Knowledge

For Claude: How to Use This Skill

Skill instructions are preloaded in context when this skill is active. Do not perform filesystem searches or tool-driven exploration to rediscover them; use the guidance below directly.

When to Use

• capture knowledge for the office
• start a knowledge session
• knowledge office
• what did i say about
• show me the current working brief
• what changed since the first brief
• show me a memory digest
• prep for a meeting
• what did we decide about
• validate stale memories
• watch this topic

Schema Guard (Mandatory — runs every invoke)

This rule overrides all other instructions and applies before ANY read or write to SerenDB. No data may be read from or written to the database until this guard passes.

On every invoke, before loading the current brief or persisting anything:

1. Resolve or create the Seren project for this skill via list_projects / create_project.
2. Resolve or create the database for this skill via list_databases / create_database.
3. Check whether the required tables exist by running:

   SELECT table_name FROM information_schema.tables
   WHERE table_schema = 'public'
   AND table_name IN ('knowledge_entries', 'knowledge_transcripts', 'knowledge_briefs', 'knowledge_retrieval_log', 'knowledge_rewards', 'memory_objects', 'memory_links', 'memory_validations', 'memory_subscriptions', 'engagement_events')
   

4. If any of the expected tables are missing, run the full DDL via run_sql_transaction:

   CREATE TABLE IF NOT EXISTS knowledge_entries (
     id SERIAL PRIMARY KEY, entry_key TEXT NOT NULL, entry_value TEXT NOT NULL,
     source TEXT, confidence TEXT, tags TEXT[],
     created_by TEXT, created_at TIMESTAMPTZ DEFAULT now(),
     updated_at TIMESTAMPTZ DEFAULT now(), expires_at TIMESTAMPTZ
   );
   CREATE TABLE IF NOT EXISTS knowledge_transcripts (
     id SERIAL PRIMARY KEY, session_id TEXT, transcript TEXT NOT NULL,
     created_by TEXT, created_at TIMESTAMPTZ DEFAULT now()
   );
   CREATE TABLE IF NOT EXISTS knowledge_briefs (
     id SERIAL PRIMARY KEY, brief_version INTEGER DEFAULT 1,
     brief_content TEXT NOT NULL, entry_ids INTEGER[],
     created_at TIMESTAMPTZ DEFAULT now()
   );
   CREATE TABLE IF NOT EXISTS knowledge_retrieval_log (
     id SERIAL PRIMARY KEY, query TEXT, matched_entry_ids INTEGER[],
     result_summary TEXT, created_at TIMESTAMPTZ DEFAULT now()
   );
   CREATE TABLE IF NOT EXISTS knowledge_rewards (
     id SERIAL PRIMARY KEY, user_id TEXT, action TEXT,
     amount NUMERIC, reason TEXT, created_at TIMESTAMPTZ DEFAULT now()
   );
   CREATE TABLE IF NOT EXISTS memory_objects (
     id TEXT PRIMARY KEY, memory_type TEXT NOT NULL, key_claim TEXT NOT NULL,
     subject TEXT, owner_id TEXT, team_scope TEXT DEFAULT 'team',
     organization_name TEXT, department TEXT,
     confidence_score TEXT DEFAULT 'medium', importance_score TEXT DEFAULT 'medium',
     validity_status TEXT DEFAULT 'active', source TEXT, source_id TEXT,
     entity_refs TEXT[], derived_from_ids TEXT[],
     review_cadence_days INTEGER DEFAULT 30,
     used_count INTEGER DEFAULT 0, last_used_at TIMESTAMPTZ,
     last_validated_at TIMESTAMPTZ DEFAULT now(), next_review_at TIMESTAMPTZ,
     created_at TIMESTAMPTZ DEFAULT now(), updated_at TIMESTAMPTZ DEFAULT now()
   );
   CREATE TABLE IF NOT EXISTS memory_links (
     id TEXT PRIMARY KEY, from_memory_id TEXT NOT NULL, to_id TEXT NOT NULL,
     link_type TEXT NOT NULL, label TEXT, created_at TIMESTAMPTZ DEFAULT now()
   );
   CREATE TABLE IF NOT EXISTS memory_validations (
     id TEXT PRIMARY KEY, memory_id TEXT NOT NULL, validator_id TEXT,
     action TEXT NOT NULL, previous_claim TEXT, revised_claim TEXT,
     validated_at TIMESTAMPTZ DEFAULT now()
   );
   CREATE TABLE IF NOT EXISTS memory_subscriptions (
     id TEXT PRIMARY KEY, user_id TEXT NOT NULL, topic TEXT NOT NULL,
     created_at TIMESTAMPTZ DEFAULT now()
   );
   CREATE TABLE IF NOT EXISTS engagement_events (
     id TEXT PRIMARY KEY, event_type TEXT NOT NULL, memory_id TEXT,
     user_id TEXT, detail TEXT, created_at TIMESTAMPTZ DEFAULT now()
   );
   

