Prompt Craft

Prompt Craft

Concept of the skill

A prompt is an executable instruction contract: stable rules and examples define the function, variable user content supplies arguments, and evals decide whether revisions improved behavior.

Coverage

This skill covers portable prompt design for LLM-backed tasks and agents:

• Instruction hierarchy and message roles: where stable policy, developer rules, user requests, examples, and variable input belong.
• Prompt anatomy: task statement, context, constraints, examples, input delimiters, output cue, refusal or escalation rule, and verification plan.
• Context placement: what belongs in the prompt template versus the surrounding context, retrieval, memory, tool output, or runtime config.
• Few-shot examples: choosing boundary examples, counterexamples, and diverse cases that teach the pattern without leaking private data.
• Output-format discipline: schema-shaped responses, enum outputs, concise prose, tables, JSON, Markdown, validation, retries, and fallback behavior.
• Positive constraints and negative boundaries: replacing vague prohibitions with explicit target behavior while retaining safety-critical refusals.
• Reasoning guidance: when to ask for concise rationale or internal carefulness, when to avoid visible reasoning, and when provider effort controls are the better lever.
• Adversarial-input resistance: separating instructions from user data, validating outputs, limiting tool authority, and routing security-sensitive controls to guardrails.
• Iterative improvement: hold eval cases fixed, change one prompt surface at a time, measure behavior, and document the delta.
• Provider differences: OpenAI, Anthropic, Gemini, and open-weight models vary in role semantics, structured-output APIs, reasoning controls, and prompt sensitivity.

Philosophy of the skill

A prompt is a specification, not a wish list. The goal is not the shortest possible wording or the most elaborate incantation; the goal is the smallest inspectable instruction contract that reliably produces the required behavior on realistic inputs.

Prompt craft is also not a substitute for system design. A model cannot reliably follow context it never received, validate JSON the runtime never checks, refuse tool calls the application allows, or prove quality on examples that were never run. Good prompt work names those boundaries. It tells the model what to do, tells the runtime what must be validated, and tells the maintainer what evidence would justify keeping a revision.

The largest failure mode is prompt-and-pray: write a prompt, run one plausible example, and ship. Prompt engineering becomes engineering only when a stable eval set, format checker, or trace review can show whether the change improved the target metric without causing regressions.

Prompt Anatomy

A non-trivial prompt usually has these layers. The order is a starting point, not a universal law; provider docs and eval results should decide the final shape.

Instruction Hierarchy And Message Roles

Place instructions where the runtime's authority model expects them. OpenAI's prompt-engineering docs describe developer messages as the system's rules and business logic, while user messages carry inputs and configuration. Anthropic docs similarly emphasize clear, direct instructions, examples, XML-style structure, and model-specific prompt tuning. Gemini docs emphasize direct structure, constraints, response format, context, and prompt iteration.

Use these defaults unless provider guidance or eval evidence says otherwise:

• Stable identity, product rules, safety boundaries, and business logic belong in the highest-authority instruction surface available.
• User input, retrieved documents, tool observations, emails, tickets, code snippets, and uploaded files belong in data sections, not in system or developer instructions.
• Examples should be visibly separated from live inputs so the model does not confuse demonstration with task data.
• Long-lived prompt objects should be versioned; every production change should be traceable to an eval, metric, or observed failure.
• Do not rely on wording alone for tool safety. Tool availability, arguments, side effects, and permissions need runtime guardrails.

Prompt Versus Context Boundary

Use prompt-craft for the wording and structure of instructions. Use context-engineering when the problem is the information environment around the prompt.

If changing the prompt cannot plausibly fix the observed behavior without changing retrieved data, tool authority, memory, schema validation, or runtime state, switch skills.

Output-Format Discipline

For structured output, the prompt should describe the contract and the runtime should enforce it. Do not treat model obedience as validation.

Use a layered approach:

1. State the output type and allowed fields or labels.
2. Prefer native structured-output or JSON-schema features when the provider supports them, especially for complex schemas.
3. Include one small example that demonstrates the exact shape.
4. Put a final output cue at the end of the prompt.
5. Validate the result outside the model.
6. Retry once with the validation error only when retrying is safe.
7. Fall back to a clear error, clarification, or human review path when validation still fails.

Example pattern:

Return one JSON object with this shape:
{ "label": "SPAM" | "HAM" | "UNDETERMINED", "confidence": 0.0-1.0 }

Output no prose outside the JSON object.

If the input cannot be classified from the supplied evidence, use "UNDETERMINED".

Avoid exact reliability claims such as a fixed breakage percentage unless you have measurements for the model, prompt, sampling settings, and input distribution in front of you.

Positive Constraints And Negative Boundaries

Negative instructions are sometimes necessary, especially for safety, privacy, authorization, and output exclusions. But a prohibition alone often leaves the model without a target behavior. Pair negative boundaries with positive replacements.

The principle is salience plus target: reduce unnecessary mention of forbidden content, and give the model an allowed behavior it can execute instead.

Few-Shot Example Selection

Few-shot examples teach pattern and boundaries. They are not a storage place for every rule.

