PART 3 SYLAS LYU · FM-01 · REV B

CLAUDE.AI

Projects, connectors and skills in the chat app.

SECTION 06

claude.ai features

For ongoing work on one subject, a well-set-up Project does most of what people build agents for: it keeps your files and standing instructions loaded so every new chat starts informed.

Projects

A Project is a workspace with its own chats, a knowledge base of files, and instructions. Facts from Anthropic's help article:

EX-07 Tighten a Project

  1. Open a Project (or create one). In project knowledge, upload the documents the work is about: briefs, reports, reference material, meeting transcripts. Use descriptive file names such as 2026-Q3_Sales_Report.pdf; retrieval and your own citations both work better with them.
  2. Click "Set project instructions" and paste a version of this, adjusted per Project:
TEXT · C-20
Subject: <what this Project is about>.
Sources: answer from the files in project knowledge first. Name the file and page, slide or timestamp for each answer.
If the files do not cover a question, say so before answering from outside knowledge, and label that answer as outside research.
Never invent page numbers, quotes or citations.
Style: plain prose. Notes as short bullet points.
  1. Start a new chat in the Project and ask: "Which file covers <topic>, and on what page?" Expected: Claude names a file and location. If it answers vaguely, check that the file uploaded and that its text is selectable (scanned PDFs without a text layer read poorly).

Memory and past-chat search

These are two separate features, both controlled in Settings. Per Anthropic's help article:

Feature What it does Plans
Search and reference chats Claude searches your earlier chats when you ask ("what did we decide about X?"); shows as a tool call Paid plans
Generate memory from chat history Claude keeps a summary of durable context (role, projects, preferences), updated about every 24 hours All plans
Project memory Each Project gets its own separate memory and summary Same as memory
Incognito chat Ghost icon on a new non-project chat; not saved to history or memory All plans

The help article fetched for this guide places both toggles and "View and edit memory" under Settings > Capabilities; other versions of the same article show a Settings > Memory section. Uncertain which your app shows; look in both.

To change memory directly, tell Claude in any chat what to remember or forget. The article says such edits apply to your next conversation without waiting for the daily update.

Profile preferences

Settings > Profile holds preferences applied to every new chat: tone, formatting, background. Changes apply to new chats only.

Artifacts

Artifacts are pages Claude publishes to your claude.ai gallery, private until you share them. Some open in dedicated editors where you can change text yourself, such as documents, slide decks and spreadsheets. Others are single web pages Claude builds, such as a calculator, dashboard or interactive explainer. Ask for the thing by name ("make a doc", "make a slide deck") and Claude picks the matching kind; documents and decks can be downloaded as Word, PowerPoint or PDF from the artifact.

BEST PRACTICE Best practices for chats and Projects

From Anthropic's prompting guide and help articles:

  • Write as if briefing a capable new colleague with no context. If a colleague would be confused by the prompt, Claude will be too.
  • Give the reason behind a rule ("no ellipses, because this is read aloud"), so Claude can apply it to cases you did not list.
  • Show the format you want with 3 to 5 examples, wrapped in <example> tags, and keep instructions, documents and questions apart with XML tags such as <document>.
  • Put long material first and your question last; Anthropic reports quality gains of up to 30 percent from that order in its tests. For long documents, ask Claude to quote the relevant passages before answering.
  • Put each instruction where it belongs: profile preferences apply to every chat, Project instructions to one Project, and skills to one repeated task.
  • Use incognito for one-off or sensitive questions. Incognito chats stay out of history and memory, though the help article says they are still retained for 30 days by default.
  • Use "Remove from project" to take a stray chat out of a Project's memory.

What claude.ai does not do

SECTION 07

Connectors in claude.ai

A connector is an MCP server that claude.ai connects to for you. It adds tools for one of your apps, and Anthropic's help article states Claude inherits your own permissions there: a file you cannot open, Claude cannot open either.

Google Drive is a typical example: once connected, its tools (search, read, create, update, share, trash) appear in your chats. The exercises below assume it.

