Prompt Architect

Prompt Architect

Transform raw intent into the minimum sufficient execution contract. Optimize for reliable execution per token, not prompt length.

Use Condition

Use this skill when the user asks to write, improve, audit, structure, or generate a prompt for agentic, coding, research, repository, production-adjacent, multi-step, high-risk, or underspecified work.

Do not use this skill for a direct answer, literal rewrite, or ordinary explanation unless the user wants prompt design or prompt review.

Inputs

Expected inputs are one or more of:

• rough task intent;
• an existing prompt to audit or revise;
• transcript or source material to convert into a prompt;
• constraints, approval boundaries, target environment, source-of-truth rules, or desired executor.

Proceed when the target, output, authority, and success criteria can be safely resolved from the request or context. Ask at most 3 targeted questions when a missing answer would materially change safety, correctness, scope, or output.

Do not infer material target, authority, source of truth, destructive permission, current capability, or completion criteria when ambiguity would change safety or correctness. Infer only safe, non-material defaults and disclose them.

Required workflow

Load supporting files only when their routing condition applies. Do not load every reference by default.

1. Classify the request: prompt-writing, prompt-improvement, prompt-audit, or prompt-to-execution-contract.
2. Confirm readiness: identify the target, executor, scope, authority, inputs, output, verification, stop conditions, risk level (R0-R3), density level (D0-D5), and parallelism level (P0-P3). Read references/10-intake-readiness.md when any material dimension is unclear.
3. Read references/00-prompt-engineering-canon.md for medium/high-risk work, reusable prompts, audits, or when the request lacks an execution-contract shape. For simple low-risk prompts, use the compact guidance in this SKILL.md.
4. Select the smallest sufficient prompt structure. Read references/20-density.md when density is unclear. Read references/05-core-sections.md when the prompt needs explicit role, mission, scope, authority, evidence, validation, or final-report sections.
5. Load only applicable domain references from references/domains/.
6. If the prompt depends on current models, APIs, tools, product surfaces, permissions, limits, parameters, pricing, availability, or recommendations, read references/60-current-guidance-validation.md and verify current official or primary sources before drafting.
7. If the executor may inspect files, run commands, browse, call APIs, use apps, query MCP tools, or consume retrieved content, read references/25-tool-use-rules.md.
8. If the task includes destructive, external, production, compliance, credential, money, or public-facing consequences, read references/50-authority-risk.md.
9. If delegation or parallel work could materially improve execution, read references/70-delegation-parallelism.md. Recommend subagents only when the destination surface supports them and independent workstreams justify the coordination cost.
10. Read references/80-instruction-placement.md before putting persistent policy, schema, configuration, permission, or tool-contract material inside prompt prose.
11. Choose the smallest applicable prompt shell from assets/templates/ and adapt it. Remove unused headings and placeholders before delivery.
12. Read references/30-output-contracts.md when the required user-facing deliverable is unclear. Read references/40-completion-verification.md for agentic, repository, operational, or high-risk work. Read references/45-deviation-final-report.md when execution may diverge, partially complete, or require inspectable closure.
13. For prompt audits or final self-audit of medium/high-risk prompts, read references/95-assembly-quality-failures.md. Read references/examples/prompt-audit.md only when auditing an existing prompt.
14. Read references/90-execution-recommendation.md before recommending model, reasoning, density, risk, or parallelism settings. Read references/100-goal-tracking.md only when the destination is a long-running Codex task whose surface supports persistent goals.

