second-brain-template

Seed a brain from a LARGE existing source by TRIAGING it, then calibrating the agent's judgment on small batches, then bulk-mining. Use when the owner says "mine my [source]", "mine this export", "calibrate the mining", "idea mining", "process my archive", or "seed my brain from this export" — pointing at a sizable export (notes-app dump, chat history, voice-note archive, bookmarks, podcast snips). Phase 0 TRIAGE: categorize the source into keep/skip bins, sample the skip-bins, get owner approval per bin, build the target list. Phase 1 CALIBRATE: small batches → present COMPACT numbered proposals (verbatim quote + the agent's call + one-line reason, no tables) → owner corrects → write every correction back into the rules (CONTENT_TAXONOMY / KEYWORD_GUIDE / INTELLECTUAL_LANDSCAPE) → draft the full Template only after the calls are approved → repeat until a batch is clean. Phase 2 SET LOOSE: process the remainder autonomously, spot-check, re-calibrate on drift. Implements "invest in structure and calibrate the

Onboard a human into THEIR own second brain — the guided interview that turns this blank template into their personalized system. Use when the user says "set this up", "let's set it up", "set up my second brain", "personalize", "personalise", "make this mine", "onboard me", "help me fill this in", "walk me through PERSONALIZE", or starts a fresh session wanting to begin. Drives the whole __FILL_FROM_USER__ fill-in via an interview. Resumable — pick up where you left off. NOT for ingesting content (use /ingest-brain-dump) — this is first-run setup.

Handles connections between related ideas. Options include creating folders, adding shared keywords, merging content, or rejecting false positives. Used when ingest-brain-dump identifies related content.

Weekly self-improvement of the HARNESS itself — not project QA. Reviews the last N days of this brain's Claude Code session traces for harness-level friction (where the owner corrected the agent, repeated permission prompts, skills that misfired or weren't used, rules/memories the agent ignored, weak sub-agent prompts), drafts a proposal of findings, and the owner reviews + applies it finding-by-finding (gated — nothing auto-applies). Uses the in-session Workflow tool to fan out over trace chunks; no external services, no API keys. Use when the owner says "/harness-review", "review the harness", "self-review", "what's the agent getting wrong", or accepts the optional offer at the end of /weekly-maintenance.

Full orchestrator for stream-of-consciousness input pasted or dictated into the Claude Code session. Pipeline: explore-second-brain (find related files) → semantic segmentation → verify-idea (dedupe check) → process-content (create Template A files) → validate_template.py → user approval → save. Use when user pastes unstructured text or dictates a voice note into chat and says "brain dump", "process this", "ingest this", "save this", "what should we do with this". One paragraph may contain 3 ideas; 5 paragraphs may be 1 idea — always semantic, never algorithmic chunking. NOT for phone inbox (use a phone-inbox ingest skill if configured). NOT for podcast snips (use a snip-processing skill if configured). NOT for bookmark exports (use a bookmarks-ingest skill if configured). For a LARGE existing source export (notes-app dump, chat history, voice-note archive), use /mine — it triages the source, then calibrates the agent's judgment on small batches before the bulk pass. Preserves the owner's exact language — no

Low-level mechanic that turns raw text into Template A/B files. Segments semantically (never algorithmic chunking), classifies each segment, creates files, runs validate_template.py. Does NOT sync from external sources — text must already be in context. Called by the ingestion skills (e.g. /ingest-example) as their file-creation stage; can also be invoked directly when user pastes content and says "make files from this", "turn this into Template A", "parse this into ideas". NOT for voice/phone inbox work (use an inbox-ingest skill), NOT for searching the brain (use explore-second-brain or semantic-search). If input has multiple distinct ideas, all are processed in one invocation.

Fast semantic search over Second Brain using local embeddings (BAAI/bge-large-en-v1.5, via `sb_embed.py`). Finds ideas by MEANING, not just keyword overlap. **Commands:** `search "query"` (top-k ranked matches, ~50ms cached), `verify "idea"` (duplicate check with DUPLICATE/OVERLAP/NOT_FOUND verdict), `similar path/to/file.md` (find files semantically close to an existing one), `stats` (index health). **Filters:** `--project "Project Name"`, `--category "Journal"`, `--top-k N`, `--json`. **Use semantic search for:** broad concept recall ("what do I have about coordination?"), rephrased-duplicate detection, unexpected-connection discovery, pre-ingestion dedupe. **Use grep instead when:** you know the exact term/phrase, doing project/category lookups by exact string, or confirming a specific filename. Grep is precise; semantic is recall-wide. Best practice: combine — semantic for broad recall, grep for precision matching. Invoked via "find ideas about X", "what do I have on X", "related to X", "similar to this i

Semantic duplicate detection for Second Brain ideas. Distills input to core concept, runs `sb_embed.py verify` (embedding-based similarity, ~50ms), reads top candidates, and returns one of four verdicts: **DUPLICATE** (>0.85 similarity — exact same idea, point to existing file, skip creation), **OVERLAP** (0.70–0.85 — related but distinct, consider merging or linking), **NOT FOUND** (<0.70 — safe to create new file), **UNCERTAIN** (ambiguous matches — surface them to human for judgment). Falls back to grep for keyword precision when semantic search is ambiguous. Core rule: SEMANTIC match, not keyword match. Two ideas are duplicates when they propose the same unique insight, regardless of wording — shared keywords with different proposals do NOT count as duplicates. Called by the ingestion skills (e.g. /process-content, /ingest-example) before file creation. Also invoke directly when user asks "is this captured?", "do I have this?", "check for duplicates", "have I thought about this before?".