The four ways tools reach Claude

Kind Where you set it up Where it runs Usable in
Directory connector Customize > Connectors, then "+" next to Connectors, or claude.ai/connectors The provider's server, reached from Anthropic's cloud claude.ai web, desktop and mobile; also Claude Code when signed in with the same account
Custom connector Customize > Connectors > "+ Add" > "Add custom connector", then a server URL Any MCP server reachable on the public internet Same as above
Local MCP server, desktop app claude_desktop_config.json Your computer Claude desktop app chat only; not claude.ai web
Local MCP server, Claude Code claude mcp add (section 11) Your computer Claude Code

Facts from the help articles worth knowing:

EX-08 Use a connector on purpose

  1. Connect Google Drive at Customize > Connectors if it is not connected yet. Start a new chat, click "+" at the lower left (or type /), hover over Connectors, and check that Google Drive is toggled on for this chat.
  2. Ask: "Find my three most recently edited Google Docs and list their titles and last-edited dates."
  3. Expected: a tool-call card showing a Drive search, then a short list. If Claude says it cannot access Drive, the connector is off for that chat or needs re-authenticating in Customize > Connectors.

EX-09 Lock down write actions

  1. Open Customize > Connectors and select Google Drive.
  2. Under Tool permissions, set the read tools to Always allow and the tools that change things (create, update, share, trash) to Needs approval.
  3. Expected: next time Claude tries to share or trash a file, it stops and asks you first.

The help article describes these per-tool settings (Always allow, Needs approval, Blocked) in the context of Team and Enterprise owner controls. Whether every option appears on an individual plan is unverified; at minimum, Claude's approval prompts give you a per-action "Allow once" or "Allow always" choice.

BEST PRACTICE Best practices for connectors

  • Use "Allow once" until you trust a tool, then "Always allow" only for read tools.
  • Know the labels in the connector directory: Anthropic's verification page says "Verified" is not a security audit, community connectors are screened lightly, and custom connectors have no Anthropic review at all.
  • Research, the claude.ai mode that searches widely before answering, uses your usage limits faster than normal chat; check its citations, and name a connected source in the prompt if it was skipped.

CAUTION Safety rules from Anthropic's custom-connector article

  • Connect only servers built and hosted by organisations you trust.
  • Read the permissions an app requests when you sign in, and refuse scopes that look unnecessary.
  • A malicious server can hide instructions in its tool results (prompt injection). Watch what tools are called and with what inputs.
  • Click "Allow always" only for a tool you would trust to run unsupervised.
  • Research is a claude.ai mode that runs many searches and tool calls on its own before answering. When using it with connectors, turn off tools that can write, since it can call tools without asking each time.

SECTION 08

Skills

A skill is a folder holding a SKILL.md file of instructions for one repeated task. Claude sees only each skill's name and description until a task matches, then loads the full instructions. This section explains how they work and how to build one well.

Anatomy

TEXT · C-21
meeting-notes/
├── SKILL.md        required: frontmatter + instructions
├── references/     optional: long reference files, read only when needed
├── scripts/        optional: code the skill runs
└── assets/         optional: templates, data files

SKILL.md has two parts: YAML frontmatter between --- lines, then Markdown instructions. Rules from the Agent Skills specification at agentskills.io:

The spec's own contrast: "Extracts text and tables from PDF files, fills PDF forms, and merges multiple PDFs. Use when working with PDF documents..." works; "Helps with PDFs" does not.

Where a skill can live

Location Path or place Loads in
Your claude.ai account Customize > Skills (uploaded ZIP) claude.ai chats; Claude Code sessions signed in to the same account (synced into ~/.claude/skills/synced/)
Personal, Claude Code ~/.claude/skills/<name>/SKILL.md (on Windows %USERPROFILE%\.claude\skills\<name>\SKILL.md) Every Claude Code project on this computer
Project, Claude Code .claude/skills/<name>/SKILL.md inside a project folder Sessions in that project
Plugin <plugin>/skills/<name>/SKILL.md Wherever the plugin is enabled, invoked as /plugin-name:skill-name

Frontmatter: what works where

Field claude.ai upload Claude Code Purpose
name Yes Yes Command name; matches folder
description Yes Yes When to use the skill
allowed-tools Yes Yes Tools pre-approved while the skill runs
license, compatibility, metadata Yes Accepted, not acted on Bookkeeping
disable-model-invocation No Yes Only you can trigger it, by typing /name
argument-hint, arguments No Yes Text typed after /name fills $ARGUMENTS
context: fork, agent No Yes Run the skill in a subagent
model, effort, paths, hooks No Yes Per-skill model, effort, file triggers, hooks

Per Claude Code's skills page, a claude.ai upload containing a Claude Code-only field fails with an error naming the unexpected key. Keep skills meant for both places to the first four rows.

Worked example: a meeting-notes skill

This one turns a meeting transcript into notes in the same format every time.

MARKDOWN · C-22
---
name: meeting-notes
description: Turns a meeting transcript into concise notes with timestamps, decisions and action items. Use when the user uploads a meeting transcript and asks for notes, a summary or action items.
---

# Meeting notes from a transcript

## Inputs
- A meeting transcript (.txt, .docx or pasted). If none is attached, ask for it. Never write notes from general knowledge of the topic.
- Optional: the meeting's name and date.

