Study guides / CCAO-F / Domain 3

Product and Model Selection · Lesson 4 of 4

3.4 — Context Limits: When to Restart, Summarise, or Persist

Recognise when a long conversation's context window is degrading Claude's grip on earlier detail, and choose between starting fresh, asking for a summary, or persisting information at the Project level.

Every conversation with Claude has a finite context window — the amount of prior conversation, pasted material, and instructions it can hold onto at once. As a single conversation grows very long — many dozens of turns, several large documents pasted in over time — Claude's grip on the earliest content can start to loosen, even though that content is technically still part of the thread. This shows up in concrete, recognizable ways: Claude forgetting or misapplying a formatting rule you set at message 3, contradicting a decision made earlier in the same conversation, re-asking something you already answered, or handling a document pasted forty messages ago less precisely than it did right after you pasted it.

Recognizing this degradation, and knowing which of three responses fits, is the practical skill the exam tests here — not the technical mechanics of context windows themselves.

Recognizing the Signal

The signal isn't subtle once you know what to look for: Claude re-explaining something already settled, applying an early instruction inconsistently as the thread grows, or getting details wrong about material that was pasted in much earlier in the same conversation. When you see this, the fix is not to repeat the instruction more forcefully or paste the same material in again in the same thread — that treats a structural problem (too much history competing for attention) as if it were a wording problem. The actual fix is choosing the right one of three moves: starting a fresh conversation, asking for a summary first, or persisting the information somewhere more durable than one long thread.

Three Responses: Restart, Summarise, or Persist

Start a brand-new conversation when most of the current thread's content genuinely isn't relevant to what comes next — for example, you used one long conversation to explore several unrelated ideas and only want to continue developing one of them. A clean slate avoids dragging irrelevant baggage forward and gives Claude a clear, uncluttered context for the next request.

Ask Claude to summarize the conversation so far, then continue in a new conversation seeded with that summary when the thread does contain decisions, facts, or context you genuinely still need — but the thread itself has grown too long and unwieldy to keep extending. A good summary compresses the parts that actually matter into a form that comfortably fits a fresh context window, instead of carrying an entire sprawling transcript forward turn after turn. This is usually the right move when you notice the degradation signals above but the conversation still holds real, needed history.

Persist information at the Project level when it isn't really specific to one conversation at all — it's a standing reference that should be available to every future conversation on this topic: a style guide, a client's background, a recurring set of product facts. Uploading it once to a Project's knowledge base (or setting it in the Project's custom instructions) is the durable fix. The test is simple: if you'd otherwise re-summarize or re-paste the same material into a fresh chat every single time you need it again, it belongs in Project knowledge instead, not in the memory of one long-running thread.

The Same Limit Affects Long Documents, Too

Context limitations aren't only about how long a conversation runs — they also affect how much a single very long document Claude can work with precisely in one pass. A 300-page report pasted in whole, when only two chapters are actually relevant to the current task, competes for the same limited attention as everything else in the conversation. Chunking the document — pasting only the relevant section, or asking for a section-by-section summary pass before working with the whole thing — keeps Claude working with a manageable, focused amount of material instead of relying on it to reliably track every detail of a massive document loaded all at once alongside a long, active conversation.

Key Concept

Three responses to a degrading long conversation, chosen by what's actually needed going forward: restart fresh when earlier content is irrelevant baggage; summarize-then-continue when real history is still needed but the thread has grown unwieldy; persist to Project knowledge when the information is a standing reference needed across many future conversations, not just this one. The same context limit is why very long documents sometimes need chunking or pre-summarizing before being pasted in.

Common Exam Distractor

Repeating or re-stating an instruction more forcefully in an already-degraded, very long thread is a distractor — it treats a structural context problem as a wording problem and rarely fixes it. So is reflexively starting a brand-new blank conversation when the earlier content actually contained decisions still needed going forward — that loses useful context instead of preserving it; summarizing first is the better move in that case.

Exam traps

Practice question

After roughly 80 messages in one long conversation, Claude begins contradicting a formatting rule the team set at message 3 and misremembers a decision made partway through. The team still needs everything decided so far as they continue working. What's the best next step?

  • A Keep working in the same conversation and simply repeat the formatting rule and the earlier decision again in the next message

    This treats a structural context problem as a wording problem. Repeating instructions in an already-degraded, very long thread tends not to fix the underlying issue and the same drift is likely to recur.

  • B Start a brand-new, empty conversation and continue from there

    A blind restart discards the decisions and context the team explicitly still needs going forward — appropriate when earlier content is irrelevant, but not when it's still needed.

  • C Ask Claude to summarize the conversation so far, then continue the work in a new conversation seeded with that summary Correct

    This preserves the decisions and context still needed while shedding the excess length causing the degradation — the right move when real history matters but the thread itself has grown unwieldy.

  • D Upload the full 80-message conversation transcript as a Project knowledge base file and keep working in the original thread

    Project knowledge is the right fix for a standing reference reused across many future conversations, not for continuing one specific still-active task — and it does nothing to fix the degraded thread the team is currently working in.

Build exercise: Practice Restarting, Summarising, and Persisting

Beginner · 20 minutes

You'll practice:

  1. In claude.ai, hold a genuinely long working conversation (aim for at least 20-30 back-and-forth messages) on an evolving task — for example, iteratively drafting and revising a short document, with a specific formatting rule stated only once near the very beginning. Periodically re-check whether that early rule is still being followed consistently.

    This lets you directly observe, rather than just read about, the kind of degradation signal the exam describes — an early instruction losing its grip as a thread grows.

    You should see: At some point, likely a formatting or stylistic drift where the early rule is applied less consistently than it was in the first few replies.

    Hints
    1. State the early rule once, clearly, and don't repeat it — the point is to see whether it holds up unaided.
    2. Pick a rule that's easy to visually check for (like a specific heading style or forbidden word).
    3. If it's still holding at 30 messages, keep going a bit further — the effect is more noticeable the longer the thread runs.
  2. Once you notice drift (or after a genuinely long thread even without obvious drift), ask Claude directly to summarize the conversation so far, capturing the key decisions and the early rule. Start a brand-new conversation, paste in that summary, and continue the same task from there.

    This applies the summarize-then-continue pattern and lets you confirm it restores consistent behavior without losing anything the team still needed.

    You should see: The new conversation, working from the compact summary, correctly follows the early rule again and reflects the decisions made earlier in the long thread.

    Hints
    1. Explicitly ask for the summary to include any stated rules or constraints, not just the general topic.
    2. Compare a reply from the new, summary-seeded conversation against a reply from just before you switched — check the rule is followed consistently again.
    3. Notice how much shorter the summary is than the full 20-30 message thread it's replacing.
  3. Take the formatting rule from your summary and, instead of leaving it only in that one conversation's history, add it to a Claude Project's custom instructions or knowledge base. Start a completely new, unrelated conversation inside that Project and confirm the rule is applied automatically without being restated.

    This distinguishes a one-off summary (good for carrying forward one specific thread's history) from Project-level persistence (good for a standing rule needed across many future, even unrelated, conversations).

    You should see: A new conversation in the Project that follows the formatting rule immediately, with no summary or restatement needed, because the rule now lives at the Project level rather than in one thread's memory.

    Hints
    1. If the rule doesn't apply automatically, confirm it was actually saved in the Project's instructions or knowledge base, not just mentioned in a chat.
    2. Test with a task genuinely unrelated to your original long conversation, to confirm the rule persists independent of any one thread's history.
    3. Write one sentence on why this rule was a better candidate for Project persistence than for a one-off summary.

Sources