examples/reference-agent/skills/soleri-vault-capture/SKILL.md
Use when the user says "save this", "capture this", "remember this pattern", "add to vault", "vault capture", or when persisting learnings from a work session. Validated capture with tier scoping, duplicate detection, and abstraction review.
npx skillsauth add adrozdenko/soleri soleri-vault-captureInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
Capture knowledge to the vault with mandatory validation before persistence. Every item goes through tier classification, duplicate detection, and abstraction review. Nothing lands in the vault without conscious scoping.
Review the conversation or user request. For each piece of knowledge, extract:
For EACH item, apply the three-question test:
| Question | If YES | |----------|--------| | Would any developer on ANY project benefit? | agent tier | | Would this apply to OTHER projects this team owns? | team tier | | Is this specific to THIS codebase only? | project tier |
Decision tree:
Universal? (WCAG, UX laws, language patterns)
→ YES → agent
→ NO → Shared across team projects? (design system, workflow conventions)
→ YES → team
→ NO → project (component rules, token policies, architectural decisions)
When in doubt: prefer project over agent. Too specific is safer than too general.
For each item, search the vault:
YOUR_AGENT_core op:search_intelligent
params: { query: "<item title and key terms>" }
Check results:
For batch scans: YOUR_AGENT_core op:curator_detect_duplicates
Validate each item is at the right level:
| Too Specific (skip) | Just Right (capture) | Too General (skip) | |----------------------|----------------------|--------------------| | "Line 42 of FileUpload.tsx needs a Button" | "Components must use Button atom for all interactive elements" | "Use design system components" | | "Changed bg-red-500 to error token" | "No raw Tailwind colors — use semantic tokens" | "Use semantic colors" | | "Fixed timeout in GuidedCodeBlock" | "setTimeout in useEffect must return clearTimeout cleanup" | "Clean up side effects" |
Ask: "Would this help someone who hasn't read today's conversation?" If NO — too specific, skip it.
Before capturing, present ALL items to the user:
## Capture Review
| # | Title | Type | Tier | Duplicate? | Abstraction | Action |
|---|-------|------|------|------------|-------------|--------|
| 1 | [title] | pattern | project | No match | Good | Capture |
| 2 | [title] | anti-pattern | agent | Similar: [id] (0.45) | Good | Update existing |
| 3 | [title] | workflow | project | No match | Too specific | Skip |
Wait for user approval. Do NOT proceed without explicit confirmation.
For each approved item:
YOUR_AGENT_core op:capture_knowledge
params: {
entries: [{
title: "<title>",
description: "<description>",
type: "<type>",
domain: "<domain>",
severity: "<critical|warning|suggestion>",
tags: ["<tag1>", "<tag2>"],
why: "<reasoning>"
}]
}
For quick single captures: YOUR_AGENT_core op:capture_quick params: { title: "<name>", description: "<details>" }
For each captured entry:
YOUR_AGENT_core op:curator_groom params: { entryId: "<id>" } — normalize tagsYOUR_AGENT_core op:curator_enrich params: { entryId: "<id>" } — LLM enrichmentYOUR_AGENT_core op:curator_contradictions — check for conflicts with existing entriesAfter capture, check EVERY response. Auto-detection frequently assigns "agent" tier with LOW confidence to project-specific knowledge.
YOUR_AGENT_core op:vault_set_scope
params: { id: "<entry-id>", tier: "<correct-tier>" }
If capture returns a proposalId, entry is queued for review:
YOUR_AGENT_core op:governance_proposals
params: { action: "list" }
For cross-project knowledge worth sharing:
YOUR_AGENT_core op:memory_promote_to_global
params: { entryId: "<id>" }
After all captures:
## Captured
| # | Title | ID | Tier | Type | Links |
|---|-------|----|------|------|-------|
| 1 | [title] | [id] | project | pattern | 3 auto-linked |
Skipped: [count] (too specific / duplicate)
Updated: [count] (merged into existing)
Verify with YOUR_AGENT_core op:admin_health and YOUR_AGENT_core op:admin_vault_analytics.
why field (makes entries not actionable)| Op | When to Use |
|----|-------------|
| search_intelligent | Duplicate detection before capture |
| curator_detect_duplicates | Batch duplicate scan |
| capture_knowledge | Persist after approval |
| capture_quick | Fast capture for simple items |
| curator_groom / curator_enrich | Post-capture quality |
| curator_contradictions | Find conflicts with existing entries |
| vault_set_scope | Fix tier after capture |
| governance_proposals | Check governance queue |
| memory_promote_to_global | Share cross-project |
| admin_health / admin_vault_analytics | Verify health after capture |
testing
Triggers: "terse mode", "be brief", "less tokens", "fewer tokens", "compress output", "caveman", or invokes /terse. Token-efficient responses with full technical accuracy.
tools
Triggers: "compress this file", "compress CLAUDE.md", "compress memory", "shrink this", "reduce tokens in file", or invokes /compress. Compresses natural language files to save input tokens.
testing
Triggers: "release", "bump version", "publish packages", "cut a release", "version bump", "npm publish". Bumps monorepo versions, commits, tags, pushes to trigger CI release. Use deliver-and-ship for quality gates.
development
Triggers: "implement X", "build Y", "fix Z", "add feature", or any work task needing planning + execution. Full orchestration loop: plan, execute, complete with vault context and brain recs.