DC CAP Enterprise AI  ·  Week 2

A long chat keeps everything,
but attention is lost.

Every turn of a conversation lands in the window, which is why a long chat gets smarter about your work. Past a point it also gets forgetful. This week you manage the window across time, on your own real work.

Week 2 of 8 About 30 min · at your pace Nothing to turn in
Week 2 · the setup

Here's how Week 2 goes

Week 1 packed the window for one message; this week you keep it working across a whole conversation.

1Watch the window fill: see a long chat drift in real time
2Carry context forward: end one chat, start the next one smarter
3Meet memory: the context that persists, and who curates it
4Run one real task: a safe file, a full brief, one iteration
Move 1 · the idea

The window fills

A conversation is the window growing turn by turn. Watch a Monday-morning chat below: six turns, three tasks, and one rule that quietly stops holding.

The context window · compressed for this demo; a real one holds hundreds of pages
You · turn 1For everything in this chat: keep drafts under 150 words, plain language, no scholar names.
ClaudeGot it. Short, plain, no names.
You · turn 2[Attached: fall retention summary, aggregates only] Draft a fall kickoff note for our team.
ClaudeDraft, 142 words: "Team, our fall numbers are in and the trend is good…"
You · turn 3Tighten the middle paragraph and make the ask clearer.
ClaudeDone. 128 words, ask moved up.
You · turn 4New topic: give me five agenda ideas for Friday's team meeting.
Claude1. Show one real task you ran this week. 2. Trade favorite prompts. 3. …
You · turn 5Turn idea three into a one-paragraph description.
Claude"Bring a draft that missed, and we'll find the missing context together…" Under 150 words, as you asked.

Five turns on, the rule from turn 1 is buried under everything that followed. Still steering, less and less. And the cap you set for drafts just clamped an agenda blurb, too: old constraints drag into new work.

You · turn 6Back to the kickoff note. Add a closing section on spring plans.
ClaudeFull revised note, 380 words: "Team, as we reflect on the fall semester and look toward spring…"

Longer, stiffer, and more than twice the cap you set in turn 1.

You'll feel this in your own chats before you measure it. Three signals it's time for a fresh chat:

Claude repeats itselfThe same point or question comes back, reworded.
It forgets an early instructionA rule you set at the top stops holding.
It drags old constraints inGuidance for the last task shows up in the new one.

Give each piece of work its own chat.

Spotting those signals is Discernment at work: you read each answer against what the window actually holds, and you call the restart.

Why does it fade?

The window didn't shrink, and the text is all still there. So what changed?

Show what's going on →

Claude weighs everything in the window, and recent turns pull hardest, so a turn-1 rule steers less and less even though it never left. Restate what matters or start fresh; the optional section below covers the research.

Move 2 · the habit

Carry it forward

Starting fresh costs nothing if you carry the decisions with you. The whole move takes three steps and about thirty seconds.

  1. Ask the old chat to wrap up. Use the prompt below; Claude writes the summary from everything it can still see.
  2. Start a fresh chat with the New chat button. The new window is empty, so every word you load gets full attention.
  3. Paste the summary in. Add the file if the work still needs it, and pick up where you left off.
Summarize the decisions we made and what's still open, so I can start a fresh chat.

That wrap-up is Description pointed at your future self: you write the opening context for the next window.

Move 3 · the feature

Memory in the app

Some context can outlive the chat. Where memory is available in our workspace, Claude carries a short set of stable facts between conversations, and you decide what's on the list.

What it is

Stable facts, carried between chats

Your role, how you like drafts, the work you keep coming back to. Those facts ride into new conversations without being retyped.

Who curates it

You do, in Settings

Week 1 said you don't pick what memory surfaces mid-chat. The list itself is yours: click your initials in Claude, choose Settings, and open Memory to see each saved fact and edit or remove any line. Read yours once this week, the same way you'd review your instructions for Claude.

Nothing about an individual scholar or family ever goes in memory. Memory is built from what you type in chats, so the gate on your inputs is the gate on memory; a saved memory enters every future window. Review the list in Settings to catch anything that shouldn't stay.

Memory is context you curate, and the gate applies to it.

Move 4 · the doing

Real work, run properly

This week the course points at your own plate. Pick one task you run every week or month, and run it through Claude properly one time.

A file, checked firstOpen it before you attach it. Aggregates and internal material enter; anything pointing at one scholar gets described instead.
A five-part briefRole, context, task, example, format, straight from Week 1.
One iterationRead the draft, name the gap, send one specific fix.
Watch, then try
You[Attached: AI program one-pager, no scholar data] You're writing to the KPMG program team that funds DC CAP's AI training. Using the attached one-pager, draft this month's update: open with where the program stands, and close with the capstone window closing September 25. Warm and professional, under 150 words.
ClaudeSubject: Program update: on schedule

Hello [contact name],

Every DC CAP team is running the training on its own real tasks this week. Earlier weeks stay open for anyone catching up, and the capstone window closes September 25. Thank you for making this work possible; happy to walk through the site any time.

[Your name]
YouThe date reads right against the file. This reader skims: cut it to three sentences and drop the walkthrough offer.
ClaudeHello [contact name]: every DC CAP team is running the training on its own real tasks this week. Earlier weeks stay open for catch-up, and the capstone window closes September 25. Thank you for making this work possible.

The second ask did the real work. Checking the draft against the file, and against what you know about the reader, is Discernment; the fix took one sentence.

Your turn:
  1. Pick the task: a recurring update, summary, or draft that touches a document.
  2. Open the document and check it first: nothing pointing to one scholar or family.
  3. Attach the file with the + beside the message box, and warm up with one question the document can answer. Check the reply against the page.
  4. Send the five-part brief. Read the draft against the file, name one gap, ask for one fix.
  5. End the chat with the wrap-up prompt either way: fresh chat, paste the summary in, ask for one refinement.
One task run properly is the whole assignment this week.
Your win

What you can do now

You can watch a window fill, restart without losing work, and curate the context that persists. Check what landed:

This is your own check that the week landed. It stays in your browser — nobody else sees it.

Support, whenever you need it

Reach out to Preston anytime; office hours run 9–10 AM ET every weekday. Bring the task you ran, or the moment a chat went stale on you. We are stronger when learning together.

What's nextWeek 3 · Projects: context that stays
Start Week 3 →
Optional

Under the hood

Skip it and lose nothing.  Three questions people usually ask about long chats and memory.

Why do things fade in the middle?

Claude weighs everything in the window when it predicts the next word, and the weighting concentrates at the very start of the window and on the most recent turns. In a long chat, the stretch between gets the least attention, so an instruction from twenty turns back still sits in the window while steering less and less. Researchers call the pattern "lost in the middle," and it shows up across AI models, without anyone designing it in. Restating what matters, or starting fresh with a summary, moves the important material back to where attention is strongest.

What does memory store, and who controls it?

Memory keeps short, stable facts, like your role or a standing formatting rule, rather than transcripts of your conversations. Where memory is available in our workspace, Settings shows the full list, and you can edit or delete any entry. Two habits keep it useful: review the list occasionally so it stays current, and run the gate before anything gets saved, because each memory enters every future window. A stale or unsafe memory misleads every chat at once, which is exactly why you're the curator.

Why does a fresh chat often beat a marathon?

A marathon chat carries every constraint, tangent, and dead end you've passed through, and Claude weighs all of it: edits meant for the kickoff note bleed into the Friday agenda. A fresh chat opened with a carry-forward summary holds only what the task needs, so every word gets full attention.