Skip to content

SKILL.md

Latest commit

 

History

History
269 lines (203 loc) · 11.8 KB

SKILL.md

File metadata and controls

269 lines (203 loc) · 11.8 KB
Raw

Anytype Mind

Personal knowledge base powered by Anytype -- an external brain for work notes, decisions, performance tracking, and AI context.

All data lives in Anytype as typed objects with properties and relations. AI agents interact with Anytype via the Anytype MCP Server (@anyproto/anytype-mcp).

Prerequisites

  • Anytype desktop app running (API at localhost:31009)
  • Anytype MCP server configured in your AI tool (see setup/anytype-mcp-config.json)
  • API key created in Anytype Settings > API Keys
  • Default space selected (stored in .space.md -- run bash setup/bootstrap.sh to set)

Default Space

The file .space.md (gitignored) stores the selected Anytype space ID. Read it at session start to know which space_id to pass to every MCP call. If it doesn't exist, prompt the user to run bash setup/bootstrap.sh.

To switch spaces: bash setup/bootstrap.sh --switch-space

Quick Reference

What How
Read an object MCP searchSpace by name, then getObject by ID
Create an object MCP createObject with type_key, name, body, properties
Update an object MCP updateObject with object ID and changed fields
Search MCP searchGlobal or searchSpace with query, types, sort
List by type MCP listObjects with type filter or property filters
Add relation Set an object-format property on the source object

See references/anytype-mcp.md for full tool patterns.

Custom Slash Commands

Defined in .claude/commands/. See references/ for supporting docs.

Command Purpose
/standup Morning kickoff -- load context, review recent changes, surface tasks, priorities
/dump Freeform capture -- dump anything, gets routed to the right Anytype objects
/wrap-up Full session review -- verify objects, relations, suggest improvements
/humanize Voice-calibrated editing -- make object body sound like you, not AI
/weekly Weekly synthesis -- cross-session patterns, North Star alignment, uncaptured wins
/capture-1on1 Capture 1:1 meeting transcript into structured Anytype object
/incident-capture Capture incident from Slack channels/DMs into structured objects
/slack-scan Deep scan Slack channels/DMs for evidence
/peer-scan Deep scan a peer's GitHub PRs for review prep
/review-brief Generate review brief (manager or peer version)
/self-review Write self-assessment for review tool
/review-peer Write peer review
/vault-audit Audit object integrity, relations, orphans, stale context
/project-archive Mark a project as completed, update collections

Data Model

All knowledge is stored as Anytype objects with typed properties. See references/type-schema.md for the full schema.

Object Types

Type Key Purpose Key Properties
work_note Project work, deliverables status, quarter, project_name, related_team
incident Incidents and outages status, quarter, ticket, severity, incident_role
one_on_one 1:1 meeting notes quarter, related_person
decision Decision records status
person People in your org title, related_team
team Teams and groups --
competency Skill definitions --
pr_analysis PR deep scan evidence related_person, review_cycle
review_brief Review period briefs review_cycle
brain_note Operational knowledge --
brag_entry Achievement records quarter
thinking_note Scratchpad drafts --

Properties

All objects support these base properties (built-in or custom):

  • description (text) -- ~150 chars, always filled
  • vault_tags (multi_select) -- classification tags
  • status (select) -- active, completed, archived, proposed, accepted, deprecated
  • quarter (select) -- Q1-2025, Q2-2025, etc.

See references/type-schema.md for the full property list per type.

Collections

Objects are organized into Anytype collections that replace folder-based navigation:

Collection Filter
Active Work type=work_note, status=active
Archived Work type=work_note, status=completed
Incidents type=incident
1:1 Notes type=one_on_one
People type=person
Teams type=team
Brag Entries type=brag_entry
Evidence type=pr_analysis
Competencies type=competency
Brain Notes type=brain_note

See references/collection-views.md for collection patterns.

Session Workflow

Starting a Substantial Session

The SessionStart hook loads context via Anytype MCP: North Star goals, active work, recent changes, open tasks.

Shortcut: Run /standup for a structured morning kickoff.

If doing it manually:

  1. Search for brain_note named "North Star" -- ground suggestions in current goals
  2. Search work_note objects with status=active -- see current projects
  3. Search brain_note named "Memories" -- index of memory topics
  4. Search for action type objects with done=false -- pending items

Ending a Substantial Session

When the user says "wrap up" or similar -- invoke /wrap-up automatically.

If /wrap-up is not invoked, at minimum:

  1. Update completed projects: set status: completed on work_note objects
  2. Verify new objects have relations to at least one other object
  3. Update brain_note objects (Key Decisions, Patterns, Gotchas) with key learnings
  4. Check if wins should be added as brag_entry objects
  5. Offer to update the "North Star" brain_note if goals shifted

Creating Objects

  1. Always set properties: at minimum description (~150 chars) and type-specific required properties per references/type-schema.md.
  2. Use the correct type_key for the object being created.
  3. Write body in markdown following references/markdown-conventions.md.
  4. Add relations immediately -- every object must relate to at least one other object.
  5. Add to collections if not auto-included by type filters.

