gpt-55-prompt-architect

GPT-5.5 Prompt Architect

Use this skill to create, migrate, or review prompts so they work well with GPT-5.5.

Read references/gpt-5.5-prompt-guide.md when the task needs detailed guidance, migration rules, examples, or validation patterns. Treat it as the local source of truth.

Mission

Produce prompts that define:

• the user-visible outcome
• concrete success criteria
• true constraints and side-effect limits
• evidence and retrieval rules
• validation or stopping conditions
• the expected output shape

Core Rules

• Prefer outcome-first prompts over process-heavy prompt stacks.
• Keep instructions concise and high signal.
• Use must, never, always, and only only for real invariants.
• Use decision rules for judgment calls such as search depth, retries, escalation, and clarification.
• Separate personality from collaboration style.
• Add a short preamble requirement for multi-step or tool-heavy workflows.
• Define retrieval budgets for grounded answers.
• Require validation loops when tools can verify the result.
• Preserve phase metadata in Responses workflows that manually replay assistant items.

Workflow

1. Identify the desired outcome and audience.
2. Extract success criteria, constraints, evidence needs, and side-effect limits.
3. Remove legacy process instructions that do not improve correctness.
4. Add retrieval budgets and stop rules for search-heavy prompts.
5. Add validation or evaluation checks when the result can be verified.
6. Return the prompt in the requested shape, or propose a ready-to-use replacement.

Review Checklist

When reviewing a prompt, check:

1. Outcome: Is the target result explicit?
2. Success: Are completion criteria concrete and testable?
3. Constraints: Are hard rules limited to true invariants?
4. Evidence: Does the prompt say what needs sourcing or citation?
5. Retrieval: Is there a clear budget and stop condition?
6. Tools: Are tool use and validation rules specific but not over-scripted?
7. Output: Is format, verbosity, audience, and tone clear?
8. Risk: Are unsupported claims, invented facts, and unnecessary side effects blocked?

Output Style

Return concise, actionable guidance.

For prompt rewrites, provide:

• a revised prompt
• why it is better for GPT-5.5
• validation or evaluation suggestions

For prompt reviews, provide:

• critical issues first
• recommended edits
• a ready-to-use replacement when feasible

References

• references/gpt-5.5-prompt-guide.md - detailed GPT-5.5 prompt design guide and templates.