VictorGjn

context-engineering

Build and query a token-efficient, provenance-tracked company brain across code, docs, and connector streams. Five capabilities ship as one skill: multi-source indexer (14 languages via tree-sitter AST + markdown heading trees, schema-versioned, incremental); Source ABC (connector contract — connectors live elsewhere, never in this skill); EntityStore (three-tier raw/events/wiki layer with full provenance, semantic-shift consolidation, drift/split/merge auditor); depth-aware packer (2 levels — full / pointer, 95% budget utilization, multi-hop reasoning, query-as-lens reranking, anti-hallucination filters); MCP server. Anabasis `find-links` reference implementation. Use when: an agent needs broad file awareness within a token budget, building or querying a wiki/EntityStore over code + human knowledge, extracting features from a repo (code-to-knowledge), packing entity pages with provenance, multi-hop reasoning across linked concepts, indexing a workspace or GitHub repo. Do NOT use for: single-file reads, when

granola-scribe

Reads Syroco's Granola meeting transcripts and emits events to the context-engineering brain via wiki.add MCP. Ingests customer calls, prospect calls, partner syncs, and internal meetings; assigns entity_hint based on Syroco's known company list; segregates by scope (customers / prospects / partners / internal-product / internal-eng / competitors). First worked example of the scribe pattern in `agent-skills/scribes/SPEC.md`. Tailored to Syroco's actual MCP setup, meeting patterns, and brain corpora.

hubspot-scribe

Reads Syroco's HubSpot CRM (companies, deals, contacts, activities) and emits events to the context-engineering brain via wiki.add MCP. Different shape from granola/slack scribes: consumes property diffs + activity logs, not chunked transcripts. Maps HubSpot company.domain to brain entity_hint; scope derived from HubSpot lifecycle stage. Third worked example of the scribe pattern. Replaces the HubSpot prioritization phase of `product-signals-pipeline` per the migration doc — but moved upstream: hubspot-scribe pushes raw deal context to the brain; consumer skills (fleet-radar-skill, opportunity-builder) apply the prioritization logic at delivery time.

slack-scribe

Reads Syroco's Slack project channels (`#project-*`) and emits events to the context-engineering brain via wiki.add MCP. Scope-aware: each channel is mapped to a known client and lands in the right scope (customer-expansion / new-biz / partners). Filters operational chatter; keeps meeting recaps, alerts, blockers, competitive intel, fleet changes, and product feedback. Second worked example of the scribe pattern in `SPEC.md`. Replaces the Slack fetch phase of `product-signals-pipeline` per `MIGRATION_FROM_PRODUCT_SIGNALS_PIPELINE.md`.

voyage-context-enricher

First enricher skill for Syroco's company brain. Subscribes to events from granola-scribe, slack-scribe, hubspot-scribe (and any future customer-touching scribe). For events whose entity is a customer with active voyages, looks up the voyage context at the event's timestamp: voyage_id, vessel, route, voyage_phase, deal_state_at_time, speaker_role. Emits enriched events into a parallel enrichment log; consumer skills read both and join. This is the keystone of the 3-layer model in `SKILL_INTEGRATION.md` §2: a captain signal in Slack is barely a signal without knowing which voyage they were referring to. This enricher closes that gap.

agent-patterns

Select and implement the right agentic architecture pattern for a task. Use when designing a multi-step AI workflow, choosing between chaining/routing/parallelization/orchestrator-workers/evaluator-optimizer, or when the user asks how to structure an agent system. Do NOT use for single LLM calls or prompt writing (use prompt-craft).

coordinator-pattern

Decompose complex tasks into parallel sub-agent work using the Research → Synthesis → Implementation → Verification pattern. Inspired by Claude Code's Coordinator Mode. Use when a task is too large for a single agent pass, needs multiple perspectives, or benefits from parallel execution. Works with any agent that can spawn sub-tasks.

dembrandt

Orchestrates a multi-dimensional UI/UX task through the full dembrandt design pipeline (brand → tokens → layout → components → UX polish → accessibility/performance gate), loading the right sub-skill at each stage. Use when designing or reviewing a non-trivial UI (a screen, flow, or design system), going from brand to UI end-to-end, or auditing an existing interface across multiple dimensions. Do NOT use for single-dimension tasks (e.g. "review my color palette" — load `algorithmic-color-palette` directly).

install-department

Install a department's slice of the company brain via multi-agent probes of the head's connected tools (Notion, HubSpot, Linear, Slack, Granola, Gmail, etc., via Syroco Connect / Pipedream) plus a structured methodology interview. Produces a Department Spec the runtime can index — covering tools-in-use, roles, cadence, decision flow, taxonomy, automations, and metrics. Use when a department head bootstraps their function into the brain — typically the first install per function (Marketing, Sales, Product, Engineering, Finance, CSM, Support). Do NOT use for one-off knowledge capture, code-context indexing (use context-engineering), or updating an already-installed department.

knowledge-dream

Periodic knowledge consolidation for AI coding agents. Synthesize scattered project learnings into durable, well-organized memory files. 4-phase process: Orient → Gather → Consolidate → Prune. Inspired by Claude Code's autoDream engine. Use when project memory is stale, after a burst of work, or as a recurring maintenance pass.

proactive-brief

Generate concise, actionable status briefs from project state and recent activity. Delta-based reports: what changed, what needs attention, what to do next. Inspired by Claude Code's KAIROS Brief Mode. Works as a manual prompt, a slash command, or a cron job.

prompt-craft

Write and refine production-grade prompts and tool definitions for Claude. Use when the user needs to craft a system prompt, improve an existing prompt, design tool schemas (ACI), or apply techniques like few-shot, XML structuring, role prompting, or chain-of-thought. Covers Claude 4.6 specifics. Do NOT use for skill authoring (use skill-author) or agent architecture (use agent-patterns).

skill-author

Create and refine high-quality Agent Skills (SKILL.md + bundled resources). Use when the user wants to create a new skill, improve an existing one, or turn a workflow into a reusable skill package. Covers structure, progressive disclosure, conciseness, feedback loops, and the quality checklist. Do NOT use for skill validation (use skill-check) or for prompt engineering (use prompt-craft).

shrink-vector-store

Shrink an embedding/RAG vector store 4–32× via int8 or binary quantization with a float rescore pass, preserving recall and provenance metadata. Use when a vector store is too large to be laptop-resident, query cost/latency is too high, or embeddings need to be quantized for FAISS/Qdrant/usearch. Do NOT use for embed-time ingestion failures (e.g. a provider 'too many tokens' 400 — that is an upstream chunking bug, not a storage-size problem) and do NOT enable the dark TurboQuant 4-bit path without the gate below.

entitystore

The schema-agnostic EntityStore engine — raw/events/wiki tiers, Source ABC, semantic-shift consolidator, contradiction auditor, depth-aware packer, and MCP. Carved from context-engineering. Reads the entity schema BY PATH (it ships no schema of its own); the canonical schema + entities live in company-brain, and domain Sources/connectors live in syroco-product-ops. Use when building or querying a provenance-tracked entity brain over any corpus, consolidating events into wiki entity pages, auditing for contradictions/drift, or packing entity context within a token budget. Do NOT use for code-context packing or code-knowledge-graph visualisation (that's the context-engineering skill, which continues independently — entitystore does not replace it).