How to use the Research Wiki agent skill (research-wiki)

What the research-wiki skill makes the agent do

Most literature searches are used once and then discarded. The research wiki is meant to accumulate knowledge: each paper read, idea tested, experiment run and claim checked is stored as a page that later sessions can query.

The wiki has four entity types: papers, ideas, experiments and claims. Each lives in its own folder (papers/, ideas/, experiments/, claims/). Relationships such as extends, contradicts, inspired_by, tested_by, supports and invalidates are stored in graph/edges.jsonl.

• Initialise a wiki for the current project
• Add a paper, optionally by arXiv ID
• Query what is already known
• Check the wiki for problems (lint)
• Show wiki statistics

Install the research-wiki skill

You can install the skill in one of two ways. Download the ZIP from the skill page and unzip it into ~/.claude/skills/ for all projects, or into a project's .claude/skills/ folder for that project only.

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

• Option A: unzip into ~/.claude/skills/ (available to every project)
• Option B: unzip into the project's .claude/skills/ folder (one project)
• Option C: upload the ZIP under Settings → Capabilities → Skills in the Claude apps

Install the ARIS helper script the skill depends on

The skill does not run on its own. All wiki operations go through the helper script tools/research_wiki.py, which is part of the ARIS research-automation repository. The skill looks for it in .aris/tools/, then tools/, then in the repository path given by ARIS_REPO.

If the helper cannot be found, the skill stops with an error. The error message suggests rerunning tools/install_aris.sh or tools/smart_update.sh from the ARIS repo, exporting ARIS_REPO, or copying research_wiki.py into tools/.

• Python 3 must be available to run the helper
• Helper lookup order: .aris/tools/research_wiki.py, tools/research_wiki.py, then $ARIS_REPO/tools/research_wiki.py
• Without the helper, /research-wiki hard-fails; other skills that update the wiki only warn and skip

Trigger the skill with natural requests

The skill activates when you use phrases such as 知识库, research wiki, add paper, wiki query, 查知识库, or when you ask to build or query a persistent field map. You can also call it directly with a subcommand: init, ingest, sync, query, update, lint or stats.

Plain requests work well. Name the paper or topic and say what you want stored or retrieved.

• "Set up a research wiki for this project"
• "Add paper: <title>, arxiv 2501.12345"
• "Wiki query: what have we tried for this problem?"
• "查知识库" to look up what the wiki already holds

Example workflow

Initialise the wiki once per project. The helper creates the folders, index.md, log.md, gap_map.md, query_pack.md and graph/edges.jsonl, and records the initialisation in log.md.

Ingest papers as you read them. The helper fetches metadata when an arXiv ID is given, creates a page with a generated slug, updates the index and query pack, and appends to the log. Add typed relationships afterwards with the add_edge command.

Safety notes: scripts and network access

The skill is allowed to run shell commands (Bash) and uses web search and fetch tools. It also lists optional Codex MCP tools. Its helper script runs locally, and metadata lookups query the arXiv API when an arXiv ID is given.

Before a note is saved as an idea, claim or experiment, the skill passes it through a capture filter. Notes that describe environment failures, transient errors or claims that a tool cannot do something are rewritten or dropped so they do not become false long-term memory.

• Runs scripts: yes, the helper and capture filter run through Bash
• Uses network: yes, for arXiv metadata and optional web search or fetch
• Review the wiki periodically; stored notes can carry errors that later sessions repeat

Tips for a useful wiki

Ingest papers with a one-line thesis taken from the abstract so later queries have something concrete to match. Keep the wiki per project so it reflects one research question.

Treat the wiki as a draft memory. Check important entries against the original papers before relying on them in writing.

• Give each paper a clear thesis when ingesting
• Record failed ideas too; the wiki is designed to stop you repeating them
• Do not hand-edit the Connections section; it is generated from the graph
• Run lint after large batches of changes