Choose examples that:

• Match the real task closely enough to transfer.
• Cover category boundaries, ambiguous cases, and known failure modes.
• Vary length, style, vocabulary, and input shape so the model does not learn a superficial cue.
• Include at least one negative or counterexample when confusion between categories is likely.
• Demonstrate the exact output contract, including labels, field names, citation style, or refusal path.
• Are newly authored or scrubbed so private, customer, credential, or local-path content is not embedded in a reusable prompt.

Use zero-shot for simple, well-defined tasks. Use one example to establish shape. Use three to five examples for nuanced classification or style. More examples can help exemplar-driven tasks, but if the prompt needs many examples to work, consider whether the task should be decomposed, retrieved examples should be selected dynamically, or the behavior belongs in a fine-tuned model, rules engine, or eval-guided workflow.

Reasoning Guidance

Do not ask the model to reveal hidden chain-of-thought. Ask for the visible artifact you need: a concise rationale, assumptions, cited evidence, decision table, calculation trace, or verification checklist. For models with explicit reasoning controls, tune the provider's effort or thinking budget instead of trying to coerce hidden reasoning through prose.

Use visible reasoning when it improves auditability:

• Decisions that need a short justification.
• Classification tasks where a label without evidence is hard to trust.
• Safety or authorization checks where the deciding rule must be inspectable.
• Multi-step tasks where the user benefits from a compact plan or progress update.

Suppress visible reasoning when it harms the product:

• Strict machine-readable output.
• User-facing answers where only the final result is needed.
• Sensitive workflows where exposing intermediate reasoning may reveal policy, private context, or attack surface.
• Latency-sensitive workflows where internal effort controls are the better knob.

Defence Against Adversarial Input

Any prompt that processes user-controlled or retrieved content is exposed to prompt injection. OWASP describes the core problem as natural-language instructions and data being processed together without clear separation, with impacts such as safety bypass, data exfiltration, prompt leakage, and unauthorized tool use.

Prompt-level defenses:

• Delimit user data and retrieved content with explicit section labels or tags.
• State that delimited content is data to analyze, not instructions to follow.
• Keep system, developer, and policy instructions separate from user-controlled text.
• Use allowlisted labels, schemas, and tool names where possible.
• Tell the model how to respond to attempted overrides, prompt extraction, hidden instructions, or tool manipulation.
• Avoid rendering model output as trusted HTML or Markdown without sanitization in applications that display it.

Runtime defenses that prompt-craft must not pretend to solve alone:

• Input validation and sanitization.
• Output monitoring and schema validation.
• Least-privilege tool permissions.
• Human-in-the-loop controls for side effects.
• Remote-content sanitization and retrieval hygiene.
• Security evals and adversarial tests.

When user input can cause tool calls, financial actions, data access, credential exposure, or customer-impacting changes, verify with guardrails.

Iterative Improvement Loop

Prompt changes need evidence. Use this loop for any prompt that matters:

1. Define success criteria before editing: accuracy, format compliance, refusal quality, latency, cost, tool-use precision, or user satisfaction.
2. Hold a representative eval set fixed. Include edge cases, hard negatives, adversarial cases, and known failure examples.
3. Run the current prompt and record baseline behavior.
4. Change one surface at a time: wording, context order, example set, output contract, role, reasoning guidance, or provider parameter.
5. Re-run the same eval set under the same model and sampling conditions.
6. Keep the change only if it improves the target metric without unacceptable regressions.
7. Record the changed surface, evidence, date, model, and residual risks.
8. If the failure is not prompt-controllable, route to context-engineering, guardrails, eval-driven-development, debugging, or model/runtime configuration.

A prompt that has not been measured may be a good draft. It is not yet an engineered prompt.

Provider-Specific Checks

Before finalizing a prompt, verify the provider/runtime details that affect behavior:

Provider docs can conflict in emphasis because models differ. Treat provider guidance as a hypothesis to test against your own eval set, not as a permanent universal rule.

Source Notes

• OpenAI grounds message authority, developer versus user message separation, structured prompt sections, prompt versioning, few-shot learning, eval-linked publication, and agentic prompt practices.
• Anthropic grounds success criteria before prompt engineering, clear and direct instructions, examples, XML-style structure, role prompting, effort controls, and measuring prompt changes against evals.
• Google Gemini grounds constraints, response formats, structured-output feature preference for complex JSON, context insertion, prompt iteration, content-order experiments, and parameter tuning.
• OWASP grounds prompt-injection attack classes and primary defenses such as input validation, structured separation, output validation, least privilege, monitoring, and human controls.

Verification

After applying this skill, verify:

• The prompt has one clear task statement and an explicit output contract.
• Stable instructions are separated from variable user input and retrieved content.
• User-controlled content is delimited and treated as data, not commands.
• Required context is present, and missing context is routed to context-engineering rather than hidden in wording.
• Few-shot examples, if used, cover boundaries and known failure modes without private data.
• Structured output is backed by provider features or runtime validation when correctness matters.
• Negative boundaries have positive replacement behavior where possible.
• Reasoning guidance asks for visible rationale only when useful and does not require hidden chain-of-thought disclosure.
• Prompt-injection and tool-use risks are checked with guardrails when user-controlled content can trigger actions or data access.
• The revised prompt was measured against a stable eval set or clearly labeled as an unverified draft.
• Eval, routing, freshness, and verification claims are no stronger than the evidence from this run.

Do NOT Use When