llm-prompting

Effective LLM Prompting

Use when getting an LLM to produce reliable output — structured or free-form — rather than hoping a vague instruction happens to work. Prompts are code: they drift, they regress, and they need tests.

Pair this with the llm-json-output skill when the target must be machine-parseable JSON. Rule of thumb established there: if the provider supports a native structured-output / JSON-schema mode, use that first and treat free-form prompting as the fallback. This skill is about making the fallback (and any free-form generation) as reliable as possible.

The five-part skeleton

1. Role — one line of expertise/framing: "You are a senior Python reviewer."
2. Task — one clear imperative sentence stating the deliverable.
3. Input — the data the model operates on (small, self-contained, as ground truth).
4. Output format — exact schema, delimiters, or shape; state "no other text".
5. Few-shot examples — 2-3 ideal in/out pairs only when ambiguity exists.

Worked example, applied end-to-end

Goal: extract {port, enabled, name} from a config blob, drift-free.

[SYSTEM]
You are a config parser. Extract exactly three fields from the input into a
JSON object. Follow the schema, never invent values, never add prose.

[USER]
Input:
  service: metrics, listen: 0.0.0.0:9100, tls: off
Return exactly: {"name": string, "port": int, "enabled": bool}
Example: "grafana" -> {"name": "grafana", "port": 3000, "enabled": true}

Then validate: json.loads, type-check the three fields, and if anything is missing or malformed re-prompt with the exact error (see llm-json-output's retry loop). The negative example below ("tls: off" → enabled: false) sharpens behavior more than a second positive does — show one counter-case when the mapping is non-obvious.

Decision table: which prompting mode when

Mode	When	Why
Native structured output / JSON-schema	Provider supports it and output is machine-read	Constrains generation; no parsing hacks. Always first choice for llm-json-output work.
Zero-shot	One obvious task, correct output is unambiguous	Cheapest; no example tokens; fewer chances to overfit
Few-shot (2-3)	Ambiguous mapping, custom format, or rare edge, free-form output	Shows the exact shape; cost is example tokens and overfitting risk
Choose-then-run (agentic)	Task needs search/tools first	Same skeleton, but input comes from tool results, not the prompt