Pick the right delivery
Choose, for any given document, whether to attach, paste, summarize, or link — and explain why in one sentence.
Intermediate training
Working with Context
Feed Claude the material it needs — attachments, long documents, projects, prior context — without overwhelming it or losing important detail. Where Module 1.2 covers the mechanics, this module covers the craft.
Start here
What you'll be able to do after this lesson
Structure a long brief
Build a multi-part prompt so the important pieces sit where Claude actually reads them — and so a long source doesn't bury the question.
Know when to start clean
Recognize when a chat is past its useful life, decide what context to carry forward, and rebuild in a fresh conversation in under a minute.
Three context decisions
What goes in, how it gets in, where it sits
Relevance — what to include
Every piece of context spends from the budget. Include what affects the answer; cut what doesn't. The skill is being honest about what's actually relevant versus what feels comforting to include.
Delivery — how to include it
Attach the source when the source is what you want analyzed. Paste only when the snippet is short and self-contained. Summarize when you genuinely don't need the original. Link when the conversation lives on a surface that can fetch.
Order — where it sits in the brief
Most important first. Goal at the top, framing right after, sources labeled, structured ask in the middle, constraints at the end. The pieces aren't optional, but the order is where most prompts fail.
Trainer note: Claude reads context like a person skimming on a deadline. Put the
headline up top. The further down the page something sits, the less it shapes the first draft — and the
first draft is usually what you ship.
Three surfaces context lives on
Where context belongs, and what belongs where
The prompt itself
Goals, framing, instructions, your specific question, format requirements. Things that change per request — and things Claude must read before reading anything else.
Per requestAttached files
The sources Claude should reason over: PDFs, documents, exports, screenshots. Material that's too long to paste, or too structured to describe.
The sourcesProject context & memory
Standing instructions, voice samples, persistent files in a Claude Cowork project. Things that hold across many conversations — your voice, your team's templates, your domain.
Across sessionsContext that works
Five rules for feeding Claude
Attach the source, don't paraphrase it
If the source is what you want Claude to reason from, attach it. Paraphrasing a 14-page disclosure into two sentences gives Claude two sentences to work from.
Summary on top, detail below
The first paragraph of a long brief is the most-read part. Put the goal, the framing, and the ask there. Detail and source material come after.
Label the parts
"SOURCE-A is the policy doc. SOURCE-B is the borrower file. Refer to them as SOURCE-A and SOURCE-B in your output." Naming lets Claude reference precisely.
Use negative examples sparingly
"Do not summarize the whole guide; only the lines that touch X" is useful. A paragraph of "don't do this, don't do that" is noise. One or two negative lines is the limit.
Trim ruthlessly before sending
Every "just in case" paragraph crowds out the answer. If a line wouldn't affect the output, cut it. Lean briefs land harder.
Prompt upgrade
Weak prompt
Read this thread and tell me what to do.
Context-aware prompt
Work-ready prompt
Goal: pull the next three actions out of this borrower thread. For each, give the owner, the deadline, and the one question that's blocking. Format as a numbered list — no thread summary. Thread is the attachment labeled THREAD. If something is genuinely ambiguous, flag it instead of guessing.
Context-heavy patterns
Four tasks where context discipline pays off
Document Q&A
Long source attached, specific question up top, ask for citations and page numbers. The cleanest, highest-leverage context pattern.
Two-source comparison
SOURCE-A and SOURCE-B both attached, both labeled. Ask Claude to compare on specific dimensions you list — not "compare these documents" open-ended.
Structured extraction from messy input
Paste or attach the messy text. Define the exact schema you want. Claude turns the unstructured into the structured better than almost any other pattern.
Draft-vs-standard review
Your draft as one attachment, the reference standard as another. Ask Claude to check the draft against the standard on the dimensions you care about.
When context is going wrong
Five drift signals to watch for
Employee rule: Drift signals are not edge cases — they are routine. When you see them,
the fix is almost never "tell Claude harder." The fix is to restart in a clean chat with a leaner brief.
Push context to its limits
Six exercises to build context discipline
The goal is reps. Most employees only know what their context habits are doing once they've watched the same task succeed or fail with three different framings.
- Open the brief anatomy in this lesson. Click each pin and read what that structural piece is doing — then write your own version for a task you do this week.
- Take a 30-page document and run three different framing prompts on it. Note which framing produced the most useful first draft.
- Deliberately attach a too-long source — a 100+ page file. Note exactly where Claude starts to drift. You'll feel the budget run out.
- Structure a long brief with headers and labeled sources (SOURCE-A, SOURCE-B). Ask Claude to use the labels in its output and verify it did.
- Carry context cleanly into a fresh chat. Decide ahead of time what to copy over. Start clean. Note what you needed and what you didn't.
- Write a one-page guide to context for your team — the three rules you'd put on a wall. Trade with a teammate and merge.
Completion standard
You've finished this module when you can pick the right delivery method in five seconds, structure a long brief that puts the goal on top, and restart a drifting chat without losing the thread of the work.