Supply Context: Files, Projects, Memory, Connectors, Skills
This content was developed with AI assistance and is regularly reviewed for accuracy.
Snapshot: October 2026. Model names, prices, and product features change monthly; What's New in AI is the maintained chronology.
The biggest improvement most people can make to their AI results has nothing to do with wording. It is giving the assistant the material it needs: the actual document instead of a description of it, the standing facts about your job instead of a fresh explanation every session, live access to the calendar instead of a pasted schedule. Anthropic's engineers call this context engineering and describe it as the natural progression of prompt engineering. This lesson walks through the five places context can live and how to use each one well.
Learning focus
- Attach files, links, and images instead of pasting or describing them
- Set up a project with standing instructions so you stop re-explaining your situation
- Turn memory on, review what it stores, and keep the wrong things out
- Connect the accounts your task needs with the narrowest useful scope
- Package how-to knowledge as a reusable skill, and know when a skill beats an instruction
Why context beats wording
The goal, in Anthropic's phrase, is the smallest set of high-signal tokens that makes a good result likely. "Smallest" does not mean short; a 40-page source document is high-signal if the task is about that document. It means you are not padding the conversation with your own retyping, your own summaries, or instructions the model does not need.
Think about what a new colleague would need on day one: the files, the folder where things live, the standing rules, access to the systems, and a walkthrough of how your team does the recurring tasks. Context supply is exactly that list, mapped onto features of the assistant you use.
The five layers
| Layer | What lives there | Lasts for | Example |
|---|---|---|---|
| Conversation | Files, links, images you attach now | This chat | The contract you want reviewed |
| Project | Standing instructions plus a set of files | Every chat in the project | Your team's style guide and last three reports |
| Memory | Facts and preferences the assistant learned about you | Across all chats | Your role, your audience, how you like summaries |
| Connectors | Live access to your mail, calendar, drive, chat tools | Until you revoke it | This week's calendar, the latest thread |
| Skills | Packaged how-to knowledge for a recurring task | Whenever the task comes up | "How we write a weekly status report" |
Beginners use the first layer and stop. The remaining four are where the time savings and the quality gains are.
Attach, do not paste
Paste a spreadsheet into a chat and the model gets a wall of text with the structure flattened out. Attach the file and it gets the columns, the sheets, and the formulas. The same goes for PDFs, slide decks, images, and code. With million-token context windows on the current models, a whole report, a folder of meeting notes, or a year of invoices fits in one conversation, so the old advice to chunk and summarize is mostly gone.
Three habits make attachment work well:
- Attach the relevant files, not everything. Accuracy on document questions falls as the context grows; the best models stay around 1% fabrication at modest lengths but exceed 10% at very long ones in a 2026 study. Pick the sources that matter and name which one is authoritative.
- Give links and let it fetch. For public pages, paste the URL and ask the assistant to read it rather than copying the text yourself. For a screenshot or a whiteboard photo, attach the image.
- Say what the file is for. "Use the attached policy as the source of record; flag anything in my draft that contradicts it" tells the model how to weigh the material.
Projects and standing instructions
A project (the name varies: Projects in Claude and ChatGPT, Gems in Gemini) is a container for related conversations with its own instructions and its own files. Everything you put there is available in every chat inside it, and in most apps the project keeps separate memory from your general chats. Claude's free tier includes up to five projects; paid tiers remove the limit.
Treat project instructions as a standing brief: who you are, who the audience is, the house style, the sources that count, and the checks you always want. Then keep it short. Anthropic's guidance for its own instruction files applies to any project: ask of every line, "would removing this cause mistakes?" and cut it if not. Long lists of rules get partially ignored because the important ones drown in the noise; a page that tries to cover every edge case performs worse than ten clear sentences.
Set up one project for each recurring area of your work, such as a client, a course you teach, or a report you own. The playbook Build Your Brief and Skill Library walks through it step by step.
Memory
The major assistants now remember across conversations. Claude includes memory on its free tier, and Anthropic reported in August 2026 that it merged memory from chat and from agent sessions into one record, so what the assistant learns in one place carries to the other (reported). ChatGPT's memory draws on past chats; in the EU and UK it is off by default. Google's Gemini has its own equivalent.