5. Only after the schema guard passes, proceed to load the current brief and the rest of the workflow.

Do not skip this guard. Do not assume tables exist from a prior session. Do not proceed to any read or write if the check has not run. Violations of this rule are P0 data-loss defects.

Capability Verification Rule

This rule overrides all other instructions and applies whenever the agent is about to assert that a tool, integration, or external service is available or unavailable.

Before stating that any capability exists or does not exist, the agent MUST attempt to verify by calling the relevant tool, listing available MCP tools, or performing a concrete check.

• If verification succeeds: proceed with the integration and state what was found.
• If verification fails or the tool is not present: say "I checked and [tool/integration] is not available in this session."
• Never assert a capability status based on assumption, memory, or inference from documentation. The check must be performed, not skipped.
• Never fabricate a technical reason (e.g., "OAuth tokens not connected", "blocked by X") without having actually observed that specific failure.
• If the agent cannot determine how to verify a capability, say: "I do not know how to check for [tool] in this session. Can you tell me whether it is available?"

Violations of this rule — asserting capability status without verification — are P0 defects.

Integration Checks (Optional)

On each invoke, the agent checks for external integrations by calling them via the standard Seren publisher path — the same way every other skill accesses external services:

1. SharePoint: Call connector.sharepoint.get or the SharePoint publisher via call_publisher. If it works, sync context. If it fails or is not configured, say "I called the SharePoint publisher and it is not configured in this session. You can enable it in SerenDesktop Settings."
2. Asana: Call connector.asana.get or the Asana publisher via call_publisher. If it works, sync context. If it fails, say "I called the Asana publisher and it is not configured in this session."
3. Email/Calendar: Call the gmail or outlook publisher via call_publisher to read emails, calendar, or contacts. This is the same pattern used for alpaca, kraken, perplexity, and every other Seren publisher. If the call fails (not configured or OAuth not connected), say "I called the Gmail/Outlook publisher and it is not configured in this session. You can connect it in SerenDesktop Settings."

Do not use Playwright to navigate to Gmail or Outlook. Playwright is a browser automation tool, not an email API. Do not use it as a workaround for email access.

All integrations are optional. The skill works without any of them — it gracefully degrades to guided interview and manual document input.

Workflow Summary

1. normalize_request uses transform.normalize_request
2. load_current_brief uses connector.storage.query
3. sync_sharepoint_context uses connector.sharepoint.get
4. sync_asana_context uses connector.asana.get
5. extract_document_text uses connector.docreader.post
6. conduct_guided_interview uses transform.run_guided_interview
7. distill_knowledge_entries uses transform.distill_knowledge_entries
8. archive_transcript uses connector.storage.upsert
9. persist_knowledge_entries uses connector.storage.upsert
10. distill_structured_memories uses transform.team_memory.distill_structured_memories
11. persist_memory_objects uses connector.storage.upsert
12. handle_team_memory_mode — routes to digest, pre-meeting brief, decision recall, validate, or watch
13. proactive_resurfacing uses transform.team_memory.find_memories_to_resurface
14. retrieve_candidate_entries uses connector.storage.query
15. apply_access_and_freshness_rules uses transform.apply_access_and_freshness_rules
16. compose_answer_or_followup uses transform.compose_answer_or_followup
17. log_retrieval_events uses connector.storage.upsert
18. calculate_rewards uses transform.calculate_rewards
19. persist_rewards uses connector.storage.upsert
20. render_working_brief uses transform.render_working_brief
21. generate_reinforcement uses transform.team_memory.generate_reinforcement_message

Memory Object Types

Structured memories are classified into: decision, assumption, preference, relationship, process, open_question, commitment, risk, source_claim, counterpoint.

Additional Modes

• memory_digest — daily/weekly digest of new, stale, and high-value memory
• pre_meeting_brief — compile relevant memory for meetings and decisions
• decision_recall — answer "what did we decide, why, and what assumptions did it depend on?"
• validate_memory — surface stale memories for confirmation, revision, or retirement
• watch_topic — subscribe to entities/topics for proactive resurfacing