CCDV-F curriculum

Requirements arrive as business sentences. Find the single deciding phrase, separate hard constraints from soft preferences, and let the constraint settle .

Your model can change without your deploy. That makes version pinning, an evaluation gate, and treating prompts as reviewed artefacts the operational core .

The API is stateless, so your app owns history and pays for it on every turn. Branch on stop_reason before reading content. And a cache checkpoint only wor.

An LLM call is still a slow network call. Async paths, correct retry scope, idempotent downstream writes, and a single place where the model is referenced .

Choose the surface the requirement implies, keep untrusted data out of instruction space and clearly tagged, scope session state per user, and design schem.

Four artefacts change behaviour — project instructions, permissions config, model pin, and prompts. All four belong in version control, and the model pin p.

The context window is shared by everything in the request including the response. Temperature 0 reduces variation but never guarantees identical output.

The SDKs wrap REST and add ergonomics, not capability. Knowing which layer an error came from tells you where to fix it. Streaming is a chunked response.

Profile the workload on volume, boundedness, and stakes, then justify the tier from those properties. Over-serving is as wrong as under-serving. Pin the ve.

Measure, then pull levers in order: cache, trim, right-size, batch. The first two are quality-neutral; the third trades and needs an eval check. Verify cac.

Enumerable steps mean a workflow, full stop — a conversational interface is not evidence of agency. Delegate to subagents for context isolation, not speed.

The model requests, your code executes — that gap is where approval, validation, and hooks live. A hook enforces what a prompt can only request. Managed ve.

Framework choice is decided by state management needs, control granularity, ecosystem fit, and team familiarity. Popularity and recency are distractors the.

Bloat is wasted tokens; drift is lost instructions. Cost tells you which one you have. Pruning and compaction fix bloat, reinforcement and isolation fix dr.

Say what to do rather than what to avoid, and put each instruction where its scope is obvious — durable behaviour in the system prompt, tool-specific const.

Extract, validate against a schema, and retry with the error fed back. A wrong answer is fluent and carries no warning, so the check has to live in code — .

Tool descriptions are the primary mechanism Claude uses for tool selection. When misrouting comes from weak descriptions, improving them is the first fix —.

MCP turns M×N integrations into M+N. Servers expose resources, tools, and prompts. Transport follows deployment — stdio for local, network for remote — and.

Built-in, custom tool, Skill, MCP server — each is right at one scope. Skills guide, tools execute. And misrouting has two causes: confusable descriptions .

Indirect injection arrives inside content the model was asked to read, often with no malicious user present. Tagging untrusted content makes it visible as .

No single guardrail is sufficient — combine behavioural, detective, and preventive layers. Scope agent credentials to the task, because a manipulated agent.

Hooks are the only deterministic guardrail — code at a fixed point in the loop whose decision does not depend on the model. Any requirement phrased as "reg.

Keys live server-side in a secret manager, never in client code, a prompt, or a tracked file. Rotate on a schedule. Verify access at call time. And monitor.

Commands are invoked, Rules always hold. Team standards belong in a project-level CLAUDE.md, not personal settings. Headless mode is what runs in CI where .

Identify the error type, attribute it to the integration layer or the model, trace back to the first divergence rather than the obvious symptom, then match.