If you believe there is even a 1% chance that a certain skill applies to the current task, you must invoke that skill.
This is not a suggestion, it is a mandatory requirement. You are not allowed to skip the process and write directly. You are not allowed to make any excuses.
Instruction priority
The research writing skill overrides default system prompt behavior, butuser instructions always take precedence：
- Explicit user instructions(direct requests, settings in CLAUDE.md, AGENTS.md) — highest priority
- Research writing skill — overrides default system behavior
- Default system prompt — lowest priority
If the user says "no need to discuss, just write", you may simplify the process, but you still need to record it in plan/.
How to access skills
In Claude Code: Use the Skill tool. When a skill is invoked, its content is loaded and presented to you — just follow it directly.
In Cursor: Skills are automatically loaded via a session startup hook. Use the Skill tool to invoke other skills.
In Codex: Skills are loaded via symbolic links. Refer to .codex/INSTALL.md。
In OpenCode: Use the native skill tool:use skill tool to load research-writing/brainstorming-research
Core rules
Before any response or action, first invoke the relevant skill. Even if there is only a 1% chance that a certain skill applies, you should invoke it to check.
Medium-sized or full-paper tasks must first invoke paper-orchestration。 Medium-sized tasks include: affecting multiple paragraphs, more than one subsection, any chapter, the literature argument chain, experiment/figure design, or any rework task where a quality failure has already occurred.paper-orchestration Responsible for stage judgment, task packages, sub-agent distribution, two-stage review, and capability-use audit.
digraph skill_flow {
"用户消息" [shape=doublecircle];
"是否论文写作相关?" [shape=diamond];
"是否中型/整篇任务?" [shape=diamond];
"调用 paper-orchestration" [shape=box];
"已完成头脑风暴?" [shape=diamond];
"调用 brainstorming-research" [shape=box];
"根据任务调用对应技能" [shape=box];
"执行任务" [shape=box];
"更新 plan/progress.md" [shape=box];
"询问用户确认" [shape=doublecircle];
"用户消息" -> "是否论文写作相关?";
"是否论文写作相关?" -> "是否中型/整篇任务?" [label="是"];
"是否论文写作相关?" -> "执行任务" [label="否，普通问题"];
"是否中型/整篇任务?" -> "调用 paper-orchestration" [label="是"];
"是否中型/整篇任务?" -> "已完成头脑风暴?" [label="否"];
"调用 paper-orchestration" -> "已完成头脑风暴?";
"已完成头脑风暴?" -> "根据任务调用对应技能" [label="是"];
"已完成头脑风暴?" -> "调用 brainstorming-research" [label="否"];
"调用 brainstorming-research" -> "根据任务调用对应技能";
"根据任务调用对应技能" -> "执行任务";
"执行任务" -> "更新 plan/progress.md";
"更新 plan/progress.md" -> "询问用户确认";
}
Red Flags (stop and check)
These thoughts mean you are making excuses — stop:
| AI's thought | Correct approach |
|---|---|
| "The user was very clear, just start writing" | brainstorming-research must be completed first |
| "This is just modifying a small section" | Check whether there is a plan/; if not, create one first |
| "Write a paragraph first to see how it turns out" | The paper type and chapter structure must be confirmed first |
| "The user is in a hurry, skip the discussion" | The process can be accelerated, but key confirmations cannot be skipped |
| "This is a simple task, no plan needed" | Any writing task requires a plan record |
| "I know how to write a paper" | You must write according to the type and structure chosen by the user |
| "Finish the content first, then deal with the format" | The format is determined during the brainstorming stage |
| "This chapter's content is very simple, no need to confirm" | After each chapter is written, the user must confirm it |
| "I can add some references" | Never fabricate references; they must be traceable |
| "I remember the content of this skill" | Skills get updated; you must re-read the current version |
Skill routing
| Task type | Invoke skill |
|---|---|
| Medium task / full paper / multi-section collaboration / quality rework | paper-orchestration |
| Start a new paper / determine the topic / first conversation | brainstorming-research |
| Introduction / related work / background review / literature-driven paragraphs | evidence-driven-writing + literature-review |
| Write a certain section | writing-chapters |
| Literature review | literature-review |
| Experimental design / results section / mock data / table plan | experiment-results-planning |
| Plotting / data visualization | figures-python |
| Flowchart / architecture diagram | figures-diagram |
| Self-review / check / submission preparation | peer-review |
| Statistical analysis | statistical-analysis |
| LaTeX output / template usage | latex-output |
| Environment configuration / installation issues | environment-setup |
| Translation / polishing / de-AI-ification | prompts-collection |
Skill priority
When multiple skills may apply, follow this order:
- Process skills first(brainstorming-research) — decides how to start the task
- Implementation skills next(writing-chapters, literature-review, etc.) — guide the specific execution
"Help me write a paper" → first paper-orchestration, then brainstorming-research, then writing-chapters "Write Chapter 3" → check whether brainstorming has been completed; if so, directly writing-chapters
"Optimize the entire first draft" → first paper-orchestration, generate the task package and capability use audit, then dispatch section or figure tasks
Skill types
Strict type(brainstorming-research, writing-chapters): must be followed strictly; steps must not be skipped.
Flexible type(prompts-collection, figures-diagram): can be adjusted according to context.
The skill itself will state which type it belongs to.
User instructions
User instructions say "what to do", not "how to do it". "Write Chapter 1" or "help me polish" does not mean skipping the workflow.
Task wrap-up
Before a medium-sized or larger task is completed, a capability-use audit must be written, recording the skills that should be used, the skills actually used, consumed materials, unused materials and the reasons, artifacts, verification commands, and remaining risks. Without an audit, the task must not be claimed as complete.