메뉴
Create the MAPS Design phase artifact for an agent or multi-agent system. Use when defining roles, workflows, handoffs, memory, context, guardrails, approvals, escalation paths, test-first proof requirements, and system boundaries from an agent brief.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
MAPS Phase 3 Build skill that merges base /build-agent, incremental implementation, and test-driven development into one implementation workflow.
Run MAPS Phase 6 Deploy by wrapping /deploy-agent with GitHub Actions environments and Cloudflare deployment automation. Use when Codex needs GitHub to deploy a Cloudflare Pages, Workers, or Agents project; replace manual Wrangler deploys; configure workflow gates, encrypted secrets, smoke tests, release evidence, and rollback records.
Run enhanced MAPS M8 Experience Design for websites, SaaS products, dashboards, landing pages, and agent-facing product surfaces. Use when Codex needs to wrap /design-experience with website design references, component-system choices, accessibility primitives, UX/service-design reasoning, responsive layout, page hierarchy, and production-ready UI handoffs.
Run MAPS M8 Experience Design for multi-agent products, or optional A2 Experience Design for single-agent products without an M layer. Use when Codex needs to turn product intent, orchestration behavior, agent behavior, and product constraints into user journeys, surfaces, information architecture, interaction states, accessibility requirements, and system-evaluation handoffs.
Run MAPS Phase 5 Evaluate by wrapping the base Evaluate process with LangSmith, Inspect AI, and Phoenix. Use when Codex needs to create eval suites, trace-backed eval runs, safety/tool-use checks, regression evidence, eval reports, or Phase 7 observability handoff for an equipped agent.
Create the MAPS Evaluate phase artifact for an agent or multi-agent system. Use when designing scenario tests, eval datasets, rubrics, scorecards, regression checks, red-team prompts, safety checks, and release evidence.
| name | design-agent |
| description | Create the MAPS Design phase artifact for an agent or multi-agent system. Use when defining roles, workflows, handoffs, memory, context, guardrails, approvals, escalation paths, test-first proof requirements, and system boundaries from an agent brief. |
Current version: 0.5.0.
Follow semantic versioning for this skill:
When changing this skill, update Current version and add a Changelog entry with the date, version, and short summary of behavior changed.
dwyl/learn-tdd as the reference pattern for red/green/refactor, failing-test-first proof, and build-gate evidence.agents/{agent-handle}/agent-profile.md required before design, added profile-limit checks and safe profile sync fields, and required website profile pages to remain mirrors only./design-agent creates the design contract, runtime contract, build backlog, proof plan, and build gate, but must not claim implementation or operational readiness.Use this skill to turn an agent brief into a practical system design.
For role-to-agent work, this skill creates the design contract and build gate only. It does not implement the role as a working agent, does not mark the role as operating, does not grant autonomous runtime, and does not claim production readiness. Full implementation requires /build-agent; operational confidence requires the appropriate Equip, Evaluate, Deploy, Observe, and Improve evidence.
Default to Research and Recommend (R&R): research comparable agents or similar implementations first, recommend a design answer for every major design question, then ask the user to accept or override the recommendations. Do not begin by asking the user a long list of blank-slate questions.
R&R means the agent does the first pass of design judgment.
agents/{agent-handle}/agent-brief.md. If the source is a role contract, read the source role contract first and require a clear /define-agent handoff or an explicit Scott-approved direct role-to-design path.agents/{agent-handle}/agent-profile.md; if it is missing, stop and route back to /define-agent to create the profile before design.G:\My Drive\Mindshare\voice-taxonomy.md when available and recommend a voice profile for the design. If it is unavailable, state the limitation and preserve any voice fields already present in the role contract or agent profile.If research access is unavailable, state that limitation and use known patterns from the brief and MAPS catalogs as the basis for the recommendations.
Use dwyl/learn-tdd as the default TDD reference pattern when designing test-first proof for agent builds.
Reference:
https://github.com/dwyl/learn-tddEvery design must translate test-first proof into a concrete TDD build gate:
/build-agent to show failing-test evidence first, then passing-test evidence, then refactor evidence when refactoring occurs.For agent systems, map TDD proof to behavior, not only code. A valid failing test can be an eval scenario, contract test, approval-gate test, tool-call boundary test, memory-write test, handoff test, stop-condition test, or user-journey test.
