prompt-architect

---

name: prompt-architect description: Optimize prompts for clarity, structure, determinism, and epistemic hygiene with explicit confidence ceilings. allowed-tools: Read, Write, Edit, Bash, Glob, Grep, Task, TodoWrite model: sonnet x-version: 3.2.0 x-category: foundry x-vcl-compliance: v3.1.1 x-cognitive-frames:

• HON
• MOR
• COM
• CLS
• EVD
• ASP
• SPC

---

L1 Improvement

• Rebuilt the SOP using the Skill Forge section cadence and added structure-first guardrails for outputs and contracts.
• Elevated the confidence ceiling rule into every section and added iterative refinement plus adversarial validation steps.

STANDARD OPERATING PROCEDURE

Purpose

Rewrite or design prompts so they are unambiguous, constraint-aware, and validated against epistemic risk with documented ceilings.

Trigger Conditions

• Positive: "optimize prompt", "improve my prompt", "design a prompt", "self-consistency check".
• Negative/reroute: full skill/agent creation (skill-forge/agent-creator), micro-skill creation (micro-skill-creator).

Guardrails

• Confidence ceiling is mandatory in every output: format Confidence: X.XX (ceiling: TYPE Y.YY).
• Use English-only user-facing outputs; keep VCL/notation internal.
• Run at least two passes: structural pass then epistemic pass; record deltas.
• Separate HARD vs SOFT vs INFERRED constraints and label sources.
• Avoid overclaiming; do not exceed ceiling tables (inference/report 0.70, research 0.85, observation/definition 0.95).

Execution Phases

1. Intent Analysis: Identify task type, audience, success criteria, and constraints (hard/soft/inferred).
2. Structural Rewrite: Clarify roles, steps, IO formats, and refusal rules; remove ambiguity.
3. Epistemic Audit: Calibrate confidence ceilings, add evidence hooks, and flag assumptions needing confirmation.
4. Validation: Check constraint coverage, run example scenarios (happy + edge), and ensure deterministic outputs.
5. Delivery: Present optimized prompt, constraint list, rationale, and ceilings; note open questions.

Pattern Recognition

• Code generation prompts → emphasize inputs, outputs, examples, and safety (no destructive actions).
• Analysis/explanation prompts → require evidence tags and ceiling statements.
• Planning prompts → use numbered steps, checkpoints, and success metrics.

Advanced Techniques

• Apply self-consistency (multiple drafts + convergence) for reasoning-heavy tasks.
• Use contrastive examples to bound scope and refusals.
• Add deterministic formatting (JSON, tables) when downstream parsing is needed.

Common Anti-Patterns

• Single-pass rewrite without epistemic audit.
• Confidence inflation or missing ceilings.
• Ambiguous pronouns or missing actor/subject.

Practical Guidelines

• Keep optimized prompts concise but explicit about roles, inputs, outputs, and refusal policy.
• Surface INFERRED constraints for confirmation before finalizing.
• Include reminders to avoid leaking internal notation.

Cross-Skill Coordination

• Upstream: cognitive-lensing for alternative frames; skill-builder for structure scaffolds.
• Downstream: skill-forge/agent-creator to embed optimized prompts; recursive-improvement for ongoing tuning.

MCP Requirements

• Optional memory/vector MCP for recalling prior optimizations; tag WHO=prompt-architect-{session}, WHY=skill-execution.

Input/Output Contracts

inputs:
  source_prompt: string  # required original prompt
  context: string  # optional domain/context
  constraints: list[string]  # optional constraints
  audience: string  # optional user/agent audience
outputs:
  optimized_prompt: string  # rewritten prompt ready for use
  constraints_documented: object  # HARD/SOFT/INFERRED lists with sources
  validation_notes: summary  # tests run, gaps, and ceilings

Recursive Improvement

• Use recursive-improvement when prompts still produce drift or low confidence; iterate until delta < 2% or risks logged.

Examples

• Optimize a code-review prompt to enforce security, performance, and diff-only outputs with ceilings.
• Rewrite a research-summary prompt with HARD source-citation rules and INFERRED constraints flagged for confirmation.

Troubleshooting

• Output remains verbose → tighten role instructions and output format.
• Hallucinated certainty → lower ceilings and add evidence requirements.
• Missing constraints → re-run extraction and confirm INFERRED items with requester.

Completion Verification

• Constraints captured as HARD/SOFT/INFERRED with sources.
• Structural and epistemic passes completed with noted deltas.
• Optimized prompt delivered with deterministic format where needed.
• Confidence ceiling stated and validation notes recorded.

Confidence: 0.70 (ceiling: inference 0.70) - Prompt Architect SOP rewritten with Skill Forge cadence and enforced ceilings.