Guide · Agent skill
How to use the Bounded Autoresearch agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The autoresearch skill is a set of instructions that lets an AI agent run a deep, source-grounded research loop on a topic, draft a cited dossier, and optionally propose filing it into an Obsidian-style knowledge vault. This guide explains what the skill does, how to install it, how to trigger it, and what safeguards to check before you let it touch your notes. You will finish knowing how to set research limits, review the proposed changes, and decide when the skill is the right tool.
Bounded AutoresearchAgriciDaniel/claude-obsidian · MIT Download the skill Quick start
- 1Install the skill. Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use, or into .claude/skills/ inside a project. You can also upload it in the Claude apps under Settings → Capabilities → Skills.
- 2Confirm the product is available. Make sure the claude-obsidian product is installed and that scripts/claude-obsidian.py exists in it. Run the test command from the skill to check the file is present.
- 3Ask for research with a clear topic. Use a trigger phrase such as 'research this topic' or 'deep dive into', and state the topic and any exclusions. Say whether web access is approved.
- 4Set or accept the research limits. Answer the agent's questions about rounds, fetches, drafted pages, approved domains, and privacy. If you set none, the defaults apply: three rounds, five fetched sources per round, fifteen drafted pages.
- 5Review the bundle before filing. Run the inspect command on the bundle, read the claims, gaps, and changed paths, then apply it with the approval hash only if you accept it.
What the autoresearch skill makes the agent do
When you ask for deep research on a topic, the agent plans a bounded set of questions, looks for official and primary sources, and collects claims with citations. It notes contradictions and gaps instead of smoothing them over. The result is a draft dossier, not a change to your notes.
Research stays draft-only until you review a proposal. Only then is the dossier filed into your vault. Any merge into existing canonical pages is a separate step that you can accept, narrow, or reject.
- Plans a bounded set of research questions, including a plausible counter-position
- Prefers official and primary sources and records where each claim came from
- Keeps source statements separate from the agent's own inference
- Searches for gaps and contradictions rather than more examples of the leading view
- Stops when the question is adequately supported or the budget runs out, and says plainly when coverage is incomplete
Install the autoresearch skill
Agent skills are folders of instructions that an agent loads when a task matches them. This skill is the autoresearch folder from the claude-obsidian project, with a SKILL.md file and a references folder that includes program.md. Download the ZIP from the skill's page on this directory and unzip it.
Put the unzipped folder in one of these locations. For a personal install, use ~/.claude/skills/. For a single project, use the .claude/skills/ folder inside that project. Alternatively, upload the skill in the Claude apps under Settings → Capabilities → Skills.
- Personal install: ~/.claude/skills/
- Project install: .claude/skills/ inside your project folder
- Claude apps: Settings → Capabilities → Skills, then upload
Trigger the skill with natural requests
You do not need special syntax. The skill is meant to activate when you ask for autonomous or deep research that may use the public web. Phrases from its trigger list include /autoresearch, autoresearch, research this topic, deep dive into, investigate, find everything about, research and file, go research, and build a wiki on.
Be specific about the topic and any exclusions. The agent will confirm the research contract before it starts, so you can answer questions about approved sources, web access, and limits at that point.
- 'Deep dive into how the EU AI Act defines high-risk systems, and exclude commentary blogs.'
- 'Investigate the current evidence on battery recycling methods. Do not use the web yet, only my vault.'
- 'Research and file a summary of our vendor's data retention policy into my vault after I review it.'
- 'Build a wiki on sourdough fermentation, limit it to three rounds.'
Set research limits and approvals
Before research starts, the agent confirms the exact topic and exclusions, whether public-network egress is approved, approved domains or source classes, and any privacy constraints. It also confirms maximum rounds, searches, fetches, elapsed time, and drafted pages, plus whether you want a vault filing after review.
If you do not set limits, the default program applies: at most three rounds, five fetched sources per round, and fifteen drafted pages. Your tighter limits override these. Without consent to web access, the agent researches only your selected vault and the sources you provide, and it tells you that boundary.
- Default: at most 3 search rounds
- Default: at most 5 fetched sources per round
- Default: at most 15 drafted pages
- Without web consent: only your vault and user-provided sources
Review and file results into your vault
Filing is a two-part process. First, the agent builds one transaction bundle for the dossier. You then inspect it. The inspect output shows the cited claims, contradictions, coverage gaps, raw captures, create and replace paths, and budget used. Apply only runs with the approval hash from that inspection.