Four steps get memory working for you rather than against you:
- Turn it on and seed it. Most apps can generate an initial memory from your past chats. Review that summary and correct it.
- Test recall. Open a new chat and ask what the assistant knows about your work. Fix anything wrong at the source rather than in the chat.
- Edit periodically. Memory accumulates stale facts, such as a project you finished or a role you left. Delete them.
- Use incognito for the exceptions. Every app has a temporary or incognito mode that is not remembered. Use it for sensitive topics and for one-off tasks that would pollute your profile.
What should never be in memory, such as other people's personal data and anything you would not want surfaced in an unrelated chat, is covered in Protect.
Connectors
A connector gives the assistant live access to another service: Gmail, Google Drive, Calendar, Slack, Microsoft 365, and hundreds more through reviewed directories. For builders the technology underneath is the Model Context Protocol (MCP); for you it is a settings page where you authorize an account. Connectors are available even on Claude's free tier, and the agent modes in every major app depend on them.
Connectors replace the worst kind of pasting. Instead of copying your week into a chat, you ask "what does Thursday look like, and where can I fit a 45-minute prep block?" Instead of forwarding a thread, you say "summarize the open questions in the Acme thread and draft replies." The quality difference is large because the assistant sees the real, current data.
Two cautions, which Protect develops. A connector grants a set of tools, and some of those tools write, send, or delete. Agents and scheduled tasks can use every tool a connector exposes, and Anthropic's own documentation for scheduled routines warns that an included connector's write tools can be used without asking, so remove connectors a task does not need. And only connect services from a reviewed directory or your own IT team; verify every server before connecting is the vendor's advice, not just ours. The Build track's Connect Your First MCP Server covers setup in detail.
Skills
The fifth layer is the newest. A skill is a packaged piece of how-to knowledge, in practice a folder with a SKILL.md file describing when to use it and how, plus any templates or scripts it needs. Anthropic released Agent Skills as an open standard in December 2025, and by spring 2026 dozens of agent products, including offerings from OpenAI, Microsoft, and Google, had adopted the format (reported). You do not write code to make one; the apps include a skill-creator that interviews you and writes the file.
The rule for choosing between a project instruction and a skill is simple: instructions are for rules that apply every time; skills are for knowledge that is only relevant sometimes. "Always write in British English" is an instruction. "How we produce the monthly board pack, with the template and the checks" is a skill. The assistant loads a skill when the task matches its description, so you are not paying for it in every conversation.
Skills are also how your best briefs stop being one-offs. Once a brief has worked three times, turn it into a skill and the next person on your team gets the same result. The Build track's Skills, AGENTS.md, and CLAUDE.md page explains the standard in depth.
The old context-window advice, reframed
| 2023-2024 advice | 2026 habit |
|---|---|
| Be concise to save tokens | Attach the full source; be concise about your own commentary |
| Summarize the chat before it "forgets" | Start a fresh chat per task; keep decisions in a project, not a thread |
| Break the document into chunks | Attach the whole document; ask targeted questions |
| Keep notes outside the tool | Let memory and projects hold them; review them monthly |
| Paste today's schedule or inbox | Connect the calendar or mail with read-only scope |
One piece of old advice survived in a new form: one task per conversation. Mixing unrelated tasks in one thread, the kitchen-sink session, pollutes the context for all of them. When a thread has drifted, do not fight it; open a new one inside the same project, where the standing context is already waiting.
Key takeaways
- Context beats wording: the assistant needs your actual files, standing facts, and live data more than better phrasing
- Attach relevant files and links instead of pasting; name which source is authoritative
- Projects hold a short standing brief; memory holds facts about you; review and prune both
- Connectors supply live data but also grant write tools, so scope them narrowly and connect only reviewed services
- Skills package how-to knowledge for recurring tasks; instructions are for rules that apply every time
Next steps
Continue to: Check: Verify, Iterate, Restart
- Set up your first live connection with Connect Your First MCP Server
- Build your projects and skills with the playbook Build Your Brief and Skill Library