How to Build an Opus 5 + Obsidian Research System That Replaces Hours of Manual Reading Every Week

Source: "How to Build an Opus 5 + Obsidian Research System That Replaces Hours of Manual Reading Every Week" — CyrilXBT, Jul 29, 2026. Adapted synthesis by Hermes (DeepSeek V4 Flash via Zen) following the article's own protocol: key ideas, not a copy.

The core idea

Most research work is not thinking: it is reading, extracting, and cross-referencing by hand. This system automates that mechanical layer: Obsidian as the permanent store, an efficient model as the processing engine, and a pipeline that turns raw material into linked, queryable knowledge. The hours spent re-reading are returned to actual thinking.

Model choice

The article picks Opus 5 (same price as Opus 4.8, more than double the benchmark score, ~26% fewer tokens per Harvey) over Fable 5 (deeper, ~2x price): a weekly pipeline is dozens of repeated passes — volume work, not one-off depth. Fable 5 is reserved for genuinely hard questions.

Adaptation: my Opus 5 equivalent is DeepSeek V4 Flash (via Zen): cheap and fast for the repeated ingestion and synthesis passes. Same rule: efficient model by default, deep model only when the question justifies it.

Vault structure

• me/raw — sources exactly as ingested, never edited
• me/wiki — processed, linked knowledge organized by topic
• questions/ and digests/ — open questions and weekly summaries (to be created)
• The original's protocol file (CLAUDE.md) is, here, the obsidian-ingest skill of Hermes, encoding the same rules

Ingestion protocol (adapted)

1. Extract only genuinely new claims, not a summary of the whole document
2. Check me/wiki before creating: extend an existing note rather than duplicating
3. Link every claim with wikilinks
4. Contradictions are never silently overwritten — flag them in both notes
5. Open questions go to questions/, never guess

Weekly digest

~500 words, only what was added or updated in the last 7 days: 3-5 key findings with links, contradictions, open questions, and unusually active topics. Scheduled (Hermes scheduled tasks): Monday starts with a 2-minute digest instead of 2 hours of re-reading.

Querying and contradictions

Ask the whole wiki with citations, not memory; value compounds over time. Contradictions force a human decision (outdated source, changed reality, or context to add) instead of resolving silently.

Common mistakes

Skipping the protocol, summarizing the whole vault instead of the week, treating the synthesis as automatic truth, and using the expensive model out of habit.

Technical adaptation

The original connects Obsidian to Claude via the Local REST API plugin + MCP. Here the equivalent already exists: the obsidian-ingest skill + the vault HTTP API (port 27123, Bearer key) + the public UI at https://obsidian-http.test.legido.com. Flow: "paste me a URL" → raw to me/raw, linked synthesis to me/wiki, index updated.

Next steps

Scheduled weekly digest, questions/ and digests/ folders, cross-vault queries. See Research System.