Prompt engineering
Use roles, hard constraints, output shapes (JSON, tables, step lists), and few-shot examples so models behave consistently and comparably across runs—version it like a checklist and treat it as the first layer of every agent task.
Category · Prompt engineering
Prompt engineering
5 skills Category 1 of 20
This category covers how to state intent clearly and wire knowledge in: reproducible templates, retrieval augmentation and context trimming, turning team conventions into triggerable SKILL docs, and exposing MCP tools so agents can call external systems. A practical order is to lock output shape and boundaries first, then add RAG and window management, then extend tooling.
In the hub sequence, this is the start of “prompts & agents”; pair it with code review and agent engineering to close the loop from prompts to review, orchestration, and evals. The five cards below match the main hub for this category.
Quick links
In depth
RAG retrieval
Tune chunk sizes, embeddings, reranking, and citation boundaries; define how to say “not found” so hallucinations are not mistaken for facts. Ties tightly to knowledge-base updates and access control.
Context window
Summarize history, prioritize files, and trim safely in long chats or huge repos—keep what decisions need, control latency and cost.
SKILL authoring
Encode repeatable workflows as structured skills: triggers, inputs/outputs, guardrails, and examples—works with Cursor rules and MCP tools so tacit conventions do not drift.
MCP server
Expose databases, internal APIs, and filesystems as MCP tools with auth, rate limits, and clear error semantics—the hard boundary of what “actions” are allowed beyond prompting.