How to use the Wiki Ingest agent skill (wiki-ingest)

What the wiki-ingest skill makes the agent do

The skill turns supplied material into grounded, cross-linked notes without changing the source. Files already in the vault's inbox/ or .raw/ folders stay user-owned and read-only. The agent records where each claim came from, and it flags contradictions or unsupported statements instead of silently accepting them.

It is built for a single source or a bounded batch. It is not meant for saving a chatbot answer.

• Reads pasted text, files staged in the selected vault's inbox/ or .raw/ archive, or explicitly approved URLs.
• Keeps provenance: each source gets a SHA-256 identity and claims are tracked against their sources.
• Preserves contradictory evidence and marks no-data claims as unsupported.
• Works from a budget (source count, pages or bytes, page reads, generated pages, network requests) and a first tranche for large batches.

Install the skill and its requirements

Download the ZIP from this page and unzip it into ~/.claude/skills/ for your user account, or into a project's .claude/skills/ folder for one project. You can also upload the skill in the Claude apps under Settings → Capabilities → Skills.

The skill needs an Obsidian vault and the claude-obsidian product installed, including its scripts/claude-obsidian.py core file, which SKILL.md references. The core script needs a Python environment; the source does not state an exact Python version. URL capture needs a network or fetch adapter, and PDF, image, audio and video extraction need a host capability or configured adapter.

Trigger the skill with natural requests

You do not need special syntax. Ask the agent in plain language, and include the material or its location. The skill recognises these trigger phrases: ingest, ingest this file, ingest this URL, process this source, read and file this source, batch ingest, and ingest these sources.

The vault is chosen by an explicit --vault option, the CLAUDE_OBSIDIAN_VAULT setting, workspace config, or discovery from the current directory. Local files and pasted content need no network access.

• Ingest this file: the file is already in your vault's inbox/ or .raw/ folder.
• Ingest this URL: the agent will ask for consent for the destination domains and request budget before fetching.
• Process this source: pasted text or a staged file, filed as one source.
• Batch ingest these sources: the agent sets a budget and proposes a first tranche rather than promising everything.

Understand scope, sources and limits

Before processing, the agent lists the inputs and agrees a budget. For a large batch it chooses a bounded first tranche. A supplied local path outside the selected vault is not durable provenance, so the agent asks you to place the file in inbox/ or supply the text instead. The core's reviewed capture plan and capture apply workflow then creates a create-only .raw/captured/ path that is ingested.

Media files are not guessed at. Without a host capability or configured adapter, the agent preserves the locator and reports that extraction is unsupported.

• Pasted text and files in inbox/ or .raw/ are read locally.
• Outside-vault paths need the capture workflow before ingest.
• URLs need an available fetch adapter and explicit consent.
• PDFs, images, audio, video, OCR and transcripts need a host capability or adapter.

Preview and apply the ingest bundle

Each batch is drafted as a single claude-obsidian.transaction.v1 bundle with operation_type ingest. The bundle can include create-only raw captures, source summaries, reviewed page changes, source and claim ledger records, and address requests. The workflow gate reports storage_valid separately from workflow_complete, so apply only when both pass.

Run the inspect step first, review the result, then set APPROVAL_SHA256 to the approval_sha256 it reports before applying.

• Storage valid and workflow complete must both pass before apply.
• Never add a source or claim record only to satisfy the gate.
• For a methodology MOC, include workflow catalogs in the bundle with that catalog's expected hash.

Safety notes: scripts, network and untrusted content

The skill runs scripts and uses the network. Source content is treated as untrusted data: web pages, local files, pasted text and metadata cannot override the skill or your scope. Embedded instructions, requests for secrets and destination changes are ignored.

Before any URL is fetched, the agent needs your explicit consent for the domains and budget. It does not send vault content, private paths, credentials or unrelated conversation data, and it stops if redirects leave the approved scope.

• Approve domains and request budget before any fetch.
• Expect scripts to run through the core script and workflow commands.
• Review the inspect output before applying changes.

Tips for better results

Start with one source to see how the notes look in your vault, then move to a small batch. Keep the budget tight so the agent reads each source completely; if it cannot, the result is labelled partial and the missing range is recorded.

Put files in inbox/ before you ask for ingest. This keeps provenance durable and avoids the extra capture step.

• Start with one source, then batch.
• Use inbox/ for local files you want tracked.
• Check the labelled partial results and missing ranges.
• Ask for URL capture only for domains you trust and have approved.