Chats, Projects, files, and Research

Bare chat is a scratchpad that forgets your house rules the moment you open a new thread. Projects are a labeled drawer: instructions, files, and related chats that should keep the same context. Research is a longer multi-source dig with citations, not a faster paraphrase of the first web hit.

This is Part 4 of Learn Claude from scratch. Part 3 got you through a first half hour on web and mobile with one real output. Here we sort the furniture: when a single chat is enough, when a Project pays for itself, how files behave in each place, and how Research fits on paid plans. You will leave with a starter Project kit you can copy without turning setup into a second job.

For model basics, see What are LLMs, ChatGPT, generative AI, and more. For retrieval and grounded-answer habits in analytics work, the Practical AI series (especially context and RAG-shaped thinking) pairs well with Projects. Browse paths on Learn.

What you will leave with

• A clear split: bare chat versus Project versus one-off file upload
• What project knowledge and project instructions actually do
• Plan reality: Projects exist on Free (with caps); large knowledge and RAG scaling are stronger on paid
• How Research differs from ordinary web search and when to spend the usage
• A starter Project kit (name, instructions, file list, first three chats)
• Mistakes that create silent wrong answers from stale PDFs and bloated knowledge bases

The map: chat, Project, files, Research

Four ideas people mash together. Keep them separate in your head.

Anthropic’s help center describes Projects as self-contained workspaces with their own chat histories and knowledge bases. You can upload documents, set project instructions, and run focused chats inside that box. Free accounts can use Projects with a maximum of five projects. Paid plans raise capacity and make large knowledge bases more practical, including retrieval-augmented behavior when content grows large.

When bare chat is enough

Start in bare chat when the work is disposable or unique:

• Rewrite this email once
• Explain a concept I will not revisit this month
• Brainstorm names I will paste into a ticket and forget
• Debug a one-line formula with no house style
• Try a prompt pattern before you decide it belongs in a Project

Bare chat is also the right place to fail safely. If the thread gets messy, start another. You have not polluted a knowledge base yet.

Rule of thumb: if you cannot name the Project in five words, stay in bare chat. “Stuff” is not a Project. “Q3 board narrative” is.

When a Project earns its keep

Create a Project when at least two of these are true:

• You will return more than twice this month
• The same instructions should apply every time (tone, audience, refuse rules)
• The same documents should be available without re-uploading
• You want related chats grouped so you can find them
• On Team or Enterprise, you may share the workspace with others (sharing is a paid org feature)

Examples that usually deserve a Project:

• Course or certification prep with a fixed reading list
• Product launch messaging with brand voice notes
• Internal analytics glossary and metric specs for draft SQL and explanations
• Client engagement (with careful redaction rules in the instructions)
• Book or paper notes you will query for months

Examples that usually do not:

• One resume tweak
• A single travel itinerary
• A joke thread
• A throwaway translation

Project instructions versus project knowledge

Two different knobs:

• Project instructions: standing orders. How to write, what to refuse, who the audience is, which vocabulary to prefer. Keep them short enough that you will maintain them.
• Project knowledge (files): source material. Specs, PDFs, glossaries, outlines, CSV samples. Claude uses them as background for chats inside the Project.

Instructions without files make a polite chatbot with a personality. Files without instructions make a pile of paper with no house rules. Use both for serious work.

On larger knowledge bases, paid plans can move into retrieval mode so Claude searches project content instead of stuffing everything into one context window. Treat that as a capacity tool, not magic. If the wrong section is retrieved, you still get a wrong draft. Ask for citations back to your file names and section headings when accuracy matters.

How files fit

You can attach files in a chat or add them to a Project’s files area. Behavior differs.

Chat uploads

Use when the document is relevant to this task only. Anthropic documents generous per-file size limits for chat uploads (on the order of hundreds of MB per file, with caps on file count per chat and PDF page counts). Images and multi-modal PDF pages have their own rules: shorter PDFs can include visual analysis; very long PDFs shift toward text extraction. Check the current upload article when something fails; limits move.

Common supported types include PDF, DOCX, CSV, TXT, HTML, RTF, EPUB, JSON, and images such as JPEG, PNG, GIF, WebP. XLSX often requires code execution and file creation enabled in capabilities settings.

Project files

Project file limits are different (smaller per-file size in Anthropic’s docs, with total content needing to fit usable context or retrieval). Project knowledge is for material you expect to reuse. Prefer clean, current sources over dumping an entire drive.

Practical hygiene:

• One canonical metric spec beats five conflicting decks
• Date the filename: brand-voice-2026-06.md
• Remove files you replaced so the model does not blend old and new
• Prefer text-extractable docs over scanned image-only PDFs when you need exact wording
• Do not put credentials, private keys, or full production dumps into knowledge “for convenience”

What files are good at (and not)

Research: multi-source digs with citations

Research is available on paid Claude plans (Pro, Max, Team, Enterprise) on web, desktop, and mobile. It is not the same as a quick web search. Anthropic describes Research as an agentic multi-search mode: Claude runs a sequence of searches that build on each other, then returns a thorough answer with citations you can check. Web search must be enabled for Research to work.

