HivemindOS manual

Shared skills

A skill is a reusable playbook for an agent. It can teach a writing style, a research method, a deployment workflow, a media pipeline, or any other repeatable way of working.

Hive Superbrain keeps shared skills in your Obsidian vault so supported agents can use the same reviewed instructions. You do not need to copy a skill separately into every runtime.

Use a shared skill

Open the Skill Browser to see what is installed and browse optional packs. You can also ask HivemindOS to find, install, create, or improve a skill for a specific job. Optional catalog skills are installed only when you choose them.

When a skill applies, the agent reads its instructions before acting. A skill can add a workflow or safety check, but it does not grant permission to spend money, publish, delete data, or take another outward action you did not request.

What is included

Fresh installs include a small core shelf for safe engineering work, capability discovery, shared memory, compiled knowledge, workflow creation, and Obsidian-native Markdown, Bases, and Canvas files. Larger or specialized packs remain optional so the shared shelf does not fill up with methods you never use.

The Packaged Skills guide lists built-in and optional skills. The Skill Browser is the easier place to install them.

Your existing runtime skills are safe

HivemindOS makes shared skills visible to supported runtimes through managed copies. It preserves runtime-local skills that you installed yourself. If a local skill has the same name as a shared skill, HivemindOS does not overwrite the unmanaged local folder.

When an imported plugin skill explicitly names a helper file it needs, HivemindOS keeps that helper with the shared skill. Unreferenced plugin files stay in the original plugin, so shared copies remain focused and portable.

If the shelf starts showing duplicates or old test skills, use the health tools described in Sync and health.

Open the skill projection and catalog reference

Skills/README.md
Skills/<slug>/SKILL.md

Agents should treat the shared shelf as the first skill source. Runtime-local skill folders are supplemental overlays for runtime-specific or personal skills. Agents should read Skills/README.md first, then the relevant SKILL.md.

Runtime Projection

Setup projects the shared shelf into supported runtime skill roots as HivemindOS-managed cache folders. This lets runtimes that only scan their own skill directory, such as Codex, still see shared-brain skills like agent-reach as native skills.

Projection is non-destructive:

HivemindOS may replace a runtime skill folder only when it carries .hivemind-skill-source.json with managedBy: "hivemindos" or shared/bundled HivemindOS provenance.
Unmanaged runtime-local skill folders are preserved and skipped on slug collision.
On collision, the shared shelf remains the logical primary source for HivemindOS routing, while the existing runtime-local folder remains available as a supplemental local skill.
Uninstall removes only HivemindOS-managed projection folders unless the user separately removes a runtime’s own skills.

Provider Inventory

HivemindOS can inspect local and remote runtime skill providers, including common Codex, Claude, Hermes, Gemini, OpenClaw, and Aeon locations.

Provider inventory and import logic live in:

src/lib/services/obsidian/brain-skills.ts

Auto-Sync

Auto sync policy is stored outside the vault:

~/.hivemindos/skill-auto-sync.json

Policies can auto import, auto update, track removals, and optionally allow provider deletion handling.

Shared brain managed mirrors must not get imported again as brand new provider skills. That is how duplicate loops start.

Provider inventories must classify runtime mirrors as already handled before import or reconcile. In particular, an Aeon provider skill under ~/.aeon/skills or /root/.aeon/skills whose directory is already provider-prefixed, such as aeon-*, aeon-aeon-*, claude-*, or hermes-*, is a mirror artifact rather than a new shared-brain skill. The dashboard importer rejects these again even when an older remote collector sends them as importable.

Auto-Installed Brain Skills

Setup copies a small default pack from packaged-skills/auto-install/ into the shared vault Skills/ shelf.

That default pack includes harness-engineering, an attributed fixed-worker experiment method for improving agent context and tools with outcome evidence, and hive-skill-autoresearch, the review-gated procedure used when repeated skill failures qualify for measured improvement. Skill autoresearch preserves the installed skill while native HivemindOS agents—or optional Evo experiments—evaluate candidates through the same harness contract.

That default shelf includes engineering-discipline, HivemindOS’ risk-scoped engineering orchestrator. Users who want the fuller method library can install the HivemindOS Engineering Discipline pack in the Skill Browser. Its manifest resolves every skill from the canonical packaged auto-install and optional directories, adds selected obra/superpowers methods, archives older HivemindOS-managed adaptations before refreshing them, and leaves unmanaged colliding skills untouched.