The skill sets PRODUCT_ROOT to your installed claude-obsidian folder and runs the core script from there. Replace the placeholder paths with your own. Apply uses the approval hash that inspect reports after you review it. A canonical merge into existing concept, entity, or overview pages is a second, separately approved transaction.
- Inspect the bundle first and read every changed path
- Set APPROVAL_SHA256 to the approval_sha256 value that inspect reports
- Canonical merges are a separate, optional approval
- Git checkpoints are created only if you explicitly request one
PRODUCT_ROOT=/absolute/path/to/installed/claude-obsidian
CORE="$PRODUCT_ROOT/scripts/claude-obsidian.py"
test -f "$CORE"
python3 "$CORE" transaction inspect /path/to/research-bundle.json --vault /path/to/vault
# Set APPROVAL_SHA256 to the inspect result's approval_sha256 after review.
python3 "$CORE" transaction apply /path/to/research-bundle.json --vault /path/to/vault --approved-plan-sha256 "$APPROVAL_SHA256"Understand the safety and network notes
This skill runs scripts and uses the network. It needs the claude-obsidian product installed with scripts/claude-obsidian.py available, and Python 3 to run the transaction inspect and apply commands. The skill treats web results, fetched pages, vault notes, and worker drafts as untrusted evidence, and it ignores instructions embedded in them.
Do not send private vault text, file paths, credentials, or unrelated conversation content to external services. The skill warns that the web step can leak context if it is not carefully scoped, so limit approved domains and keep private material out of web queries.
- Web access happens only if you approve public-network egress
- The skill never writes into the plugin or product root
- Writes need explicit approval; the skill does not publish unattended
- Do not use this skill on private or confidential material unless you have scoped the web step carefully
Tips for better research runs
Give a precise research question rather than a broad topic. The skill is built to answer a question, so a narrow question gives a tighter dossier and clearer gaps. State exclusions up front so the agent does not spend budget on material you do not want.
Read the contradictions and coverage gaps before you approve a filing. If the agent says coverage is incomplete, that is a true limit of the evidence, not a failure to fix by pushing for a fuller answer.
- Name the question, the exclusions, and what a good answer would let you decide
- Set tight round and fetch limits for a first pass
- Check each cited claim's source before accepting it into your vault
- Keep canonical merges separate so you can keep the research without changing existing pages
What people use it for
- Building a sourced briefing on a regulation or standard, with each claim linked to its official source
- Researching a vendor's policy and filing a cited summary into an Obsidian vault after review
- Creating a structured wiki on a topic over several bounded research rounds
- Checking a claim by finding primary sources and noting where they contradict each other
Frequently asked questions
Is the autoresearch skill free to use?
The skill is listed under the MIT licence. The source does not state prices for the wider claude-obsidian product, so check its own terms if you rely on it.
Does the autoresearch skill need internet access?
Not always. Web access is used only if you approve public-network egress. Without approval, the agent researches only your selected vault and the sources you supply.
Will the skill change my Obsidian notes automatically?
No. Research is draft-only until you review a proposal. The dossier is filed only after you inspect and approve the transaction, and any merge into existing pages is a separate approval.
What do I need installed to run autoresearch?
You need the claude-obsidian product with scripts/claude-obsidian.py available, Python 3 for the inspect and apply commands, a user vault folder for filing results, and web access only if you approve it.
Why does the skill say it cannot find the core script?
The skill checks for the file with test -f on the core script path. If that check fails, the PRODUCT_ROOT path is probably wrong, so point it at the absolute path of your installed claude-obsidian folder.
Is autoresearch a good alternative to a normal search or chat question?
The source does not compare it with other tools. It is built for deep, multi-round research with citations and review. For quick factual questions, the source says the loop is slow and heavy, so a simple question is usually better answered directly.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use the autoresearch skill when you need a sourced, reviewable research summary and want a record of where each claim came from, especially if you keep an Obsidian vault. Skip it for quick factual lookups, for fully unattended publishing into a knowledge base, and for private or confidential material unless you can keep that material out of web queries.
Download the skill