| name | research_plan_composer |
| description | Compose a natural-language research plan from the old content diagnosis, opportunity hypothesis, and candidate probe skills. |
| stage | planning |
Skill: research_plan_composer
Mission
Write the Agent Research Plan.
The plan must sound like the agent reasoned from this specific old content. It must not look like a checklist of selected tools.
Product Decision It Supports
This skill decides what the agent will investigate before generating a task card or no-action report.
Inputs
- old content diagnosis
- comment signal diagnosis
- opportunity framing
- candidate probe skill instructions
Reasoning Moves
- Explain how the agent interprets the old content.
- Explain the ReCast opportunity hypothesis.
- Select the research moves that can confirm, weaken, or redirect the hypothesis.
- Reject irrelevant paths with concrete reasons.
- Generate search intents and queries only for research moves that need external evidence.
Rules
- Frontend users should see the natural-language plan, not a skill menu.
- Backend state must retain internal skill references and rejected skills.
- A rejected skill is useful only if the reason is specific to the old content.
- Do not load every possible probe into the visible plan.
- Do not say the old conclusion is wrong before evidence and testing.
- Research may use a wider time window to understand context, but change/update/drift judgement must use the old post's
posted_at as the baseline.
- For each research move, distinguish whether it is collecting background context, audience signal, competitive context, availability context, or post-publication change evidence.
- If a research move is meant to support a change trigger, its
change_baseline_date must be the old post's posted_at.
- A useful topic direction is not automatically a ReCast signal. The plan should test whether the old content itself has reusable value.
- Do not introduce specific model/version names into search queries unless the old post explicitly named them or earlier evidence has established them. If the old post only says "ChatGPT", search for ChatGPT, not an assumed model such as GPT-4o.
Output Contribution
Contribute:
- plan narrative
- research steps
- internal skill references
- rejected skills with reasons
- expected output type