Intelligent weekly maintenance with parallel review agents. Updates counts/dates, spawns 3 Explore agents (Doc Coherence, New File Quality, Project Pulse), auto-applies mechanical fixes, surfaces only judgment calls. Use when user says "weekly maintenance", "maintenance", "weekly sync", or on Sundays.

Persist session progress before context compaction or session end. Use when user says "wind down", "checkpoint", "save progress", "done", or "finished". Auto-triggered via PreCompact hook. PostCompact hook verifies execution.

Evolve the golden few-shot examples your brain's processes use. The owner names a task + a recipe; the skill runs the evolution cycle (present outputs, collect the owner's scores, track versions, propose mutations) and saves a versioned golden set the other skills read at prompt-build time. Tasks: classification (taxonomy routing, used by /ingest-brain-dump + /mine), summarization (Template A/B summaries, used by /process-content), keyword_extraction (used by /process-content), template_creation (full Template A from raw input), segmentation (idea-boundary detection in long text), and delegation (orchestrator → sub-agent prompts; see Orientation_Docs/SUBAGENT_DELEGATION_PRIOR_ART.md). Recipes: Quick Curation (hand-pick 5), Bootstrap (generate 20, filter to top 5), Deep Evolution (meta-prompt iteration), Escape Local Maximum (aggressive mutation + wildcards), Hybrid (bootstrap then evolve). Invoke: "/golden-evolver", "/golden-evolver classification", "evolve examples for keyword extraction", "quick curation fo

Updates orientation docs after actions that affect them. Spawns subagent to check all docs and propose updates. Triggered after file operations, status changes, structure changes.

Register a discovered capability, tool, pattern, or gotcha in the right discovery layer for future Claude instances

Eight-axis codebase cleanup sweep. Spawns 8 parallel general-purpose subagents, each owning one cleanup dimension (dedupe/DRY, type consolidation, unused code, circular deps, weak types, defensive try/catch, legacy/fallback code, AI slop). Each subagent researches, writes a critical assessment, and implements all high-confidence recommendations in its lane. A final verify+repair gate then re-runs the target's tests/typecheck against a pre-cleanup baseline and repairs or reverts any regression. Accepts an optional target path argument (defaults to current working directory). Use when user says "/code-cleanup", "code cleanup", "clean up the codebase", "run code cleanup", or names a specific repo and asks for the 8-agent sweep. Do NOT auto-fire on generic words like "cleanup" alone — this skill edits code across 8 dimensions in parallel and should be invoked deliberately. Prompt origin: shaw makes magic (public community share).

Proactively surfaces related ideas from the Second Brain while the owner is thinking out loud. Uses semantic search + grep (via scripts/sb_embed.py) to connect new thinking to existing ideas without being asked. Activates when the owner discusses concepts (AI, singularity, coordination), mentions people, explores themes, or brainstorms. Does NOT activate for file operations, debugging, explicit-instruction execution, or logistics — this is for ideation, not task-completion. **Sensitivity modes** (SETTINGS.md): AGGRESSIVE (current default — fire proactively), MODERATE (fire only on strong semantic matches), CONSERVATIVE (fire only on exact-keyword overlaps). Caps at 3 connections per activation. Watches for rejection signals ("not relevant", "stretch", owner ignores) — after 3+ rejections in a session, proactively offers to dial down sensitivity. Goal: surface USEFUL connections that advance thinking, not overwhelm.

Lint a draft sub-agent prompt against the 10-slot skeleton before dispatching via Task/Agent tool. Catches the high-frequency failure modes: negative framing, numbered algorithms, pre-warned failure modes, step budgets, character-sheet roles, missing schema pin, relative paths, negative examples. Returns a scored audit (PASS / FLAGS) and a revised draft that addresses every flag. Full rationale: your delegation prior-art doc. Invocation: `/delegate-check`, "lint this sub-agent prompt", "audit this Task prompt", "check this delegation", or automatic pre-flight before any Agent() dispatch when the orchestrator has time to iterate.

