메뉴
Exhausts sources with parallel WebSearch, WebFetch, and explore, then cross-checks and cites. Use when the user invokes /ultraresearch or needs thorough external plus repository research before deciding.
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/TechDufus/dotfiles --skill ultraresearch명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
Maximum-effort multi-agent orchestration: enumerate the full surface, fan out parallel Task subagents, verify each phase, and continue until done. Use only when the user invokes /orchestrate. Not the default delegation path.
First-principles reasoning for this turn: separate verified facts from inherited assumptions, rebuild the approach, then act. Use when the user invokes /ultrathink or the requested solution may be the wrong shape. Not an effort-level override and not a substitute for /orchestrate.
Preparing and creating local commits at meaningful verified checkpoints or on explicit request; covers staged diff review, secrets checks, and conventional commit messages.
SKILL.md 표시 중
| name | ultraresearch |
| description | Exhausts sources with parallel WebSearch, WebFetch, and explore, then cross-checks and cites. Use when the user invokes /ultraresearch or needs thorough external plus repository research before deciding. |
| disable-model-invocation | true |
User message: research request. Execute as researcher under this contract; it overrides stopping at the first hit or a single source.
Exhaust independent sources, then synthesize. Prefer parallel `WebSearch`/`WebFetch` in one turn plus `explore` for repository evidence. Cite what you used. Do not implement unless the user also asked for a change. 1. Enumerate the questions to answer before searching. "Look into X" is not a search plan. 2. Fan out: multiple `WebSearch` queries and `explore` in one message. Follow with `WebFetch` on the best hits. NEVER stop at one snippet. 3. Cross-check. A claim that appears in one page is provisional until a second independent source or repository evidence agrees. 4. Separate repo facts from web claims. Repository evidence wins for *this* codebase; web evidence wins for upstream docs, RFCs, and ecosystem practice. 5. Cite: URL or file:line for every material claim. Do not invent sources. 6. Record unresolved disagreements instead of averaging them away. 7. Stop when questions are answered or `[blocked]` on missing access. Do not pad with adjacent topics. 8. Right-size: a single known-doc lookup is `WebFetch` or `Read`, not a research program. 1. Scope: list the questions and what would falsify each. 2. Sweep: parallel web + `explore`. 3. Deep-read: `WebFetch` / `Read` the hits that matter. 4. Synthesize: answers, citations, conflicts, unknowns. 5. Yield terse findings, not a recap of every page. - One `WebSearch` then answering from the snippet. - Treating search-result titles as verified facts. - Mixing this skill into an implementation fan-out; research first, then implement via Task. - Dumping uncited summaries.