Llm app engineering

LLM App Engineering

Design LLM applications as systems, not prompts: what goes into the context window, what the model is allowed to do, and what happens when it's wrong. The model is the one component you don't control — engineer everything around it. This skill designs and builds; proving a change helped is ai-evaluation's job, and no design below is "done" until it has an eval harness.

Workflow

Step 1: Choose the Simplest Sufficient Architecture

Escalate only when the previous rung measurably fails:

1. Single prompt — the task fits in one call with instructions + input.
2. + RAG — the answer needs knowledge that doesn't fit in the prompt or changes independently of deploys.
3. + Tools — the model must act (query, search, write) — not just answer.
4. Agent loop — the path to the goal is genuinely unpredictable in advance. Most "agents" that fail are workflows that should have been a fixed pipeline with one or two LLM steps.

Step 2: Engineer the Prompt as an Interface

• Structure: role and constraints → domain context → task → output format. Put stable content first (prompt caching) and the volatile input last.
• State what to do on failure: "if the context doesn't contain the answer, say so" — the unhappy path is where LLM apps lose user trust.
• Version prompts in git like code; a prompt change is a deploy and goes through the same eval gate (ai-evaluation).

Step 3: RAG — Design Retrieval Before Generation

Most bad RAG answers are retrieval failures wearing a generation costume.

• Chunking: split on document structure (headings, sections, one ticket+resolution per chunk), not fixed character counts; attach metadata (source, date, product area) for filtering and citations.
• Retrieval: hybrid (BM25 + vector) as the default — pure vector search misses exact identifiers, SKUs, and error codes; filter by metadata, then rerank the candidate set (retrieve ~20–50, keep the top 5–10).
• Context budget (decide, don't default): how many chunks fit, in what order (most relevant near the question), the relevance floor below which a chunk is dropped, and the behavior when nothing clears it — say "I don't know" rather than pad the context with noise.
• Grounding prompt: answer only from the provided context, cite which source supports each claim, refuse when unsupported.

Step 4: Agents — Design the Tool Surface and the Loop

• Tools: few and sharply distinct beat many and overlapping — every tool description competes for the model's attention on every step. Name by intent (search_orders, not api_call); make each tool's failure return a message the model can act on, not a stack trace.
• Decomposition: long tasks fail as one monolithic loop with 25 tools; split into phases or sub-agents, each with a scoped toolset, and pass structured summaries between them — not raw transcripts.
• Guardrails: cap iterations, budget tokens per task, and validate side-effecting tool calls (or require confirmation) — an agent retrying a failed write in a loop is the classic production incident.

Step 5: Memory — Externalize State

The context window is a scarce cache, not a database. For long-horizon work: keep a plan/scratchpad file the agent re-reads and updates; summarize or compact stale history instead of letting it scroll off; store durable facts (user preferences, decisions) outside the window and retrieve them like RAG.

Step 6: Wire In Evaluation Before Shipping

Hand the design to ai-evaluation: a golden dataset (including unanswerable and adversarial cases), RAG-stage metrics (retrieval vs generation scored separately), agent task-success rate, and a CI regression gate. "The answers feel better" is not evidence; a design without an eval harness is a prototype.

Deeper pattern detail (chunking table, reranking, memory patterns, failure modes): references/patterns.md.

Principles Applied

• KISS: every architecture rung you skip is a class of failure you never have to debug. Workflows beat agents when the path is predictable.
• YAGNI: no vector DB before checking whether the corpus fits in the prompt; no agent framework before a plain loop with 3 tools fails.
• SRP for tools: one tool, one capability — overlapping tools force the model to guess, and it will guess differently each run.

Cross-Skill References

• ai-evaluation — golden datasets, RAG metrics, judges, eval gates (the measurement half of every step above)
• ml-model-deployment — serving, monitoring, and drift once the app ships
• api-design — the contract around the LLM feature (streaming, errors, timeouts)
• security-audit — prompt injection, data exfiltration via tools, output handling
• architecture-design — ADRs for costly-to-reverse choices (vector store, framework)