Claude Models, Prompting & Context Engineering

Knowledge of

• The three model tiers -- Haiku (fastest/cheapest), Sonnet (balanced), Opus (most capable) -- and the typical task profile each tier fits
• The quality vs. latency vs. cost tradeoff triangle as the frame for resolving model choice against a specific requirement, not in the abstract
• Mixed-model designs that route cheap, well-defined steps to Haiku and hard reasoning/orchestration steps to Sonnet or Opus
• Extended (adaptive) thinking as a capability lever that trades billed output tokens and latency for better performance on hard reasoning tasks
• The risk that a new model release can change behavior, requiring pinned versions and re-run evals before upgrading

Skills in

• Right-sizing model selection to task complexity instead of defaulting to the most capable model
• Designing mixed-model pipelines that assign each step to the tier that fits its difficulty and latency/cost budget
• Reserving extended thinking for steps whose reasoning depth genuinely pays for the added tokens and latency
• Pinning model versions and re-running evaluations before adopting a new model release

Knowledge of

• The separation of stable rules/role/tone/constraints (system prompt) from the specific request and per-request data (user message)
• Why this system/user separation also keeps the stable prefix cacheable
• Templates as a way to parameterize variable parts while keeping the scaffold constant, improving consistency and cache hit rates
• Prompt-level guardrails as probabilistic behavioral boundaries, distinct from deterministic controls (hooks, permission scoping, output validation) that sit around the model
• The need to delimit and sanitize untrusted input so injected text is not read as an instruction

Skills in

• Placing stable rules, role, and constraints in the system prompt and per-request specifics in the user message
• Building parameterized prompt templates that keep the scaffold constant across requests
• Writing guardrail instructions that set behavioral boundaries (refusals, content policy, escalation to a human) while recognizing their probabilistic nature
• Pairing prompt-level guardrails with deterministic controls for destructive or high-stakes actions rather than relying on the prompt alone
• Delimiting and sanitizing untrusted input passed into a prompt

Knowledge of

• Zero-shot, few-shot, and chain-of-thought as distinct techniques suited to different task profiles
• Instruction clarity and specificity as more effective than generic instruction lists
• Positive framing of instructions (e.g., 'respond only in JSON matching this schema') over long negative lists
• Placement effects: with long inputs, the key instruction is more effective near the end, after the material
• Deliberate iteration against an eval set as the mechanism for improving a prompt over time

Skills in

• Matching zero-shot, few-shot, or chain-of-thought technique to task difficulty and ambiguity
• Writing positively framed, specific instructions instead of long negative constraint lists
• Placing the key instruction after long input material rather than before it
• Iterating on prompts one change at a time and measuring against an eval set
• Avoiding reflexive addition of few-shot examples or chain-of-thought when the task does not warrant the added tokens/latency

Knowledge of

• The context window as a shared budget across system prompt, conversation history, tool definitions, tool results, retrieved documents, and the response
• Context rot/drift: bloated context degrading quality as the model loses the thread among low-signal tokens
• Pruning, compaction, and isolation as curation techniques for dropping stale output, summarizing older turns, and moving heavy subtasks into a subagent's own window
• Progressive disclosure: loading information as needed rather than front-loading everything
• The usage field (input/output/cache tokens) as the mechanism for tracking cost and catching context bloat before it hits the budget

Skills in

• Curating the context window to the smallest set of high-signal tokens rather than relying on raw capacity
• Pruning large or stale tool output that no longer informs the task
• Compacting older conversation turns into a summary recap instead of carrying the full transcript
• Isolating heavy subtasks into a subagent with its own context window to keep the main context clean
• Tracking the usage field to model cost and detect context bloat before hitting the budget

Knowledge of

• Prompt caching: reusing a large stable prefix (system prompt + policy document + few-shot block) across requests, where cache reads are cheap and writes carry a slight premium
• The ordering principle: stable content first, dynamic content last, to maximize the cacheable prefix length
• Modular prompts: composing prompts from reusable, independently versioned fragments (role, policy, format spec) instead of duplicating text
• Agent Skills as reusable instructions/procedures authored once and reused across apps and teams without a running service
• The distinction between prompt caching (reusing a stable prefix within a model call) and retrieval (fetching external documents)

Skills in

• Ordering prompts so stable content comes first and dynamic content last to maximize the cacheable prefix
• Composing prompts from independently versioned, reusable fragments rather than duplicating text
• Packaging reusable capabilities as Agent Skills for reuse across applications and teams
• Avoiding truncation of a needed policy document to save tokens when caching would preserve both correctness and savings
• Distinguishing when to reach for prompt caching versus retrieval based on whether content is a stable prefix or an externally fetched document