Object Types and Where They Go

  • Project work? -- create work_note with status=active
  • Completed project? -- work_note with status=completed
  • Incident? -- create incident object (use /incident-capture)
  • 1:1 meeting? -- create one_on_one object (use /capture-1on1)
  • Decision? -- create decision object
  • Person info? -- create/update person object
  • Team info? -- create/update team object
  • Achievement? -- create brag_entry object
  • Codebase knowledge? -- create brain_note object
  • Scratchpad? -- create thinking_note object (promote or delete later)

Relations -- This Is Critical

Every object must have at least one relation. An object with no relations is a bug.

Use object-format properties (related_person, related_team) to create relations between objects. Additionally, mention related objects by name in the body text.

Relation patterns:

  • work_note <-> decision: bidirectional relations
  • work_note -> competency: link to competencies demonstrated
  • work_note -> team: link to team(s) involved
  • work_note -> person: link people involved
  • one_on_one -> person: always link the meeting participant
  • brag_entry -> work_note: every entry links to evidence
  • brain_note -> source: every memory links to where it was learned

Thinking Workflow

Use thinking_note objects for drafts and reasoning before creating final objects.

  1. Create a thinking_note with descriptive name
  2. Reason through the problem in the body
  3. Promote findings to proper typed objects
  4. Delete the thinking_note -- it served its purpose

Memory System

All persistent knowledge lives in Anytype as brain_note objects:

Object Name Purpose
North Star Living goals document, read at session start
Memories Index of memory topics with links to other brain_notes
Key Decisions Architectural and workflow decisions
Patterns Recurring conventions discovered across work
Gotchas Things that have bitten before
Skills Custom commands and workflows

When asked to "remember" something:

  1. Search for the appropriate brain_note by name
  2. Update its body with the new knowledge
  3. Add a relation to the context object
  4. Update the "Memories" brain_note index if a new topic was created

Tags Convention

Use the vault_tags multi_select property:

  • Type tags: work-note, decision, perf, thinking, north-star, competency, person, team, brain
  • Index tags: index, moc

Other metadata uses dedicated properties (status, quarter, review_cycle, etc.), not tags.

Properties for Querying

These properties enable search and collection filtering:

  • review_cycle: h2-2024 -- find all review material for a cycle
  • related_person: <object_id> -- find all evidence related to a person
  • related_team: <object_id> -- find all objects related to a team
  • status: active -- find active projects
  • quarter: Q1-2026 -- find all work for a quarter
  • ticket: TICKET-123 -- find incident by ticket number
  • severity: high -- incident severity
  • incident_role: incident-lead -- your role in an incident

Subagents

Specialized agents in .claude/agents/ for heavy operations.

Agent Purpose Invoked by
brag-spotter Finds uncaptured wins and competency gaps /wrap-up, /weekly
context-loader Loads all context about a person, project, or concept Direct
cross-linker Finds objects missing relations /vault-audit
people-profiler Bulk creates/updates person objects from Slack profiles /incident-capture
review-prep Aggregates all performance evidence for a review period /review-brief
slack-archaeologist Full Slack reconstruction /incident-capture
vault-librarian Object integrity, orphans, stale content /vault-audit
review-fact-checker Verifies claims in review drafts against vault sources /self-review, /review-peer

Hooks (Claude Code only)

Hook When What
SessionStart On startup/resume Load North Star, active work, recent changes, tasks via Anytype MCP
UserPromptSubmit Every message Classifies content and injects routing hints
PostToolUse After MCP writes Validates object properties and relations
PreCompact Before context compaction Backs up session transcript
Stop End of every session Lightweight checklist reminder

Anytype MCP Reference

The Anytype MCP server exposes these endpoint groups:

  • Search: Global and space-specific search with type/property filters
  • Objects: CRUD operations with markdown body support
  • Types: Manage custom object types
  • Properties: Manage typed properties (text, date, select, multi_select, object)
  • Tags: Manage select/multi_select tag values
  • Lists: Manage collections and their views
  • Templates: List and read templates

See references/anytype-mcp.md for detailed tool call patterns.

Rules

  • All data lives in Anytype -- never create standalone .md files for vault content.
  • Every object needs a description property (~150 chars).
  • Every object must have at least one relation to another object.
  • Prefer Anytype search over broad listing when looking for specific objects.
  • When asked to "remember" something, update the relevant brain_note in Anytype.
  • Check for existing objects before creating new ones (search first).
  • Respect Anytype rate limits: burst 60, then 1/sec sustained.
  • The Anytype desktop app must be running for the MCP server to work.

Cross-Platform Support

This project works in:

  • Claude Code: Copy SKILL.md to CLAUDE.md. Hooks load from .claude/settings.json.
  • Codex: Copy SKILL.md to AGENTS.md. Configure Anytype MCP in Codex settings.
  • Cursor: MCP configured in .cursor/mcp.json. Commands load from .claude/commands/.