Models, prompts and context — study roadmap
Models, prompts and context — study roadmap
Build a reproducible model-and-prompt decision for an incident-review assistant. Work through the four topic lectures, then use the local project to test the boundaries with synthetic data. This package covers Unit 2's five official objectives through ten explicit house skills; the house skill IDs below are teaching subdivisions, not additional official objectives.
| Order | Topic | Evidence you should produce |
|---|---|---|
| 1 | Model selection | Representative task set, recorded configuration and joint quality/latency/cost decision |
| 2 | Prompts and guardrails | Separate instructions and data, response contract and independent validator |
| 3 | Examples and reasoning | Targeted demonstrations, held-out comparison and checkable decision support |
| 4 | Context and reuse | Complete token ledger, preserved continuation state and justified reuse mechanisms |
1. Choose a candidate through a task experiment
Read the model-selection lecture and the selection guide. Define the task before choosing a model. Include meaningful exceptions, record supported settings such as effort, and evaluate all required operating limits. The retained guide's model examples are dated 5 September 2026; check current versions and availability before a real integration.
Exit check: explain why a strong overall result cannot waive a separately required exception floor. Project stage 1 exercises CCARP-U2.T1.LO1.S1 and .S2.
2. Make prompt boundaries explicit
Read the prompt-template lecture and prompting guidance. Identify the stable instructions, variable evidence and expected response shape. Then read prompt-injection mitigations: external text must remain untrusted data even when it contains convincing instructions.
Exit check: show a parseable output that your application rejects because its identity or proposed action is wrong. Project stage 2 exercises CCARP-U2.T2.LO2.S1 and .S2. Delimiters help organize content; independent checks still govern consequential use.
3. Add support for a measured failure
Start with a clear zero-shot contract. Use representative few-shot examples to demonstrate confusing boundaries. For reasoning-heavy decisions, test structured support through observable outcomes and evidence. Do not make successful integration depend on revealing private internal reasoning.
Exit check: distinguish a real task improvement from example leakage, a changed grading rule or a longer explanation. Project stage 3 exercises CCARP-U2.T3.LO3.S1 and .S2 using fixed outputs; it does not claim to measure a real model.
4. Budget context without losing meaning
Use the context-window guide to account for the complete request and response. Compare the original task state with a proposed summary: can the next step still identify the target, evidence, pending issues and approval scope?
Then distinguish prompt caching, modular project rules and Skills. They address input reuse, maintainability and specialist activation respectively; one mechanism does not prove another is active.
Exit check: produce a complete token ledger, reject an over-broad summary and explain where stable versus variable content belongs. Project stage 4 exercises CCARP-U2.T4.LO4.S1, .S2, CCARP-U2.T4.LO5.S1 and .S2.
Practice and revisit
Use each lecture's companion quiz immediately after that topic, then revisit its cards after a delay. The 28 quiz items are separate from the 36 exam-bank questions. The reference document supports review; the project requires applying the decisions. All model performance, token ceilings, cost units and policy rules supplied in the exercises are synthetic house fixtures. The full seven-domain exam bank and mock papers have separate completion gates.