How to use the researchwrite agent skill for research proposals

What the researchwrite skill makes the agent do

Without the skill, a request like 'write my proposal' usually gets a single block of text. The skill changes the order of work. The agent first asks for or builds a research canon (hard facts and constraints) and an evidence table that maps each claim to its source. It then writes an argument map, and only after that drafts each section against a contract listing allowed and forbidden claims.

After drafting, the skill runs a multi-stage review: content experts first, then language cleanup, then format and citation checks, then scoring against a threshold. It also supports partial-proposal scoping, so you can write one stage at a time without the scope growing, and exporting to .md and .docx.

• Modes: compose (start from a topic or rough idea), revise (improve an existing section), hybrid (expand an existing draft).
• Output folder: outputs/researchwrite/<project-slug>/ with files such as 01_research_canon.md, 02_evidence_table.md, 03_argument_map.md and exports/.
• The skill is designed not to upgrade facts: it should not change 'may indicate' into 'proves' without evidence.

Install the researchwrite skill from the ZIP

Download the ZIP for the skill from this page. Then choose one of the two install routes below.

For a personal install that works in every project, unzip the download into ~/.claude/skills/. For a single project, unzip it into that project's .claude/skills/ folder. Both routes use the same SKILL.md file.

In the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.

• Personal: ~/.claude/skills/
• Single project: .claude/skills/ inside the project folder
• Claude apps: Settings → Capabilities → Skills, then upload

Check that the agent loaded the skill

Start a new conversation after installing, then ask for a task that matches the skill's description, such as drafting a 开题报告 from your materials. If the agent follows the evidence-first steps (asking for sources, building a table before prose), the skill is active.

If it behaves like a normal chat assistant, the skill probably was not found. Check that the folder containing SKILL.md sits directly inside the skills folder you chose, not nested one level deeper by mistake. The source does not describe a dedicated check command, so do not expect one.

Trigger the skill with natural requests

You do not need special syntax. Name the skill, or describe a proposal task in plain language. The skill's own examples include a request to review a discussion section in paper mode, and requests in Chinese that split a grant topic into scientific questions and research content without writing the full text.

Be explicit about what you want. The skill asks follow-up questions only when the answer changes the workflow, so state the mode, the stage, and the materials you have.

• Use researchwrite to review this discussion section, paper level.
• 我有一个基金题目，帮我先拆科学问题和研究内容，不要直接写正文。
• This introduction is scattered. Restructure it using the proposal-first approach.
• Turn this draft into a proposal framework and mark the evidence gaps.
• Quickly scan this email (uses the quick level, which only flags P0 issues).

Prepare what the skill needs

The skill depends on your materials. It does not search the literature or confirm facts. Give it your sources, data, and notes, and state any facts or wording that must not change. If evidence is missing, the skill is designed to flag it for you rather than write around it.

It needs an agent runtime that loads SKILL.md skills. Python or similar is only needed for an export script that the manifest refers to, and that script is not included in the files reviewed.

• Source materials: literature, data, notes, figures, references you have verified.
• Target length, language and output format.
• An agent runtime that loads SKILL.md skills.

Safety notes: scripts, network and verification

According to the review, this skill does not run scripts and does not use the network. The export step and some reference files are mentioned in the instructions, but the scripts and templates they point to were not among the reviewed files, so treat those features as unverified until you have them.

The license needs a check. The listing states Apache-2.0, while the SKILL.md frontmatter says MIT. Read the license file in the source repository before you rely on either.

Most important: verify every citation and claim yourself before submitting to a funder or journal. The skill works from your evidence and cannot confirm facts on its own.

• Runs scripts: no (per review).
• Uses network: no (per review).
• Export script referenced in manifest.yaml: not included in reviewed files.

Tips for better results

Start with the evidence, not the prose. Give the agent your papers and notes first, because the skill is built to build the table before the text. Keep each run to one stage, for example one section or one review pass, to avoid scope creep.

Use the review levels. Choose the proposal level for a 开题报告 or research plan, and the internal or quick levels for less formal material. Expect the review to stop when it hits a plateau, an expert conflict or missing evidence, because the skill treats those as reasons to stop.

• Give the skill your own sources and say which facts are fixed.
• Ask for a revision brief if you plan to hand the work to another agent.
• Remember that the skill deletes claims it cannot support rather than explaining them away.