gbalabanli/prompt-improvement

Prompt Improvement

Use this skill to convert rough requests into intentional prompts that reduce ambiguity, preserve user intent, and improve output reliability.

Workflow

1. Run intake and classify prompt type.

• Capture the user's raw request, expected deliverable, domain, and context.
• Classify prompt type: coding, reasoning, extraction, research, or general.
• Restate intent in one sentence before refinement.

2. Run understanding self-check.

• Score understanding 0-5.
• If score <= 3, ask clarifying questions before building the final prompt.
• If score >= 4, proceed with explicit assumptions.

3. Ask targeted clarification questions.

• Ask only high-impact questions that change implementation direction or output quality.
• Ask 1-3 questions per round, for up to 2 rounds.
• Prioritize unresolved fields in this order:
  • model family (gpt, reasoning, codex) when known
  • goal
  • scope (in-scope and out-of-scope)
  • hard constraints (must-have limits)
  • optimization constraints (what to optimize for)
  • inputs/context (files, APIs, examples)
  • verification expectations (tests, lint, citations, sanity checks)
  • output contract (format and required sections)
  • success criteria (what "done" means)
• Use clarifying-question-bank.md.

4. Build a prompt spec.

• Normalize user answers into structured fields:
  • prompt type
  • model family
  • role
  • goal
  • context
  • inputs
  • scope
  • hard constraints
  • optimization constraints
  • verification
  • output contract
  • success criteria
  • non-goals
• Detect contradictions and resolve them explicitly.
• If unresolved ambiguity remains after 2 rounds, proceed with bounded assumptions.
• Use prompt-spec-template.md.

5. Generate the improved prompt using structured delimiters.

• Produce one final prompt with explicit sections and XML delimiters.
• Keep instruction order stable: role -> objective -> context -> inputs -> constraints -> execution rules -> output format -> verification.
• Default to canonical XML structure from codex-openai-prompt-improvement-spec.md.
• For coding, include minimal-diff and validation requirements.
• For extraction, enforce strict schema and "no guessing" behavior.
• Do not add unrelated implementation details not requested by the user.

6. Apply model-specific steering.

• For gpt: be explicit on structure and verbosity constraints.
• For reasoning: keep prompts direct and do not require chain-of-thought exposition.
• For codex: require reproducible validation steps and explicit quality bar (tests/lint/type checks where relevant).

7. Validate quality before finalizing.

• Score the prompt using quality-rubric.md.
• Required pass conditions:
  • intent preserved with no scope drift
  • explicit objective and output format
  • constraints are testable, not vague
  • verification expectations are included when relevant
  • wording is concise and token-efficient
• If score is below 8/10, revise once with targeted fixes.

8. Deliver refinement output.

• Return:
  • clarifying questions asked (if any)
  • compact prompt spec
  • final improved prompt
  • quality check score
  • assumptions and remaining risks

Output Rules

• Always distinguish hard constraints from optimization constraints.
• If key details are missing, ask questions first instead of guessing.
• Keep questions concise and actionable.
• Keep the final prompt deterministic, intentional, and parseable.
• Prefer XML-tagged sections unless user requests another schema.
• Preserve user intent; do not widen task scope without explicit user approval.
• Do not claim certainty when unresolved assumptions remain.

Defaults

• Clarification rounds: 2
• Questions per round: 1-3
• Quality threshold: >= 8/10
• Target comprehension confidence: >= 99% when required fields are complete
• Output style: xml-structured
• Prompt variants: final only (optional compact variant when requested)

References

• clarifying-question-bank.md
• prompt-spec-template.md
• quality-rubric.md
• codex-openai-prompt-improvement-spec.md