Day 1e · 30 minutes
Context
Engineering
"What are best practices for prompting to get more strategic outputs?" — this session is the full answer.
30 min
Katherine
Cowork + CLI
The most common AI complaint
I've tried AI and it doesn't really work for me.
Almost always a context problem.
Not a capability problem. Not a model problem.
The AI didn't know enough to give you what you needed.
The pattern — every time
Generic input → generic output.
Strategic input → strategic output.
The thing nobody explains
Your AI has working memory.
It fills up.
One question
~5% of working memory
A long conversation
~60% of working memory
Every message, every file, every answer adds to the total. Once it's full, something gets pushed out.
"Every new token introduced depletes this budget by some amount"
— Anthropic Engineering, 2025
Claude's working memory holds roughly 150,000 words — about 500 pages. Sounds huge.
A 3-hour working session fills it.
Why conversations fill up so fast
Every turn sends
everything again
"The Messages API is stateless, which means you always send the full conversational history to the API."
— Anthropic Docs
A 10-message conversation: 10+8+6+4+2 = 30 messages' worth of context sent.
Fresh starts and focused chats aren't a workaround — they're the response to this constraint.
Here's what fills that working memory
Context = everything the AI
knows when it answers
Background
Who, what, why — the situation. Client, project, what this work is part of, why it matters now.
Specifications
Output requirements: format, length, audience, channel, budget. The concrete shape of what you need.
Voice & Examples
Tone, style, and show-don't-tell samples. "Like Patagonia's copy" beats "warm and friendly" every time.
Intention
What decision does this output enable? Who reads it? What changes? This is what makes output strategic, not just descriptive.
Limits
What NOT to do. Hard constraints, non-negotiables, things to avoid. Fences matter as much as directions.
A vocabulary check — what you already call these things
The prompt is the smallest part.
Context — everything the AI sees when it answers
Background
brand overview · client brief · "the situation" · account history · about the brand
Specifications
deliverables · scope · format brief · copy matrix · requirements · "what we're making"
Voice & Examples
references · comps · mood board · "write like this" · brand voice guide · samples
Intention
campaign objective · "the job this piece does" · brief behind the brief · success criteria
Limits
brand guardrails · never-do list · "off-strategy" · legal constraints · client sensitivities
The Brief =
A pre-packaged context bundle — background, specs, voice, and intention in one document. When you paste a brief into an AI tool, you're loading most of your context at once. The question is whether the brief was complete to begin with.
+ The Prompt
activates the context
The specific ask you send in the moment. Can be one sentence ("write the subject line") or your entire brief pasted in. Also called: the ask · instruction · message · "what I typed"
Writing a better prompt matters less than building better context. The prompt asks; the context determines what's possible.
Not prompt engineering — context engineering
Context engineering is the delicate art and science of filling the context window with just the right information for the next step. Too little or of the wrong form and the LLM doesn't have the right context for optimal performance. Too much or too irrelevant and costs go up and performance comes down. Doing this well is highly non-trivial.
Andrej Karpathy · Former Director of AI, Tesla · Former OpenAI · June 2025
"Context engineering is iterative — the curation phase happens each time we decide what to pass to the model."
Anthropic Engineering Blog · Jan 2026
Old frame
Prompt Engineering
"Write a better sentence"
→
New frame
Context Engineering
"Design the full information environment"
Same AI. Same task. Totally different output.
Prompts that generate text
vs. prompts that generate decisions
Without context
"Write a campaign brief for a product launch."
Generic brief. Any product, any brand, any audience. Placeholder language throughout. You rewrite 90% of it.
With context
"Write a Q2 launch brief for Bondi Botanics's eco skincare range. Tone: like Patagonia's copy — environmentally committed but never guilt-tripping, not like a sustainability manifesto. Differentiator: 100% local Australian sourcing vs. Frank Body. Budget: $50k. Channels: Instagram, OOH, micro-influencer. The brief will be presented to the founding team Thursday — they care most about whether the message differentiates from Frank Body. Output: three-section brief (objective, audience insight, creative territory). One page max."
Specific, actionable, decision-enabling. You edit 10%, not 90%.
The difference isn't more words. It's the right information — especially Intention: what decision does this output enable?
What happens even when you're doing it right
Context drifts.
Even good sessions break.
Stanford/Meta research (ACL 2024): Performance drops more than 30% when relevant information is buried in the middle of a long context. Your AI prioritises the start and end — not the middle. A U-shaped attention curve, not a flat line.
