Distill

Skill marsmike/agentic-toolkit/plugins/obsidian/skills/distill

Vault-first toolkit for Claude Code — curated plugins over an Obsidian-compatible knowledge vault, governed by an explicit contract, gated by evals, continuously delivered.

Install

npx -y skills add marsmike/agentic-toolkit --skill distill

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

10 days oldThe repository was created 10 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Process captures into vault knowledge — triage the inbox, distill one file, or file a conversation insight. Use when working with 01_Capture/.

SKILL.md

3.2 KB, as published. Nobody here has run it

Knowledge Distillation

Transforms raw captures from 01_Capture/ into integrated, linked knowledge notes. Filesystem-first throughout — no CLI or embeddings store is required to run this workflow; vault/CLAUDE.md's "Distilling captures" section states the two non-negotiable rules (two-phase checkpoint, source preservation) that hold even without this skill loaded. Read rules.md before starting.

Modes

Mode	Use for	Reference
Distill (default)	Full pipeline: capture → search → PARA placement → enrichment → write	workflow.md
Triage	List and prioritize the inbox before distilling	workflow.md — Triage section
Insight	File a conversation synthesis as a new capture, then distill it	workflow.md — Insight section

Hard requirements (non-negotiable, see rules.md for the full list)

Two phases, one checkpoint. Analyze and propose, then stop for review before writing anything. Skip only on an explicit --auto/non-interactive instruction.
Search before writing. Run scripts/search.py against the capture's key terms — it degrades gracefully with no embeddings store present, but it must run; a distill pass with zero search is a name for "guessing at what already exists." When a gaiafield binary is available, scripts/graph.py's graph_context() also adds graph-derived backlink/bridge candidates the text search alone missed (see workflow.md); its absence never blocks this step. When the binary supports gaiafield v2, graph.inferred_candidates() adds a separately-labeled, report-only block of statistical candidates (workflow.md's "Inferred candidates" section) — these are never auto-applied, ever; see contract/KNOWLEDGE_API.md's v2 section, rule 1.
Every distilled note carries its source — a *Source: ...* line in the body and a source: frontmatter field — and gets status: distilled.
The capture leaves 01_Capture/ after distilling — archived (default, to 05_Archive/<Origin>-Captures-<YYYY-MM>/) or deleted (duplicates/empty stubs only).
Ambiguity goes to the DLQ, not a guess. If placement, source, or search results are genuinely unclear after one honest attempt, write a dead-letter note to 00_Memory/dlq/ via scripts/vault_utils.write_dlq_note() and say so in the Phase 1 report, rather than silently picking an answer.

Quick start

uv run --project "$CLAUDE_PLUGIN_ROOT/scripts" python3 "$CLAUDE_PLUGIN_ROOT/scripts/search.py" "key terms from the capture" --top 10 --json

Then follow workflow.md's numbered steps.

References

Workflow — the full distill/triage/insight procedure
Rules — PARA placement logic, enrichment levels, the DLQ convention, cluster mode