second-brain

Second Brain

Bootstrap, maintain, or operate an LLM-managed personal wiki using the LLM Wiki at Scale pattern. The LLM incrementally builds a persistent, interlinked wiki from raw sources — knowledge compiled once, kept current, compounding with every ingest.

Three branches. Read the disclosed file before executing.

---

Shared Rules

• TaskCreate visible. Use TaskCreate to make stages visible to the user on every branch.
• Scripts propose, LLM disposes. Deterministic scripts hand clean worklists; the LLM exercises judgment. Never blind-apply script output.
• Anti-bleed rule. "Lean" applies ONLY to operational files (AGENTS.md, playbooks, scripts). Content generation — ingestion, entity extraction, concept pages — must be relentless: sprawling, exhaustive, granular. A rich source spawns 10-15 pages. If you produce one summary page, you are committing premature completion and stranding knowledge. The schema is lean so the content can be massive.
• Handoff. Day-to-day ingest and query run through the vault's own AGENTS.md (or CLAUDE.md) playbooks once live — not this skill.

---

Branch: Setup

When: User wants a new vault, or to continue a partial bootstrap. Read: references/setup.md Completion: Every phase (Discovery → Scaffold → Interview → Inject → Build → Orient) completed. Zero {...} placeholders remaining in generated files. User oriented with core commands and a suggested first action.

---

Branch: Maintain

When: User asks to lint, health-check, rebuild index, rotate logs, calibrate confidence, or normalize tags. Read: references/maintenance.md Completion: Lint loop exits with zero criticals. Index rebuilt. Log entry appended. Every worklist item accepted, modified, or declined by the user.

---

Branch: Operate

When: User wants to build or schedule an autonomous wiki-maintenance Operator. Read: references/operator.md Completion: Operator prompt saved with zero {{...}} placeholders. Schedule wired via the schedule skill. User told the cadence and how to manage the trigger. Scope: Pure wiki upkeep, scheduled and unattended — lint, the three safe fixers, index rebuild, log rotation, source-queue and scratch visibility — run locally against the vault with no external connectors and no MCP. It never resolves judgment calls (broken links, contradictions, tag drift) and never auto-ingests; both stay queued for a supervised session. Optionally researches open questions into raw/scratch/ for later review — never into wiki/, and off by default.