The HivemindOS Hive skills are included by default:

Skill	Why it exists
`photo-library-keyworder`	Guides the bundled Photo Library through read-only inventory, representative calibration, resumable vision indexing, human review, and protected Lightroom-compatible XMP export. It requires no Python or source checkout.
`harness-engineering`	Holds the worker constant, compares baseline and treatment runs, separates available/retrieved/invoked/relevant context, and keeps only interventions supported by outcome and proof evidence.
`create-zero-human-company`	Turns a business goal or existing repository into a durable, approval-gated company record, verifies its crew and operating setup, and leaves autonomy stopped until explicitly launched.
`bankr-app-development`	Teaches agents to create and publish Bankr apps with the iframe/backend SDK, minimum permissions, wallet identity, public data, schedules, secrets, x402 actions, cache-resistant hosted shells, exact live verification, and coordinated rollback.
`hive-assimilate`	Requires agents to search pinned sources, the shared brain, user projects, local/private indexes, and public GitHub before building software from scratch.
`hive-pulse`	Gives agents a bundled last-30-days signal brief across social, developer, market, GitHub, and web sources using a pinned MIT licensed engine, a setup-installed `hive-pulse` command shim, and Hive safety rules.
`hive-quant-research`	Gives agents a research-only, schedulable quant workflow with typed hypotheses, lagged Rust simulation, independent Python statistics, fail-closed overfitting gates, and hashed local artifacts.
`hive-capability-search`	Finds available tools, skills, apps, agents, credentials by key name, and delivery channels such as slash commands, API routes, MCP tools, CLIs, and dashboard surfaces for a task.
`hive-remote-capability-use`	Runs remote connected apps and fleet capabilities selected by capability search, using fresh discovery, app-proxy routing, private file transfer, artifact verification, and side-effect gates.
`hive-skill-fusion`	Turns a capability request into a reusable shared-brain skill.
`hive-workflow-fusion`	Composes multi-step hive workflows from skills, apps, agents, and tools.
`hive-aeon-fusion`	Converts reusable hive workflows into AEON-ready agent duties when appropriate.
`hive-brain-memory`	Teaches agents typed recall, canonical durable-memory heads, operational-event separation, review-gated pattern proposals, and evolution with superseded-history preservation.
`hive-brain-compiled-wiki`	Teaches agents the HivemindOS compiled-brain workflow: entity/concept/summary writes, compiled-wiki search, graph-native MCP reads, wiki health, and human collective shared-brain contribution rules.
`hive-skill-autoresearch`	Turns repeated, evidence-backed skill failures into a reviewed optimizer task with a measured baseline, four candidate variants, regression floors, and no automatic replacement.
`notebooklm`	Teaches agents to use the optional local NotebookLM MCP integration with explicit notebook IDs, private machine-local authentication, verified artifacts, and confirmation for destructive or outward actions.
`wrapup`	On an explicit wrap-up request, deduplicates or evolves durable typed memory and archives a concise redacted session summary in the user’s verified NotebookLM AI Brain notebook.
HyperFrames suite (`hyperframes` + 18 siblings)	Routes explicit HTML / HyperFrames video work into bundled HTML, CSS, media, and seekable-animation workflows. Generic creation requests first present an actionable choice between cloud AI, local AI, and HTML / HyperFrames; ordinary discussion stays conversational.

The Obsidian Native Brain Pack is also included by default:

Skill	Why it exists
`obsidian-markdown`	Write Obsidian Flavored Markdown with wikilinks, embeds, callouts, properties, tags, math, Mermaid, and footnotes.
`obsidian-bases`	Create and edit native `.base` YAML views over vault notes.
`json-canvas`	Create and edit Obsidian `.canvas` maps, boards, and flowcharts.
`defuddle`	Optionally extract clean markdown from web pages when the local Defuddle CLI is installed.

These are curated from kepano/obsidian-skills. HivemindOS does not auto-install the upstream obsidian-cli skill because the shared vault already carries safer, HivemindOS-aware Obsidian CLI skills and write policy.

GitHub Pages packaged-skill docs live under docs/packaged-skills/. That section splits HivemindOS-owned Hive skills from curated third-party packaged skills and optional catalog skills.

