was agreed.

The briefed version needs the transcript. So walk the actual steps, because this is where the work hides.

**There is no audio path into Claude.** Not on any surface, not on any plan. The recording sitting in your Downloads folder is inert. Every route ends in a text file that a human exported by hand.

![A diagram titled getting this morning's call into Claude. Step zero states the recording in your Downloads folder is inert because there is no audio or video path into Claude on any surface or plan. Google Meet needs Business Standard or above and takes three steps if you hosted. Zoom needs cloud recording and takes five to six steps. Microsoft Teams only lets organisers download by default, so a participant has to go and ask someone. The closing note says the paste itself is roughly 48,000 tokens, under 5% of a one-million-token window, so space was never the constraint.](/blog/images/how-to-use-claude/no-audio-path.webp)

Sources for each row, because the plan tiers are the part people get wrong: [Google's transcript requirements](https://support.google.com/meet/answer/12849897?hl=en) and its [launch post listing the excluded editions](https://workspaceupdates.googleblog.com/2022/10/google-meet-transcripts.html), [Zoom's cloud-recording rule](https://support.zoom.com/hc/en/article?id=zm_kb\&sysparm_article=KB0061101), and [Microsoft's page saying organisers and co-organisers download the transcript](https://support.microsoft.com/en-us/office/start-stop-and-download-live-transcripts-in-microsoft-teams-meetings-dc1a8f23-2e20-4684-885e-2152e06a4a8b).

Then you paste it in, along with the email thread the request came from, and that part costs nothing. Five to nine manual steps stand in front of every single one of those pastes, after every meeting, for the rest of your working life.

That is the honest state of "just give Claude more context". It is correct advice with an invisible price tag.

## How Kai hands a task to Claude with the context attached

We build [Kai](/blog/ai-executive-assistant), an AI executive assistant that sits on your calendar, your email and your meetings. It is the reason I think about this problem at all, and it takes a specific position on it: Claude is very good at the work, so we do not try to replace it. We try to stop you from being the courier.

When a task in Kai came out of a meeting or an email thread, the **Do it with Claude** button compiles a briefing and hands it over. What rides along is specific, and it is worth naming rather than describing: the task itself with its notes, due date and priority, the linked meeting's summary and its transcript, or the email thread's summary if it came from email instead, plus your preferences and the ids Claude needs to fetch more over [Kai's MCP connector](/kai-mcp). Before it sends, a sheet lets you state what "done" means and what the hard limits are.

Two of those fields are yours and nothing compiles them for you, because nobody else can answer them. **Done means** is the finish line in your own words, so the model knows what it is aiming at and when to stop. **Hard limits** is the rail: drafts only, don't send anything, don't touch the live file. Both travel inside the prompt, so they hold wherever it lands.

![An annotated screenshot of Kai's Review prompt sheet. The Done means field is filled in with a real objective and is labelled as the finish line you write yourself. The Hard limits field below it is labelled as the rail that travels inside the prompt. Underneath, the compiled briefing is already written out, starting with a line naming the person it is helping and a DELIVERABLE section, labelled as something Kai assembles from the task and the meeting before you ever see it.](/blog/images/how-to-use-claude/review-prompt-sheet.webp)

With Kai's connector live in that chat, Claude writes what it produced back into the task's notes, signed so you can see who did it, and the task picks up a chip telling you to go and read it.

There are two places that briefing can land, and the difference is worth understanding because it is not the difference people assume. Sending it to the Claude app gives Claude the briefing and nothing else we hold. Copying it into your own MCP client gives Claude the identical briefing plus every server and file that client already reaches, which is why I use that route for anything that touches our site or our own documents. Kai sends the same prompt both times and cannot tell which world it landed in.

![A diagram showing Kai compiling one briefing from the task, the meeting summary, the transcript or email thread summary, preferences and reference ids, which then splits into two destinations: the Claude app, where Claude works from the briefing plus its own memory and your connectors, and your own MCP client, where Claude works from the same briefing plus every server and file that client already has. Your Notion, Drive and wiki are marked as never included.](/blog/images/how-to-use-claude/two-paths.webp)

**Now the limits, and there are several.** Kai launches the executor, it never runs the work. You fire the prompt yourself and it runs on your own Claude subscription, not ours. There is no run status: Kai finds out the job came back by noticing a new note appeared, so it cannot tell you a job is in progress, and it cannot tell a real result from an apology. It cannot even tell whether the connector is live in that chat, which is why the briefing hedges and says the tools "may be connected".

And the biggest one, which is the same hole this whole article is about, pointed at us. **Kai hands over the context Kai holds. It cannot reach the context you hold.** Your Notion, your Drive, your team wiki: none of it reaches that briefing. If your own Claude has a Notion connector of its own, that seam closes on your side, not ours. Nobody has built it on ours.

## Where this stops working

Three limits that apply to Claude itself, and that most guides skip.

**It runs on a clock, not on a trigger.** Claude can run scheduled work: Cowork tasks fire hourly, daily, weekly or on weekdays on every paid plan. What it does not do, on any plan a normal user has, is fire because a new email arrived or a Slack message landed. If you have read that Claude can react to events, check whether the article was about ChatGPT.

