Internal Session Report

Three skills, one audit, 26 wikilinks back in place.

A vault accumulates entropy the way an old house does — quietly. This session built a tooling layer for the maintenance work that was getting deferred, then ran the first audit and repaired the navigation layer of one project subtree.

Date2026-05-27

Surfaceclaude-code-terminal

Duration~75 minutes

Scopevault-wide infrastructure + projects/shur audit

Outcomes at a glance

New skills scaffolded

459

Files audited

174

Real orphans surfaced

Broken MOC wikilinks found

26+

MOC wikilinks repaired

New INDEX hub created

Clusters in value-flows graph

Hygiene reports produced

01The gap

I asked whether an existing Obsidian skill covered three things: organizing the vault, cleaning up connections, and finding insights. The honest answer was that the kit existed in fragments — InfraNodus for insights, vault-ops for note management, mapupdate for the ecosystem graph — but no single workflow walked the vault to surface orphaned notes, broken wikilinks, or stale references.

A 25,000-node Obsidian vault doesn't fail loudly. The link graph silently degrades, Bases queries return partial results, and the cost of finding anything climbs a percentage point per quarter.

The fix wasn't a single Swiss-army tool. The problem decomposes into three orthogonal concerns: a broad audit (find what's broken), a surgical repair (fix what's fixable), and a semantic scan (find what's structurally missing in a project's prose). Three skills, not one.

02The build

Three new slash commands, each with its own SKILL.md, scoped to a different concern. Read-only by default; modification requires an explicit --fix flag plus user confirmation. Output reports land in system/vault-hygiene/ with Bases-queryable frontmatter so they aggregate over time.

Audit

/vault-sweep

Walks the vault to detect orphan notes (zero inbound wikilinks), broken wikilinks (targets that don't resolve), and duplicate titles. Produces a Bases-friendly markdown report. Read-only.

system/vault-hygiene/YYYY-MM-DD-vault-sweep.md

Insight

/vault-insight-scan

Builds an InfraNodus knowledge graph from a folder's prose, surfaces structural gaps, isolated clusters, and gateway concepts. The targeted, project-scoped version of infranodus-expert.

system/vault-hygiene/YYYY-MM-DD-insight-scan-<slug>.md

Repair

/connection-cleanup

Surgical wikilink audit with file:line precision. Categorizes every broken link into auto-fix-case / confirm-typo / stale / dead. With --fix, applies trivial repairs after explicit confirmation.

system/vault-hygiene/YYYY-MM-DD-connection-cleanup-<slug>.md

Composition

The pattern

The three chain naturally. Sweep finds the scope, cleanup repairs the wikilinks, insight-scan tells you what's structurally missing in the prose before you build the next INDEX.

sweep → cleanup --fix → insight-scan → INDEX.md

03The findings

First run scoped to projects/shur/ — 459 markdown files. The raw numbers looked terrifying (748 broken wikilinks, 194 orphans, 41 duplicates). After filtering, three real signals emerged.

The conceptual-markup discovery

The biggest finding wasn't a number — it was a discovery about the vault itself. Wikilinks in this vault serve two distinct purposes: file references (the conventional Obsidian use), and InfraNodus-style entity tagging where [[ReelShort]] is semantic markup, not a file link. Of 2,439 wikilink occurrences in the scope, 984 are conceptual markup. A naive broken-link audit conflates them and produces useless noise.

The skill now pattern-matches source files to disambiguate. Files matching *-ontology.md, vertical-*-ontology/, SESSION-TRANSCRIPT*, or state/brand-profiles/ have their broken wikilinks classified as conceptual-markup rather than file-target.

What's actually broken

MOC
36 broken wikilinks in MOC / INDEX / README files Highest read traffic in the vault. Mostly path-style links like [[../INDEX]] or [[slack-agent/README]] — Obsidian resolves by basename, not relative path.
Brand
Load-bearing concept variants [[ShurAI]] referenced as bare shur (16x), [[SHUR IQ]] as case-variant of ShurIQ (9x), [[ontology]], [[consensus scoring]], [[palantir]] — referenced enough that a canonical reference note would tie the discourse together.
Orphans
10 sibling files under value-flows/ with no INDEX agent-framework, mcp-integrations, memory-architecture, skills-system, flywheel-effect, self-serve-bi, etc. Single highest-leverage organizational fix in the project subtree.
Orphans
28 orphan meeting notes Accumulated across engagements without per-engagement meeting indexes. Recoverable only by date or daily-note cross-reference.
Filter
Python virtualenv pollution .venv/site-packages/*/licenses/LICENSE.md produced 100+ false-positive orphans. Filter additions baked back into the skill: .venv/, venv/, site-packages/, dist-info/, licenses/, __pycache__/, egg-info/.
Convention
engagement_cycle.md ×3 — not a duplicate CLAUDE.md says singleton per engagement, not per vault. All three (SHUR-2026-02 internal, SHUR-HASBRO-2026-04, SHUR-TRUDATA-2026-04) are by-design. No action needed.