Optional catalog skills are not auto-installed. Optional crypto helpers such as packaged-skills/optional/crypto/hivemindos/b20-issuer-proof/, GTM helpers such as packaged-skills/optional/gtm/athm793/local-business-scraper/, the Untested • Experimental packaged-skills/optional/gtm/hivemindos/distribution-first-app-validation/ workflow, the proposal-only packaged-skills/optional/gtm/hivemindos/paid-media-reallocation/ budget review and read-only packaged-skills/optional/gtm/hivemindos/seo-aeo-audit/ visibility review, the private local-first packaged-skills/optional/productivity/madslorentzen/ai-job-search/ workflow, writing helpers such as the security-audited packaged-skills/optional/writing/petergyang/no-ai-slop/, HivemindOS-authored production helpers such as packaged-skills/optional/brand/hivemindos/brand-book-concept-page/, packaged-skills/optional/brand/hivemindos/hivemindos-brand-visuals/, packaged-skills/optional/events/hivemindos/venue-activation-visualizer/, packaged-skills/optional/media/hivemindos/launch-video-hyperframes/, and packaged-skills/optional/ops/hivemindos/work-board-airtable-bridge/, the UI Skills design catalog under packaged-skills/optional/design/<source>/<skill>/, the 130 current MIT-licensed skills in the pinned, security-audited MengTo catalog under packaged-skills/optional/design/mengto/<skill>/, the security-audited official GreenSock GSAP set under packaged-skills/optional/design/greensock/<skill>/, and the pinned, conditionally approved Three.js Awesome Graphics set under packaged-skills/optional/design/scottstts/<skill>/ are copied into the shared brain only when the user chooses a skill or pack from the catalog. The MengTo packages exclude upstream demos and install from hashed local files; two upstream-removed legacy packages remain available at their prior pin. Three.js executable examples remain restricted at use time, and their GPL/third-party licensing boundaries remain visible after installation.

Shared skills can also be authored directly into the vault for recurring workflows. ai-ugc-production-pipeline owns the performance-creative and asset-production side of AI UGC: public-ad evidence coding, pain/reframe briefs, five-beat scripts, storyboard-first generation, version-aware MiniMax H3/Higgsfield/Seedance routing, local asset preflight, preview-before-spend gates, idempotent job manifests, word-timed captions, bounded batch assembly, cost-controlled variants, and metric-driven regeneration. It distinguishes edit-only derivatives from variants that require new generation, and its authored reusable rows live in the performance-creative GTM workflow bank. video-generator-prompting is the provider-neutral router used before video prompt compilation: it resolves the exact runtime, endpoint, task mode, duration, frame inputs, and audio behavior, then loads the reviewed model guide and records that decision in the generation receipt. minimax-h3-video-prompting owns MiniMax H3 T2VA/I2VA/FL2VA/L2VA/Ref2VA compilation, boundary-frame and full-reference semantics, native audiovisual fields, current license/territory gating, and artifact QA; it links to and paraphrases the official MiniMaxAI guides rather than redistributing them. All three are available as optional HivemindOS media packages. content-rewards-viral-app-campaign operates the distribution side: app-readiness gates, viral format banks, creator scoring and course briefs, verified-view reward controls, full-funnel diagnosis, bounded cohort and format-decay loops, and a rights-cleared organic-to-paid handoff. Its reusable operating rows live in the creator-rewards GTM workflow bank. instagram-reel-growth-workflow captures the Instagram account-study to AI Reel production loop: public profile analysis, current niche research, retention scripting, hook engineering, and a daily human-approved output workflow. video-shot-transcript is another shared media-analysis skill: it teaches agents to break local videos into unique shot/angle segments, align each segment with transcript or visible captions, and keep local media private unless an external transcription or vision upload is explicitly approved. higgsfield-api-quirks records Higgsfield model-specific API workarounds so agents use the correct Seedance 2.0 audio, aspect-ratio, and reference-slot schema instead of relying only on generic API examples.

The native pack pairs with seeded vault views under Operations/Brain Services/:

Agent Memory.base
Project Brain.base
Secure References.base
Hive Superbrain.canvas

AEON Mirror Rule

The hidden AEON runtime mirror belongs here:

Operations/Runtime Mirrors/AEON/.aeon

Human AEON profile notes stay under Agents/AEON/<profile>/. A hidden runtime basename like .aeon must not become a visible profile folder.

Test Skill Rule

Real fleet propagation tests create temporary hive-e2e-* skills. Tests must clean up their shared brain copies after proving import, update, and removal tracking.

vault-doctor also detects old active hive-e2e-* shared skills.

Skill Hygiene

Use vault-doctor when the shared shelf looks noisy:

pnpm vault:doctor
pnpm vault:doctor -- --fix

The doctor reports duplicate active skill hashes, conflict artifacts, and test leftovers.