## Steps
1. Read the whole transcript before writing.
2. Split it into the topics discussed, in the order discussed.
3. For each topic: 2 to 6 short bullets, with the transcript timestamp or line number.
4. List every decision made, one line each, in the speakers' wording where exact wording matters.
5. List every action item: who, what, and by when if a date was said.
6. List open questions: things raised but not settled.

## Output
Sections in this order: Topics, Decisions, Action items, Open questions.

## Rules
- Short bullets, not full sentences.
- Add nothing that was not said. Context you add is tagged [Added].

EX-10 Install it in claude.ai

No terminal needed:

  1. In File Explorer, open C:\Users\<you>, create a folder named skills, and inside it a folder named meeting-notes. The folder name must match the name line in the frontmatter.
  2. Open Notepad, paste the example above, and choose File > Save as. Set "Save as type" to All files, go to the meeting-notes folder, and save as SKILL.md. If File Explorer then shows it as a text document, it was saved as SKILL.md.txt; turn on file name extensions in File Explorer's View menu to check, and rename it.
  3. Right-click the meeting-notes folder itself. On Windows 11 24H2 and later, including 25H2, choose Compress to > ZIP File; earlier Windows 11 builds call it Compress to ZIP file, and Windows 10 uses Send to > Compressed (zipped) folder. Zipping the folder, rather than the file inside it, keeps the folder at the top of the ZIP, which claude.ai needs: its how-to page says it looks for <skill-name>/SKILL.md inside the archive.

Terminal route, once you have done section 4:

POWERSHELL · C-23
mkdir $HOME\skills\meeting-notes
code "$HOME\skills\meeting-notes\SKILL.md"
cd $HOME\skills
tar.exe -a -c -f meeting-notes.zip meeting-notes

tar.exe ships with Windows 10 and 11, and -a picks the ZIP format from the file name. Check the result with tar.exe -tf meeting-notes.zip: every entry should start with meeting-notes/. Avoid Compress-Archive in Windows PowerShell 5.1 for this. Tested: its built-in Archive module (version 1.0.1.0) stored the file as meeting-notes\SKILL.md with a backslash, while the ZIP specification (PKWARE APPNOTE, section 4.4.17) requires forward slashes; tar.exe stored meeting-notes/SKILL.md. Whether claude.ai's uploader rejects the backslash version was not tested.

Then, per Anthropic's help article:

  1. Check Settings > Capabilities: "Code execution and file creation" must be on.
  2. Go to Customize > Skills, click "+", then "+ Create skill", then "Upload a skill", and choose the ZIP.
  3. Make sure the new skill's toggle is on.
  4. Test in a new chat: upload any meeting transcript and write "make notes from this meeting". Expected: the reply shows the skill being used and follows the four-section format.

EX-11 Install it in Claude Code

After section 9:

POWERSHELL · C-24
mkdir $env:USERPROFILE\.claude\skills\meeting-notes
Copy-Item ~\skills\meeting-notes\SKILL.md $env:USERPROFILE\.claude\skills\meeting-notes\

Start claude in any folder and type /skills. Expected: meeting-notes in the list. Invoke it with /meeting-notes or by asking for meeting notes. Skills uploaded to your claude.ai account also sync into Claude Code when you sign in with that account (Claude Code v2.1.273 or later), so this step is only needed for skills you keep local.

Writing skills that trigger correctly

BEST PRACTICE Best practices for skills

From Anthropic's skill-authoring guidance:

  • Write the description in the third person ("Turns a meeting transcript into...") and say both what it does and when to use it. The claude.ai how-to page allows 1,024 characters, an older help article says 200; staying under 200 works with both.
  • Keep SKILL.md under 500 lines, move detail into references/ files one level deep, and give any reference file over 100 lines a table of contents.
  • Match strictness to fragility: plain guidance for open-ended work, exact steps or a script where one mistake breaks the result.
  • Write at least three test prompts before polishing the instructions, then run each in a fresh chat with the skill on and off. In Claude Code the skill-creator plugin (Exercise 22) runs the with-and-without comparison for you; on claude.ai, skill-creator is a skill under Customize > Skills and works through the prompts one at a time.
  • Use forward slashes in any paths inside a skill, even on Windows.
  • Read the allowed-tools line of any skill that comes with a repository before running Claude Code there; a skill can grant itself broad tool access.

Claude Code extras worth knowing