Context Management Decision Framework

Run the Context Management Decision Framework Diagram Fullscreen

About This Diagram

Chapter 7 introduces half a dozen context management techniques, and it is easy to come away with a list rather than a method. This diagram turns the list into a decision tree: given a task, follow the branches and you arrive at the technique that fits.

The tree has two halves. The top half decides what goes into the prompt. The bottom half — the part most people forget — decides what to do when a conversation runs long enough that the history itself becomes the problem.

How to Use

Start at "New Prompt Task" and follow the arrows down.
Hover over any node for a detailed explanation in the right-hand panel, including the symptoms that tell you which branch you are on.
Notice the convergence at "Submit Prompt." Four different technique paths meet there, which is a useful reminder that these techniques are alternatives, not a checklist.
Follow the multi-turn loop. A single conversation can pass through the history-management cycle several times.

Embed This Diagram

Place the following line in your website to include this diagram:

<iframe src="https://dmccreary.github.io/prompt-class/sims/context-management-decision-framework/main.html"
        height="882px"
        width="100%"
        scrolling="no"></iframe>

Workflow Steps

Step	What happens
New Prompt Task	Clarify what you are asking and what the model needs to know
How much context needed?	Minimal, moderate, or extensive — the first fork
Prompt Compression	Say more with fewer tokens
Relevance Filtering + Context Injection	Remove what does not help, add what does
Does it fit in context window?	Count tokens including instructions and response
Background Info + Domain Priming	It fits, so supply background and domain vocabulary
Document Summarization + Chunking	It does not fit, so reduce it first
Process in Stages	Chunked material needs multiple passes
Submit Prompt	All technique paths converge here
Multi-turn conversation?	One-shot tasks end; conversations continue
Manage Conversation History	The model is stateless — history management is your job
History getting long?	Watch for forgetting, slowdown, and rising cost
Summarize + Reset Context	Summarize, start fresh, paste the summary in

Lesson Plan

Learning Objective

Students will be able to select and justify an appropriate context management strategy for a given prompting task, including deciding when to reset a long conversation.

Bloom's Level: Apply (L3) — select, apply

Grade Level

High school through adult learners.

Duration

15 minutes

Prerequisites

Students should have read Chapter 7's coverage of prompt compression, relevance filtering, context injection, and conversation history.

Activities

Hover tour (5 min): Students read every node's detail, paying particular attention to the decision diamonds, which carry the diagnostic criteria.
Route the scenarios (7 min): Students route four tasks through the tree — a one-line rewrite request, a question about a 40-page requirements document, a 30-turn debugging session, and a task that needs three internal acronyms defined.
Recognise the reset (3 min): Ask students to describe, from experience, what a conversation feels like right before it needs a reset. Map their symptoms onto the "History getting long?" node.

Discussion Questions

The diagram treats prompt compression as the answer when minimal context is needed. Why is compression still worth doing when you have plenty of window to spare?
Four paths converge at "Submit Prompt." What does that convergence tell you about how these techniques relate to each other?
Resetting a conversation feels like losing progress. What does the "Summarize + Reset Context" step preserve, and what does it genuinely lose?
Where in this tree would a retrieval-augmented generation system sit?

Assessment

Can the student route an unfamiliar task to a leaf and name the technique?
Can the student state the symptom that triggers a context reset?
Can the student explain why the model being stateless makes history management the user's responsibility?