04The repair

The MOC layer is where every navigation path starts. 36 broken wikilinks there is the highest-leverage place to spend a half hour. Fixed across four files.

File	Fixes	Pattern
`projects/shur/INDEX.md`	14	Stripped `value-flows/` prefix from 14 wikilinks; cross-project paths converted to markdown links; 6 phantom targets marked TBD
`projects/shur/hasbro/INDEX.md`	5	Path-style → markdown links; case normalizations
`projects/shur/hasbro/competitive-intel/README.md`	4	All `[[../...]]` path-style → markdown links
`projects/shur/content-factory/INDEX.md`	3	Same pattern: path-style → markdown links

The 10 remaining flagged wikilinks are classifier false positives — references to .base files (resolved by the Bases plugin, not standard Obsidian), intentional concept-tag markup ([[infranodus]], [[ShurAI]]), and one path-style link that does resolve in Obsidian. The strict-basename classifier doesn't recognize these cases yet.

05The insight scan

Before building the value-flows INDEX, I ran /vault-insight-scan against the folder to inform the structure. Twenty-two markdown files, 141KB of corpus, fed into a new InfraNodus graph.

22 files → corpus 141KB

InfraNodus → graph vault-shur-value-flows

Analysis → 13 clusters · 150 nodes · 435 edges · modularity 0.63

Gateways → infranodus (0.39) · ShurAI (0.28) · methodology (0.27) · ontology (0.24)

Gaps → 3 structural holes identified

The three gaps

Gap	What it means
Integration ↔ Valuation	InfraNodus/MCP/DEVONthink tooling isn't connected to investor-facing valuation narratives. The technical defensibility story isn't being told.
Brand Intelligence ↔ Knowledge Dynamics	Outputs (stack rankings, flywheel) are structurally isolated from the knowledge-graph processes that produce them. The production chain isn't documented.
Brand Intelligence ↔ Valuation	Traction story isn't connected to investment thesis. Two of the three gaps converge on the valuation cluster — that's the biggest under-bridged area.

The clusters then mapped directly onto the structure of the new value-flows/INDEX.md. Methodology, Integration & Architecture, Brand Intelligence Outputs, Investment & Valuation, Operations & Economics, Future Revenue, Dated Engagement Analyses. The hub explicitly calls out the three gaps so a future reader knows what bridging notes are missing.

A graph-informed INDEX isn't just a file list. It's a diagnosis of what the project's prose isn't yet doing — what conceptual bridges the writer hasn't built.

06Skill tuning, captured

The first run was partly a calibration exercise. Three filter improvements were folded back into vault-sweep/SKILL.md so future runs benefit from today's learnings.

Excl
Python virtualenv exclusions added find + grep now exclude .venv/, venv/, site-packages/, dist-info/, licenses/, __pycache__/, egg-info/
Filt
Noise filters expanded Per-engagement singletons (engagement_cycle.md, prospect-*.md), variant-eval files (under variant-*/, eval/), numbered chapter files (^[0-9]{2}-)
4.5
New Step 4.5: Conceptual-markup classifier When source file matches ontology / transcript / brand-profile patterns, broken-link references get classified as conceptual-markup rather than file-target. Stops the 296-reference noise blast from a single ontology file.

07What's next

Four immediate next moves and a small set of skill-maturity improvements. None are urgent; all are real.

Immediate

Done
MOC layer wikilinks 26+ repaired across 4 INDEX/README files this session.
Done
value-flows/INDEX.md created Hubs the 22 files, organized by InfraNodus cluster structure, with the 3 structural gaps called out.
Done
engagement_cycle.md decision Keep all three — not duplicates, each is a per-engagement singleton per CLAUDE.md.
Done
vault-insight-scan on value-flows Graph live at vault-shur-value-flows. 13 clusters mapped onto INDEX structure.

Soon

Soon
Brand variant canonicalization Targeted /connection-cleanup projects/shur --fix for SHUR IQ → ShurIQ (9 refs) and the bare shur → ShurAI cases (16 refs). Manual review then apply.
Soon
Meeting-note strategy 28 orphan meeting notes need a linking decision: per-engagement INDEXes, flat meetings/INDEX.md, or a Bases dashboard. Punted from this session.
Soon
Sweep other project subtrees Run vault-sweep on projects/microco/, projects/totem-orchestration/, and vault-wide. Compare to today's baseline.

Skill v0.2.0 candidates

v0.2
High-density transcript detection connection-cleanup should suppress files where >50% of wikilinks are dead — almost always conceptual transcripts.
v0.2
.base file resolution Bases dashboard names (.base extension) should resolve as valid wikilink targets, not get flagged as broken.
v0.2
Path-style awareness Path-style wikilinks like [[folder/file]] resolve in Obsidian when the path is valid. Classifier should recognize this pattern.
v0.2
Vault-hygiene cadence Add a recurring schedule entry in system/schedules/personal/ — quarterly sweep, monthly per-active-project insight scan.