Menu
Research external information (APIs, libraries, best practices, existing solutions) via web search before planning or implementation.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Baseado na classificação ocupacional SOC
| name | deep-research |
| description | Research external information (APIs, libraries, best practices, existing solutions) via web search before planning or implementation. |
| model | opus |
| effort | medium |
Investigate external information related to the given task.
DR-1. Delegate to web-researcher:
Agent({ subagent_type: "web-researcher", prompt: JSON.stringify({
topic: TOPIC, context: CONTEXT,
artifact_dir: PLANS_DIR
}) })
On failed status: surface summary to user and stop.
DR-2. Read the report from artifact_path (one read, at the end).
DR-3. Present findings — output format: ## Deep Research: PERFORMED|FAILED (1 line) + artifact_path pointer (1 line) + ≤200 char summary. Do not re-emit the full report text in assistant output. The caller must not re-summarize or paraphrase these findings — DR-3 output is the complete user-facing surface.
After completing this skill:
echo "<<WORKFLOW_MARK_STEP_research_complete>>" (must be the ENTIRE Bash command — no pipes, no && chaining, no redirection)Skip this skill when no external knowledge is needed (e.g., the task is purely internal to the codebase).
If research is genuinely not needed for this task:
echo "<<WORKFLOW_RESEARCH_NOT_NEEDED: {reason}>>" (reason must be ≥3 non-space chars, not a placeholder like "none"/"skip", and contain no '>')Conduct a decision-tree interview with the user to lock in requirements, motivation, scope, and non-goals before planning.
Stage 3 of three-stage planning pipeline. Produce a file-level implementation plan via detail-planner/detail-reviewer loop, then get user approval. Inputs are confirmed intent (<session-id>-intent.md) and outline (<session-id>-outline.md) from prior stages.
Edit source code for the current task. Delegates editing and lint/typecheck/self-repair to a subagent.
Plan and write test cases with high reasoning effort. Test iteration runs in a subagent to minimize confirmations.
Explore the codebase to understand existing patterns, constraints, and relevant files before planning.
Investigate git history, docs/history.md, and GitHub issue/PR timeline since the relevant issue opened, to surface changes that may invalidate the issue's premises.