guide

Sombra — Reader Mode for AI

Sombra gives you a persistent research library. Web pages saved as clean markdown, organised into collections, distilled into dense context. Everything persists between sessions.

Skill boundaries

This skill covers general Sombra usage — saving, browsing, searching, distilling. For specific workflows, use the dedicated skills:

• Track a project (house purchase, renovation, content strategy) → project
• Track a software feature with git context → spec
• Pick up where you left off → pickup
• Deep API/library research → research

Core workflow

1. Read context before acting

browse_collections → find relevant collection → read_collection_context

The distilled context is the signal. It's dense, preserves code verbatim, and replaces forty pages of raw docs with the five hundred tokens that matter. Always prefer read_collection_context over read_collection unless the user specifically asks for full content.

2. Save as you go

When you find useful pages or work something out during a task:

save_url → move_artifact to relevant collection

Or create a note with your findings:

create_artifact with a summary of what you learned

3. Distil when a collection grows

After adding several items to a collection, offer to distil:

read_collection → write_collection_context

When distilling:

• Preserve all code examples, CLI commands, config snippets, and function signatures verbatim
• Compress everything else to dense prose
• Focus on what an agent needs to write correct code: breaking changes, key API differences, auth flows, gotchas

Use the distill_collection or distill_technical_collection prompts for guided distillation.

When to use Sombra proactively

• Before coding against external APIs or libraries — read the collection context first. It has verified, current docs. Don't rely on training data.
• When the user mentions a collection by name — browse or read that collection.
• When the user asks you to save something — save URLs or create notes.
• When the user says "research" or "look up" — search the library first before going to the web.
• When starting a new task — check if there's a relevant collection. browse_collections is cheap.

Tools reference

Save & Create

• save_url — save a web page as clean markdown
• create_artifact — create a note with markdown content

Read & Browse

• browse_collections — list all collections (cheap, use often)
• read_collection — full content from a collection
• read_collection_context — distilled summary (prefer this)
• fetch_artifact — single artifact by ID
• list_inbox — recent unsorted artifacts
• search_artifacts — search across everything
• list_archived — find deleted items

Update & Organise

• update_artifact — edit metadata or content (notes only)
• move_artifact — move to a collection
• bulk_move_artifacts — move multiple at once
• remove_from_collection — back to unsorted
• create_collection — new collection
• update_collection — edit name, description, icon
• write_collection_context — write/update distilled context
• replace_context_section — surgically update one section of context
• append_to_context — add to the end of context

Delete & Restore

• delete_artifact — archive (recoverable)
• delete_collection — archive (recoverable)
• restore_artifact — recover deleted item

Rules

• Context before code. Always read collection context before generating code against external APIs.
• Don't hallucinate sources. If it's not in the library, say so. Don't invent content.
• Preserve code in distillations. Code examples, CLI commands, and config must be preserved verbatim when writing collection context. Compress prose, never compress code.
• One collection per topic. When creating collections, scope tightly. "Pedestal rate limiting" not "Backend stuff".
• Save the signal, not the noise. When saving URLs, prefer authoritative sources: official docs, well-researched blog posts, primary sources. Skip aggregators and SEO content.