Clarify Question Loop

June 26, 2026 · View on GitHub

The clarify question loop turns a rough request into enough information to generate or package a useful Agentlas repo.

This is a public contract. Hosted web products may store sessions, meter usage, or run background jobs. Local runtimes may ask inline. The question strategy should stay portable.

When To Ask

Ask clarification when the request is missing details that change the package shape:

  • mode is ambiguous;
  • single-agent vs team selection would change the package shape and the independent ownership boundaries are unclear;
  • target runtime is unclear;
  • public vs private release boundary is unclear;
  • tools, APIs, credentials, or data sources are unclear;
  • the user wants an existing repo packaged but has not provided its path or contents;
  • the output might require destructive actions, paid APIs, public publishing, or sensitive data handling.

Do not ask when the missing detail can be safely defaulted and repaired later.

Question Budget

Ask one to five questions. Prefer three.

Each question should:

  • be answerable in one sentence or one choice;
  • change the generated files or safety boundary;
  • avoid asking for secrets;
  • avoid asking the user to choose internal implementation details;
  • include a default when there is a safe default.

Shared Questions

Use these for all modes when relevant:

  1. Target runtime: Codex, Claude Code, Gemini CLI, Cursor, generic AGENTS.md, or multiple?
  2. Public boundary: local-only, private team, public open-source, or marketplace?
  3. Data and tools: what files, APIs, websites, CLIs, or services should the agent use?
  4. Success proof: what should count as a correct output?
  5. Safety boundary: what must the agent never read, write, publish, or spend?

When single-agent vs team shape is unclear, ask this in the first batch before generation:

  1. 이 일을 한 명의 전문가가 처음부터 끝까지 맡으면 되나요, 아니면 조사/분석/검토처럼 여러 전문가가 나눠 맡고 마지막에 합쳐야 하나요?

Then follow up only as needed:

  • 나눠야 한다면 어떤 역할들이 필요한가요? 예: 조사, 분석, 검토, 승인.
  • 각 역할이 쓰는 파일, 계정, 웹사이트, 도구가 서로 다른가요?
  • 마지막에 누군가 결과를 하나로 모아야 하나요?
  • 앞 사람이 끝낸 결과를 다음 사람이 이어받아야 하나요, 아니면 동시에 해도 되나요?

Plain-Language Question Rule

Questions shown to users must be understandable by a non-technical operator. Do not ask the user to choose internal labels such as single-agent, team-builder, ownership boundary, memory/context, synthesis, or produces/consumes.

Translate internal terms before asking:

  • ownership boundary -> who needs a separate role;
  • memory/context -> the notes, materials, criteria, and progress each role must remember separately;
  • tools/permissions -> the accounts, files, websites, or tools each role may use;
  • synthesis -> putting several outputs into one final result;
  • sequential dependency -> one person handing finished work to the next person.

Mode-Specific Questions

Single Agent Creator

  • What is the one job this worker owns?
  • Who will use it?
  • Should it include memory, scheduled refresh, or self-evolution proposals?
  • What setup steps should be beginner-friendly?

Team Builder

  • What departments or roles are required?
  • Who is the final orchestrator/HQ?
  • What needs PM Soul continuity?
  • What gates are required before release: policy, eval, QA, approval, or human review?

Agentlas Packager

  • Where is the existing prompt, agent, team, repo, folder, or ZIP?
  • Should the package preserve current behavior exactly, repair weak structure, or prepare a public release?
  • Which adapters are required: Codex, Claude, Gemini, Cursor, generic AGENTS.md, or all?
  • What private files, logs, secrets, or local-only notes must be excluded?

Answer Synthesis

After answers arrive:

  1. Re-run mode classification if the answers changed the mode.
  2. Generate or repair the smallest useful package.
  3. Include public contracts: AGENTS.md, .agentlas/, mode files, skills, runtime adapters, memory map, and verification.
  4. Preserve user-provided constraints exactly.
  5. Record unresolved assumptions in the output.

Output

Return:

  • selected mode;
  • answers used;
  • assumptions;
  • generated or repaired files;
  • verification command;
  • blockers, only when user input or external state is required.