**Your chats may be training data.** On a consumer plan, Anthropic's [privacy policy](https://www.anthropic.com/legal/privacy) says it "may use your Inputs and Outputs to train and improve Anthropic AI models, unless you opt out through your account settings". If you allowed it, that material can sit in training pipelines de-identified for up to five years, and turning the setting off later stops future use but does not pull your data out of models already trained. Conversations flagged for safety review are used regardless. Team and Enterprise plans default the other way.

**Shared chats are public objects.** A shared link is a snapshot anyone holding the link can read. In July 2026, chats shared this way [turned up in Google's index](https://futurism.com/artificial-intelligence/claude-chats-publicly-accessible), reportedly including patient names and internal company documents. Treat the share button the way you would treat publishing a page.

> If you take one setup action from this article, open Settings and check the model training toggle and your Memory topics in the same visit. Both are two clicks, both are things people assume they configured and never did, and one of them decides what happens to five years of your work.

## Getting more out of it, in practice

The pattern that works is boring and it holds up. Put the durable things where they persist: reference documents in a Project, working preferences in Memory. Bring the per-event things in deliberately, because nothing else will. And notice which of the two you are re-typing, because that is your signal. Re-typing a preference means you have a settings problem you can fix in five minutes. Re-typing what happened in a meeting means you have a plumbing problem, and no prompt is going to solve it.

If most of your re-typing is the second kind, the useful reading from here is [how to take meeting notes](/blog/how-to-take-meeting-notes) that a model can actually use, [how to run a meeting](/blog/how-to-run-a-meeting) so the decisions are captured in the first place, and [how to write meeting minutes](/blog/how-to-write-meeting-minutes) that survive the week. If it is the first kind, [organising your day](/blog/how-to-organize-your-day) and [time blocking](/blog/time-blocking-complete-guide) do more for you than any model setting.

For tool comparisons rather than workflow, we keep a list of [AI tools for executive assistants](/blog/ai-tools-for-executive-assistants), a roundup of [the best AI executive assistants](/blog/best-ai-executive-assistants), a set of [AI tools for product managers](/blog/ai-tools-for-product-managers), and a guide to [preparing for meetings with AI](/blog/ai-meeting-prep). The mechanics of turning a call into tracked work live on our [action items](/action-items) page, and the exact handoff described above is written up as a recipe: [hand off a task to Claude](/use-cases/hand-off-a-task-to-claude).

## If you write code

Claude Code is a different product with a different answer to this problem, and it solves it by living in a directory: your repository, your files and your project notes are simply there. That is the version of "give it context" that actually works today, and it works because the context was already text on a disk. Six of the twenty results ranking for this search are about Claude Code, which is why this article says non-developer in its first paragraph. If that is what you came for, Anthropic's own documentation is better than anything we would write.

## FAQs

## Frequently asked questions

### Is Claude free to use?

Yes, with real limits. The free plan gives you Sonnet 5 and Haiku 4.5, web search, Memory, Artifacts and connectors, metered by a rolling five-hour window with no published message count. Opus, Projects on some readings of Anthropic's own pages, Cowork and Claude in Chrome need Pro at $20 a month, or $17 a month if you pay $200 up front for the year.

### Which Claude model should I pick for everyday work?

Sonnet 5, and stop thinking about it. It is the default on Free and Pro and Anthropic positions it for writing, analysis and research. Switch to Opus 5 for one genuinely hard task in a session rather than as a standing setting. The lever most people miss is not the model at all: it is the Effort control in the same menu, which decides how long Claude thinks before it answers.

### Why does Claude forget what we discussed in an earlier chat?

Because each conversation starts fresh by design. Memory carries preferences and standing facts about you between chats, not the content of past conversations. Projects carry the documents you uploaded into that Project. Neither carries what happened in a specific past chat, which is why people end up re-uploading the same file or writing themselves handoff notes.

### Can Claude read my Gmail and my calendar?

It can search them when you ask, once you connect them. What it will not do is watch them. Anthropic's own documentation says Claude accesses connector data only when you explicitly ask a question requiring it, and retrieves the minimum it needs. Nothing fires on its own, so a connected inbox does not mean Claude knows what happened in it this week.

### Can I give Claude a recording of a meeting?

Not directly. There is no audio or video path into Claude on any surface or plan, so the file itself is useless to it. You need a text transcript, which means exporting one from Google Meet, Zoom or Teams first, and the number of clicks depends on whether you hosted the call. If you were a Teams participant, the default permissions may not let you download it at all.

### Does using Claude at work put company information at risk?

It depends on your plan and one setting. On consumer plans, Anthropic may use your chats for model training unless you opt out, and material you allowed can be retained de-identified for up to five years. Team and Enterprise default to not training on your content. Shared chat links are snapshots readable by anyone holding the link, and some have been indexed by search engines.

## Sitemap

See the full [sitemap](https://hirekai.ai/sitemap.md) for all pages.