Anthropic finding: Placing your question after your documents — not before — improves responses by up to 30%. Order matters. Put your context first, your question last.
What happened — the Carrigan case
Researcher Mark Carrigan was using Claude to compile blog posts. He gave good instructions. The session ran long. The working memory filled and triggered a memory checkpoint (compaction). Claude resumed — but now it was writing new text from summaries instead of compiling the originals. He was doing context engineering correctly. The system still drifted.
Why it happens
Working memory is finite. As sessions run long, earlier content gets compressed into summaries. The original task instructions — and their specificity — can get lost. Anthropic describes it as "a gradient rather than a hard cliff." In Cowork, it happens silently. In CLI: /compact forces it, /clear resets.
Three things that actually help:
1. Keep critical instructions in a document you re-paste at the start of each session — not just in conversation history.
2. Build checkpoints: every 5–10 outputs, verify the AI is still doing the original task, not its summary of it.
3. Segment long tasks — don't run a 3-hour session. Run three 45-minute sessions with deliberate context resets between them.
The solution to context overload
Split the work.
Run parallel teams. (sub-agents)
In Cowork — what you can do
Start separate conversations for separate tasks
One for research, one for drafting, one for review. Each gets focused working memory. No session gets overloaded. This is the manual version of parallel teams (sub-agents).
Chat 1: Background + brief-writing task
Chat 2: Competitor research task only
Chat 3: Survey docs + summary task
In CLI — the explicit pattern
Coordinator assigns tasks to parallel teams
Use a coordinator (orchestrator) that assigns tasks to parallel teams (sub-agents) — each with a fresh, focused working memory. The coordinator combines their outputs. You see and control what each team knows.
Coordinator → Team("research X") → Team("write brief") → combine
This is the workaround professionals use. Each task gets only what it needs.
The practical session decision
Default to fresh start.
Carry your documents forward.
90% of the time — fresh start, updated docs.
- New conversation. Paste your context document and updated progress document at the top. Clean working memory, full context.
- You're not starting over — you're starting clean. The documents carry everything forward.
- "Go back and forth until you like the plan — then Claude can usually 1-shot it. A good plan is really important." — Boris Cherny, head of Claude Code ↗
10% of the time — carry forward to investigate.
- Stay in the chat when you want to understand what went wrong — did it miss a constraint, misread the brief, or just drift?
- Poke around. Ask it what it understood. Identify the failure precisely.
- Then fresh start with a better-specified brief. Carry forward is diagnostic — you still end up in a new chat.
The context document: For any recurring project, maintain a living document — client background, brand voice, key decisions, what the AI should and shouldn't do. Paste it at the top of every new session.
The progress document: A file the AI updates throughout a long session — decisions made, tasks completed, current state, what comes next. At the end of each major step: "Update our progress document with what we just decided and what comes next." When you restart fresh, paste both documents. You don't lose your place.
Context doesn't have to be typed every time
Getting context in:
a progression
Download & Upload
Export from any tool — CSV, PDF, screenshot — and attach or paste it in. Low-tech, always works, zero setup. This is your default move.
This month
Cowork power users
Direct Connections (MCP)
Think of it as a data cable between Claude and your tools. Gmail, Drive, Slack, Outreach, DocuSign are available in Cowork now. Worth setting up if you're repeatedly loading the same context by hand.
Live Data Feeds (APIs) & Browser Automation
Powerful for live data from CRMs, analytics, web pages. Requires technical setup. Not for today — but worth knowing where this goes.
The habits that make it stick
Do this. Not that.
✓ Do
Paste your context document first
Context doc
"Client: ACME. Voice: warm, not corporate…"
One conversation per task
Working memory
Research / drafting / editing in separate chats
Put documents before your question
Context order
Paste the brief, then ask your question
Start fresh when it feels off
Drift
2 minutes to reset beats 30 minutes of drift
Check the output against your brief
Verification
Does it reflect your background, intention, and limits?
✗ Don't
Rely on AI remembering
Stateless API
"As I mentioned earlier…" may fail after compaction
Run everything in one chat
Working memory
Overloaded working memory = drift and confusion
Lead with your question
Context order
AI answers before reading your 15 pages of context
Accept output that ignores your brief
Verification
Name the specific failure. Don't let drift accumulate.
Start new sessions from scratch
Context doc
That background belongs in a document, not your memory
Working memory is finite. What you put in it is a choice. Make it deliberately.