Skip to content

SKILLS.md

Latest commit

 

History

History
120 lines (94 loc) · 5.05 KB

SKILLS.md

File metadata and controls

120 lines (94 loc) · 5.05 KB
Raw

The skills, in depth

brainstack ships 14 native Hermes skills. Each one is a directory under skills/ with a SKILL.md the agent loads on demand. They are designed as a pipeline: capture feeds memory, memory feeds retrieval, consolidation keeps everything honest.

The daily loop looks like this:

capture / signal / meetings          (things come IN)
        │
memory / reference / people-crm / trackers   (things get TYPED)
        │
index / retrieval                    (things stay FINDABLE)
        │
consolidate / reflect / health       (things get STRONGER overnight)
        │
agents                               (things get DELEGATED safely)
        │
distill                              (patterns become new skills)

Intake

capture — the front door

Turns anything landing in 00-Inbox/ (PDFs, articles, transcripts, clippings) into atomic notes in under 30 seconds. Smart extraction only — full documents are never dumped into the vault. Pairs with brain inbox for batch runs. Trigger it with: "process my inbox", "put this article in the brain".

signal — the owner's voice

Captures stray thoughts, principles, and quotations from you verbatim, with an ID. No paraphrasing: the original voice is the value. Trigger it with: "signal: ship beats perfect", "save that thought".

meetings — decisions with provenance

Ingests meeting notes or transcripts, extracts decisions and action items, and links the meeting to the people and projects involved. Trigger it with: "ingest this meeting", "what did we decide with Acme?".

Typed knowledge

memory — the schema layer

The heart of the system. Every fact gets a type, a confidence, a TTL, and provenance. Conflicts resolve by recency + confidence; stale facts decay instead of silently lying to you. Governs what is stored and how it ages.

reference — durable deep-dives

Authoring standard for reference notes (articles, theses, frameworks, whitepapers) with uniform frontmatter so they stay findable years later.

people-crm — relationships as plain notes

People, organizations, projects, and interactions as typed Markdown notes with an index. A full personal CRM with zero SaaS. Trigger it with: "log an interaction with João", "what do I know about Acme?".

trackers — collections that stay queryable

Books, movies, and series with a typed schema, progress logging, and an INDEX.

Navigation & recall

index — the map

Maintains Maps of Content (MOCs) and the Tree Index so the growing graph stays browsable and new concepts are discoverable from the top.

retrieval — ask the brain first

Answers questions from the vault: hybrid semantic search plus Iron Law backlinks, with synthesis and gap analysis. The rule: consult the vault before any external API. Trigger it with: "what does the brain know about X?".

Maintenance (where compounding happens)

consolidate — the dream cycle

Session-end and nightly consolidation: dedupe, backlink repair, Tree Index refresh, NOW.md cleanup, contradiction detection, memory decay, next-day priorities. Automated by the cron installed at setup (brain dream runs it on demand).

reflect — from notes to knowledge

A 5-stage pipeline that turns raw captured content into durable, connected knowledge: extract decisions, surface patterns, flag contradictions, find gaps.

health — the vault audit

Scored report on orphans, broken links, duplicates, schema drift, index staleness, and Iron Law violations, with prioritized fixes. CLI: brain doctor.

distill — knowledge that compounds

When a problem has been solved twice, distill turns the workflow into a new reusable skill — so the third time it's solved in seconds.

Governance

agents — delegation with a boundary

Runs the multi-agent operating model: domain workers (research, network, content, finance, vault) with isolated context, context modes per work domain, and the credential boundary — work flows upward, credentials do not flow downward. Workers never publish; the orchestrator verifies and delivers. Trigger it with: "delegate this research", "set up a context mode".

The engine behind them

The skills call small Python scripts in bin/, all exposed through one CLI:

Command Script Does
brain inbox inbox-processor.py Batch-process 00-Inbox/
brain index semantic-index.py Build/refresh the search index
brain search <q> semantic-search.py Hybrid semantic search
brain compile memory-compiler.py Compile typed memory
brain decision <t> register-decision.py Log a decision with provenance
brain session consolidate-session.py End-of-session consolidation
brain dream dream-cycle.py Full nightly consolidation
brain doctor vault-health.py Scored vault health report
brain sync vault.sh Git commit + push the vault

Zero required Python dependencies. If sentence-transformers is installed in the setup venv, search upgrades from keyword TF vectors to true embeddings.