메뉴
Research Harbor and Terminal-Bench 2 patterns to design an original tb-hard task. Use before approving a design brief or when checking whether a candidate is too shallow or derivative.
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
Build or repair a deterministic Harbor task environment with explicit resources, network policy, health checks, and safe solver-visible boundaries.
Implement a legitimate Harbor Oracle solution and validate its behavior against an independent verifier without weakening the task contract.
Red-team a Harbor verifier for answer leakage, reward hacking, shared-container tampering, artifact manipulation, and semantic under-testing.
Calibrate a frozen tb-hard task using Harbor Oracle and real-agent rollouts, then distinguish authentic difficulty from ambiguity, flakiness, leakage, or infrastructure failure.
Create or evolve a canonical Harbor task inside a stable tb-hard case directory. Use after the design brief names observable acceptance criteria and at least three difficulty dimensions.
Audit a calibrated Harbor task for deterministic purchaser export, leakage, evidence completeness, and release readiness.
| name | tb2-pattern-research |
| description | Research Harbor and Terminal-Bench 2 patterns to design an original tb-hard task. Use before approving a design brief or when checking whether a candidate is too shallow or derivative. |
| argument-hint | <case_id or topic> |
| user-invocable | true |
| context | fork |
| agent | tb2-task-designer |
| model | sonnet |
| effort | high |
| allowed-tools | Read Glob Grep Bash |
Research target: $ARGUMENTS
Use official Harbor documentation and the Terminal-Bench 2 repository as a pattern library only. Do not reproduce task wording, tests, solution code, hidden fixtures, or task assets.
For a case ID, write cases/<case_id>/private/pattern-research.md; otherwise return a self-contained memo. Include up to five relevant task patterns, hard elements, expected agent failure modes, originality constraints, asset/provenance risks, and a proceed/reshape/reject recommendation.
Reject a candidate when it is primarily service installation, a known linear recipe, obscured documentation lookup, slow setup, or a static-file exercise. Explain how the proposed case manifests at least three substantive difficulty dimensions in actual solver work.