At the start of every project run, look for project-foundation.md. If it exists, read Persistent Memory Contract and use its configured notes, sources, memory, RAG, and sync rules as the project defaults. If .maps/foundation-preferences.json exists, use it as the structured preference source for automation.
When this skill creates durable knowledge, write it through the shared MAPS memory helper. The helper gives this skill its own named note under the configured notes root, mirrors that note into the configured RAG location when one exists, appends MAPS Skill Run Log, and records a RAG reindex manifest.
For role-to-agent design, preserve the source role contract as the authority source for mandate, voice, maturity, authority, non-responsibilities, forbidden actions, escalation, learning loop, and memory rules. The agent brief states what part of the role becomes runnable. The agent profile is the runtime control contract for authority, tools, memory, approvals, stop conditions, auditability, activation status, profile gates, and runtime enforcement status. Website profile pages are mirrors only and must not grant authority.
When the proposed design needs more authority than the profile allows, do not expand the profile. Instead, add an approval-required gate to the design and leave the profile authority fields unchanged.
Voice-profile selection may sharpen style, tone, challenge behavior, humor, sentence shape, and example phrasing, but it must not change the profile's authority, tools, memory, approval gates, activation status, runtime enforcement, website mirror boundary, or safe profile sync limits.
/design-agent may update only design-derived profile sync fields that do not expand authority, such as:
not built or another non-activation status.At the end of the run, call the helper after creating the primary output artifact:
python "$CODEX_HOME/skills/foundation/scripts/maps_memory.py" complete-run --project . --skill /design-agent --phase A2 --output "<primary artifact path>" --summary-file "<primary artifact path>" --memory-updates "<notes, sources, memory, or RAG updates>"
If the helper is unavailable, manually append the timestamp, skill, phase, output path, memory updates, and short note to project-foundation.md, then update this skill's named note in <notesRoot>/maps-runs/.
Before writing agents/{agent-handle}/agent-design.md, ask for any missing answers after the Research and Recommend pass. Do not silently choose workflow, memory, tools, approvals, or proof requirements.
Ask exactly one question at a time. Do not present the user with a multi-question form, checklist, or table to fill out. Use the questions below as the internal interview sequence: ask the next most important missing question, wait for the answer, then continue.
Ask:
/define-agent created the agent brief, or has Scott explicitly approved a direct role-to-design path?agents/{agent-handle}/agent-profile.md present, and should the design update it or treat it as read-only input?G:\My Drive\Mindshare\voice-taxonomy.md should be used, or should the design recommend one?/build-agent starts?If the user cannot answer, propose a recommendation and ask them to accept, revise, or mark it unknown.
agents/{agent-handle}/agent-profile.md; if it is missing, stop and route back to /define-agent.G:\My Drive\Mindshare\voice-taxonomy.md when available and recommend a voice profile; preserve existing profile gates if the taxonomy is unavailable./build-agent starts.agents/{agent-handle}/agent-design.md.When the skill is complete, tell the user explicitly. Do not end with only files changed or raw output.
Report:
/design-experience or /design-experience++ when user-facing surfaces matter, otherwise /build-agent.If the skill is blocked, say what answer, artifact, access, approval, or tool is needed before the next skill can run.
Create or update agents/{agent-handle}/agent-design.md using templates/workflow-spec.md from the installed skill when available, otherwise assets/maps/templates/workflow-spec.md from the Mojo repo when working inside this repository.
For role-to-agent designs, also create or update agents/{agent-handle}/agent-backlog.md using assets/maps/templates/agent-backlog.md when implementation slices are identified. Read agents/{agent-handle}/agent-profile.md as a required control contract. Update it only within the authority and mirror rules stated by the source role contract, project foundation, and the safe profile sync field list above.
The completed file contains:
/build-agent/build-agent startsUse assets/maps/templates/workflow-spec.md from the Mojo repo when the installed skill does not provide its own workflow-spec template.
/build-agent starts.