The main agent (you) is the conductor of this procedure. Research, judgment, and synthesis are all done by subagents. You do not read the body of the result files — you only handle paths and verification results.
Input
{topic}: research topic (extracted from the user request){items}: list of research items (items given by the user. If none, decompose the topic into 3–5 items and confirm with the user){results_dir}:.agents/context/deep-research/results(fixed){report_path}:.agents/context/deep-research/report.md(fixed)
Place the artifacts under .agents/context/ for this reason: to follow the convention that this location is the control center for agent artifacts. If the project already has a different convention, you may change {results_dir}/{report_path}to match it — but once decided, you must use it consistently in Steps 0–5.
Step 0 — Preparation
mkdir -p {results_dir}- Resume check:
{results_dir}Skip item result files that already exist in . Research only the remaining items.
Step 1 — Researcher fan-out
For each remaining item, researcher launch one subagent each in parallel.
- If spawning fails because the concurrent execution limit is exceeded, wait for the running agents to finish and then launch the remaining items.
- Use the prompt below as-is, substituting only the variables. Do not modify the structure or wording.
## 조사 항목
{item_name}: {item_description}
## 조사 주제 맥락
{topic}
## 출력 파일
{results_dir}/{item_slug}.md
이 저장소의 코드를 직접 읽어 조사하고, 에이전트 지침의 결과 파일 형식대로 위 경로에 저장하라.
결론마다 근거(파일:줄번호)를 붙이고, 확신이 없으면 [미검증]으로 표시하라.
{item_slug}: convert the item name to lowercase and hyphens (e.g., "state management" → state-management... if in Korean, use item-1, item-2 sequential number)
Step 2 — Mechanical verification (script)
Once all researchers have finished, run the verification script located in the same directory as this SKILL.md :
python3 {이 스킬 디렉토리}/scripts/validate_research.py --results-dir {results_dir} --repo-root .
- The script checks ① whether every conclusion has a source citation ② whether the cited file:line numbers actually exist.
- If any file is FAIL: delete that file and relaunch only the researcher for that item. Append the verification failure message to the rerun prompt. Retry up to 2 times. If it still FAILs, leave that item in the report as "verification failed".
Step 3 — Judgment verification (evaluator)
evaluator Launch the subagent and have it verify with the prompt below. Use it as-is, substituting only the variables. (If the environment has no evaluator agent, skip this step and state "judgment verification omitted" in the final report.)
{results_dir}/ 안의 리서치 결과 파일들을 검증하라.
각 결론이 인용한 근거(파일:줄번호)를 직접 열어 읽고, 결론이 그 근거로 실제 뒷받침되는지 판정하라.
근거가 결론을 뒷받침하지 못하는 항목을 [파일명] [결론] [이유] 형식으로 보고하라.
코드와 결과 파일을 수정하지 말고 판정만 반환하라.
- If the evaluator judges any conclusion as "not supported", rerun that item's researcher once with the judgment reason.
Step 4 — Synthesis (synthesizer)
synthesizer Launch the subagent. Use it as-is, substituting only the variables.
## 주제
{topic}
## 입력 파일
{results_dir}/ 안의 모든 .md 파일
## 출력 파일
{report_path}
입력 파일들을 읽어 에이전트 지침의 리포트 형식대로 합성하라.
원본의 근거 표기와 [미검증] 표시를 유지하라.
Step 5 — Report
Report only the following to the user (do not read and summarize the body of the result file):
- Report path:
{report_path} - Verification result per item (machine verification PASS/FAIL, summary of evaluator judgment)
- If there were retries, how many and which items