When to prefer each mode:

Research counts against the same usage pool as other chat, but it can burn through that pool faster because it retrieves many sources and writes longer answers. Turn it on from the + menu in the composer when your plan supports it, then ask a scoped question. If Claude does not seem to research, steer explicitly: “Use the research tool to…”

If you connect workplace tools (for example Google Workspace connectors where available), Research can also pull internal context you have authorized. That is powerful and easy to overshare. Prefer least privilege, and never treat a connector as a substitute for access reviews. Part 6 of this series goes deeper on privacy and paste rules; the short version is: if the doc is sensitive, ask whether it should be in Claude at all.

A Research prompt that stays honest

Use Research.

Question: Compare [A] and [B] for [specific decision].
Audience: [role].
Time horizon: sources from the last [N] months preferred.

Deliver:
1) One-page decision brief
2) Table of criteria with what each source supports
3) Open questions and conflicts between sources
4) Full citations with links

Refuse:
- Do not invent statistics
- Mark confidence low when sources disagree
- Separate fact from vendor marketing language

After the report lands, open at least two citations yourself. Research reduces hunting time. It does not replace your judgment on whether a source is junk.

Starter Project kit

Build one Project this week. Keep it small enough that you will maintain it.

1. Name and purpose

Pick a name a future you will understand:

AMS writing desk
Purpose: Draft AMS-style outlines and rewrites using house rules.
Not for: confidential client data, credentials, production dumps.

Swap “AMS” for your team or class. The “not for” line belongs in the instructions, not only in your head.

2. Project instructions (copy and edit)

You help me draft for human review. I will edit before anything ships.

Voice:
- Plain English, concrete examples
- No hype, no empty praise
- No em dashes in drafts I will publish

Defaults:
- Ask for missing audience, goal, or length when absent
- Prefer bullets and short sections
- Flag assumptions and unknowns explicitly

Refuse:
- Invented numbers, quotes, citations, or case studies
- Pretending a draft is final
- Requests that need secrets I did not provide

When project files are present:
- Prefer them over general knowledge
- Name which file you used when it matters
- Say when the files do not cover the question

3. Seed files (start with three)

Add more only when a chat keeps missing the same fact. Growth by pain, not by “maybe someday.”

4. First three chats inside the Project

Chat A: stress-test the instructions

Using only project instructions and files, draft a 200-word intro
for [topic] aimed at [audience]. Then list which rules you applied
and which files you used. If a rule blocked something, say so.

Chat B: file Q&A

From glossary.md, explain [term] in two sentences for a new hire.
If the term is missing, say "not in glossary" and propose a draft
definition for me to approve. Do not invent org-specific policy.

Chat C: rewrite against the good example

Rewrite the draft below to match the voice in good-example.md.
Keep my facts. Show a short diff-style list of changes after.
Draft:
[paste]

If Chat A ignores your refuse rules, tighten the instructions and retest before you add ten more files.

Worked example: metric glossary Project for analytics

You support a weekly retail metrics review. People keep arguing about “active store” and “net revenue.” Bare chat keeps “helpfully” inventing definitions. You build a Project.

Name: Retail metrics desk

Instructions (short form):

You draft explanations and checklists for retail metrics.
Only use definitions from project files.
If a metric is missing, say so and propose a draft for human approval.
Never invent SQL table names.
When writing SQL drafts, label them "draft for review" and list assumptions.
Refuse customer-level raw extracts; sample rows only if I paste them.

Files:

• metric-active-store.md (grain, inclusion rules, known exceptions)
• metric-net-revenue.md (gross minus discounts, tax treatment, refund lag)
• sql-style.md (prefer explicit joins, no SELECT * in shared drafts)

Chat: “Explain active store for a new ops analyst in 150 words, then give three common mistakes.” Claude should stick to your file. If it invents a fourth rule, your file is incomplete or the model drifted. Fix the file, do not only scold the chat.

When you later need current competitor pricing or public benchmark chatter, that is a Research chat (paid), not something you permanently stuff into the metrics Project without review. Keep durable definitions in knowledge. Keep transient web digests in ordinary or Research chats, then promote only the bits you verified into a dated note file.

For SQL that might leave the Project, keep How to check AI-written SQL next to you. Project context reduces invention. It does not certify a query.

Decision guide you can screenshot

• One task, no reuse → bare chat
• One document, one pass → chat upload
• Same rules and docs, many chats → Project
• Need multi-source external synthesis with citations → Research on a paid plan
• Need hard reasoning on text you already have → higher effort / thinking, not Research
• Need a live system of record → not Claude; use the database, warehouse, or ticket tool

Common mistakes

How this fits Free versus paid (without redoing Part 2)

Projects exist for Free users with a small project count cap. That is enough to learn the pattern. Paid plans matter when you live in Projects daily, need larger knowledge bases with retrieval scaling, want Research as a normal tool, or work in Team/Enterprise shared projects. If you are still deciding whether Claude is a weekly habit, build one lean Project on Free and graduate when limits block real work, not when marketing pages make you anxious.

Promoting a good chat into a Project

You will often start in bare chat because you did not know the work would repeat. That is fine. Promotion is a deliberate step, not a default.