Recurring internal-disk cleanup. Phase 1 itemizes regenerable caches, local-LLM model stores (HuggingFace / Ollama / LM Studio), dev toolchain versions, and Docker — every item shown with its size, deleted only on explicit per-item confirmation. Phase 2 (optional, runs when an external drive is present) tars the big archival folders (Claude backups, audio recordings) to the external drive, verifies entry counts, then deletes the originals. Use when user says "/disk-cleanup", "clean up my disk", "disk cleanup", "free up space", "my laptop is full", "internal disk is full", or asks for the recurring every-few-weeks cleanup. Phase 1 is safe to run alone; Phase 2 needs the external drive plugged in.

Spawn a subagent to scan Second_Brain files and find content related to new input. Uses semantic search + grep for comprehensive results. Returns a manifest of relevant files. Use before processing new content to avoid duplicates and find related ideas.

Extracts action items from content and routes to project-specific TODO files. Used by ingest-brain-dump when TODOs are identified in brain dumps.

On-demand RAM reclamation for your machine. Audits memory-hungry processes (footprint-accurate, counting swapped/compressed pages), then presents them in four buckets — KILL CANDIDATES (orphan/restartable servers, e.g. a forgotten local-model server or dev server), REVIEW (uncertain), INTERACTIVE sessions (your editor / agent CLI / browser — surfaced for eyeballing, never auto-killed), and MANAGED (supervised services, left alone). Kills NOTHING without explicit per-item confirmation, then re-checks memory and reports what was reclaimed. Use when the user says "/free-memory", "free up memory", "free up some ram", "what's eating my memory", "my machine is slow/swapping", "kill stale processes", or "clean up processes". **Calibrate it to your machine via /setup** (RAM size, which servers are disposable, which apps are interactive) — it works with safe defaults until then.

Launch the HTML Tweaker — direct-manipulation styling for LLM-generated HTML. Drag an HTML file in, fiddle sliders/toggles/color pickers for CSS custom properties, swap between aesthetic presets (Cockpit / Manuscript / Ares), and export either the restyled HTML or a Claude Code skill bundle that makes future generations match the chosen taste. Use when user says "/html-tweaker", "tweak this html", "style this html", "open html tweaker", "design playground", "I want to adjust this", or asks to fiddle/tune/restyle an LLM-generated HTML artifact.

Launch an invention from elevator pitch to spec sheet - tiered pipeline for shipping ideas from Inventions/ files

Load the essential Second Brain (large) into context. Your core original thinking: journals, inventions, predictions, writing, health, lists. Light trim removes redundant metadata; keeps Summary, Keywords, Connections, Original Text. Excludes external sources, bulk fiction, transcripts, code/builds, and docs already loaded via rules. Use when user says "load brain", "load everything", "full context", "deep context".

Generate images using Google Gemini 2.5 Flash - fast, 1024x1024, up to 3 reference images

Generate images using Google Gemini 3 Pro - up to 4K resolution, 14 reference images, advanced reasoning

Scaffolds a new Second Brain project folder via interview. Interviews the owner on name, type, tier, intent, done-criteria, and scaffold richness, then creates the canonical Second Brain project docs (ORIENTATION.md + UNOFFICIAL INTENT_SPEC.md + TODO_[Name].md) plus optional CLAUDE.md / BUILD_SPEC.md / COCKPIT.md / DESIGN.md / README.md depending on project type. Each project gets an INTENT_SPEC.md — a short vision doc you author at project start and refer back to — drafted UNOFFICIAL until you review and promote it. Use when user says "new project", "scaffold project", "start a project", "/new-project", "create a project folder", or asks to begin a competition/campaign/build/writing project that needs structured docs.

Helps you finish projects. Auto-activates when the conversation mentions any project registered in Orientation_Docs/INTELLECTUAL_LANDSCAPE.md — that is the authoritative active-project roster; pull live status from STATE_OF_SECOND_BRAIN.md. High-signal triggers include whatever active projects you have listed there. Also activates on common variations and shorthand names for those projects. Surfaces current state + last-touched + blockers + single next action, using FINISHING_FRAMEWORK.md. Archived projects — only activate when user asks by name. Philosophy: ship things, not perfect things — one concrete next action at a time.

Remove a visible watermark from an image with local OpenCV inpainting (no Docker, no hosted model, sub-second), self-calibrating across aspect ratios via a Sonnet vision locate + before/after QA loop. Built for the Gemini bottom-right sparkle but works on any visible mark. Use when the user says "remove watermark", "dewatermark", "remove the gemini watermark", "clean watermark", or points at a watermarked image / folder. Does NOT touch invisible SynthID provenance marks.

Surfaces random ideas from the Second Brain for serendipitous rediscovery. Combats the "you can't find what you don't know to search for" problem. Invoke with /surprise-me or when user wants to browse ideas without specific keywords.

Tufte-style visualization critique/design skill. Use when user says "/viz", asks to design or improve charts, dashboards, data tables, comparison views, scorecards, cockpit UIs, analytical HTML artifacts, trading/analytics dashboards, idea-ranking cockpits, feed-curation tasklists, or review-queue views. Applies data-ink ratio, chartjunk removal, graphical integrity, small multiples, sparklines, layering, provenance, and high-density/low-slop defaults.