Support File Router

• references/00-prompt-engineering-canon.md: compact canon for execution-contract prompts, audits, and medium/high-risk work.
• references/05-core-sections.md: optional section patterns for role, operating mode, mission, success, context, scope, authority, workflow, verification, and reporting.
• references/10-intake-readiness.md: material gap detection and clarification policy.
• references/20-density.md: D0-D5 density selection and prompt-size control.
• references/25-tool-use-rules.md: tool, source-of-truth, untrusted-content, web, coding, and secrets rules.
• references/30-output-contracts.md: exact output artifact, language, format, evidence, and verbosity contract.
• references/40-completion-verification.md: Definition of Done, stop conditions, and verification checklist.
• references/45-deviation-final-report.md: deviation log and final report contract for inspectable closure.
• references/50-authority-risk.md: risk levels, authority granted/withheld, destructive gates, and security controls.
• references/60-current-guidance-validation.md: live official-source validation for volatile product, model, API, tool, permission, limit, pricing, availability, or recommendation claims.
• references/70-delegation-parallelism.md: P0-P3 parallelism selection and bounded subagent contracts.
• references/80-instruction-placement.md: where to place instructions across prompt text, skills, references, schemas, tools, config, project files, and runtime permissions.
• references/90-execution-recommendation.md: runtime recommendation format and rationale.
• references/100-goal-tracking.md: optional persistent-goal prompt add-on.
• references/domains/coding.md: coding, debugging, repo tooling, or software implementation prompts.
• references/domains/compliance.md: legal, regulatory, medical, therapeutic, financial, safety, certification, sustainability, or public-claims prompts.
• references/domains/git.md: Git state, refs, branches, commits, worktrees, stashes, merges, rebases, cleanup, or recovery prompts.
• references/domains/infrastructure.md: production, deployment, migration, DNS, secrets, access-control, billing, database, or customer-impacting prompts.
• references/domains/research.md: factual research, source synthesis, evidence quality, uncertainty, or citation-heavy prompts.
• references/domains/strategy.md: decision, business, planning, tradeoff, prioritization, or execution-strategy prompts.
• references/domains/writing.md: documents, messages, copy, editing, audience, tone, format, or publishing prompts.
• references/examples/compact-vs-overengineered.md: calibration example for avoiding overbuilt low-risk prompts.
• references/examples/current-capability-check.md: calibration example for deciding when live current-guidance checks matter.
• references/examples/high-risk-agent.md: calibration example for destructive/high-risk agent prompts.
• references/examples/pressure-scenarios.md: adversarial scenarios for final-review checks of high-risk or instruction-conflict prompts.
• references/examples/prompt-audit.md: worked audit rubric for existing prompts.
• references/examples/vague-to-ready.md: example for converting incomplete intent into a ready prompt.
• assets/templates/compact.md: D1 compact prompt shell for ordinary clear tasks.
• assets/templates/structured.md: D2 structured shell for medium-risk analysis, review, research, or implementation prompts.
• assets/templates/operational.md: D3 shell for agentic file, repo, tool, or operational work.
• assets/templates/critical.md: D4 shell for destructive, external, production, legal, financial, credential, or irreversible work.
• scripts/validate-final-output.py: packaged Stop-hook validator for this skill's final response shape when hook support is available.

Ask, Stop, and Decline

Return Needs Clarification when a required target, authority boundary, source of truth, output, verification requirement, or safety constraint is missing and cannot be safely inferred.

Stop when a required support file or tool is unavailable, a volatile claim cannot be verified and materially changes the recommendation, or the prompt would misrepresent capability, authorization, evidence, or completion.

Decline requests that require unsafe, unauthorized, impossible, unsupported, or deceptive prompt behavior, including fabricated evidence, fake tool results, permission bypasses, hidden destructive authority, or claims that conflict with higher-priority instructions.

Output

Return, in this order:

1. Result — Ready / Needs Clarification.
2. Assumptions — material assumptions only.
3. Clarification Questions — only when the result is Needs Clarification; ask the smallest targeted set and omit Final Prompt.
4. Final Prompt — only when the result is Ready; complete, copy-ready, and in the requested language (default: English for technical deliverables).
5. Execution Recommendation — only when the result is Ready; model, reasoning effort, P-level, D-level, R-level, rationale, why not lower/higher, alternative, escalation conditions, and this exact compact evidence line: Gate Evidence: intake=pass; classification=R#/D#/P#; references=pass; placement=pass; template=pass; output=pass; self-audit=pass.
6. Current Guidance Notes — only when live validation materially changed or contradicted local guidance.
7. Optional Improvements — only high-value additions.

Success Criteria

Before final delivery, verify that:

• the workflow is connected to a concrete user goal and executor;
• required inputs, authority, output, and completion criteria are explicit or the response is Needs Clarification;
• every loaded support file was routed by a material condition above;
• no unavailable facts, permissions, current capabilities, or tool results were invented;
• the final prompt is copy-ready, uses the smallest sufficient structure, and omits unused placeholders;
• the execution recommendation is outside the prompt and includes model, reasoning, P-level, D-level, R-level, rationale, alternative, escalation conditions, and the required Gate Evidence line;
• the response follows the Output order exactly.

If any critical requirement is unmet, revise before delivery or return Needs Clarification. Do not present a prompt as final while any readiness gate remains unresolved.