Promote when you notice any of these:

• You re-pasted the same style rules three times this week
• You re-uploaded the same PDF twice
• You keep saying “remember, our definition of X is…”
• A teammate asks for the same explanation you already drafted once

How to promote without dragging junk along:

1. Create the Project with a clear name and a one-line purpose.
2. Write instructions from the rules you actually repeated, not from a fantasy style guide.
3. Add only the current source files. Leave exploratory dead ends in the old bare chats.
4. Run one stress-test chat inside the Project before you trust it for something sendable.
5. Archive or ignore the old bare threads so you stop editing the wrong place.

Do not try to migrate every historical message. Memory and chat search (where available on your plan) are different tools. A Project is a curated workspace. Curation means leaving things out.

FAQ

How many Projects should I have?

As few as still match real recurring jobs. Free accounts cap at five, which is a useful forcing function. On paid plans you can create more, and Team plans can share them, but sprawl still hurts. Three well-kept Projects beat twelve half-empty ones.

Can project instructions replace a good prompt?

No. Instructions set defaults. Each chat still needs a concrete task, audience, and output shape. Think of instructions as the employee handbook and the chat prompt as today’s ticket. Handbook without a ticket produces vague helpfulness. Ticket without a handbook produces inconsistent voice.

Is Research always better than web search?

No. Research is better when you need synthesis across many sources and a cited brief. Web search is better for a quick factual check. Extended thinking or higher effort is better when the hard part is reasoning over text you already provided. Anthropic’s own chooser article frames it that way: match the tool to the job, including the fact that Research burns usage faster.

Should every file go into knowledge?

No. Knowledge is for durable reference. Meeting notes from Tuesday that will be irrelevant Friday belong in a chat upload or not in Claude at all. If you put transient noise into project knowledge, you teach the model the wrong background and you will forget to clean it up.

What if Free RAG behavior differs from what I read online?

Plan features and help articles move. The durable habits do not: keep knowledge lean, date files, refuse invention in instructions, and verify anything you ship. If a large upload fails or quality drops, shrink the corpus and split Projects by topic rather than building one megabase.

Practice for this week

1. List five recent Claude (or other AI) tasks. Mark each: bare chat would have been enough / Project would have helped.
2. Create one Project with the starter kit (instructions + three files).
3. Run Chats A through C from above and note where the model ignored a rule.
4. If you are on a paid plan, run one Research question you actually need, then open two citations by hand.
5. Delete or archive any file that is not the current source of truth.

What comes next

Part 5 moves into everyday jobs: writing, summarizing, studying, planning, with prompt patterns that do not depend on a perfect Project. Part 6 covers memory, privacy, and what never to paste. Later parts hit workplace connectors and wrong-tool cases. Product flavors like Claude Code and Cowork get their own series once chat, Projects, and files feel normal.

Quick recap

• Bare chat for one-offs; Projects for recurring rules and docs.
• Instructions steer behavior; knowledge files supply source material; use both.
• Chat uploads are temporary; project files are for reuse and need hygiene.
• Research is a paid multi-source mode with citations; it spends usage faster.
• Start one small Project; grow files only when repeated pain shows the gap.
• Citations and human edit still close the loop.

Sources

Confirm plan availability and limits in Anthropic’s help center when the product UI changes.

• Anthropic Help Center, What are projects?: https://support.claude.com/en/articles/9517075-what-are-projects (workspaces, knowledge, instructions, Free cap of five projects, paid RAG scaling, Team sharing)
• Anthropic Help Center, How can I create and manage projects?: https://support.claude.com/en/articles/9519177-how-can-i-create-and-manage-projects
• Anthropic Help Center, Retrieval augmented generation (RAG) for projects: https://support.claude.com/en/articles/11473015-retrieval-augmented-generation-rag-for-projects
• Anthropic Help Center, Upload files to Claude: https://support.claude.com/en/articles/8241126-upload-files-to-claude
• Anthropic Help Center, Use research on Claude: https://support.claude.com/en/articles/11088861-use-research-on-claude (paid plans, citations, web search required, usage)
• Anthropic Help Center, When should I use web search, extended thinking, and research?: https://support.claude.com/en/articles/11095361-when-should-i-use-web-search-extended-thinking-and-research
• Anthropic Help Center, Get started with Claude: https://support.claude.com/en/articles/8114491-get-started-with-claude
• Analytics Made Simple, What are LLMs, ChatGPT, generative AI, and more: https://analyticsmadesimple.com/analytics/what-are-llms-chatgpt-generative-ai-and-more/
• Analytics Made Simple, Practical AI series: https://analyticsmadesimple.com/series/practical-ai/
• Analytics Made Simple, How to check AI-written SQL: https://analyticsmadesimple.com/tutorials/how-to-check-ai-written-sql/
• Analytics Made Simple, Learn: https://analyticsmadesimple.com/learn/

More definitions in Key Terms →

Merch on Etsy

Open store →

Data Science T-Shirt View on Etsy → Dashboard to Excel Ceramic Coffee Mug View on Etsy → Apache Flink Shirt View on Etsy → Apache Kafka T-Shirt View on Etsy →