A practical AI workflow system: Hermes for execution, reusable skills for procedures, delegated agents for parallel work, and a vault-first knowledge system for durable context.
Skills
Reusable procedures
Agents
Delegated work
Vault
Durable context
Checks
Verified output
Hermes
Codex
Subagents
Reusable workflow areas
⚙ExecutionHermes runs concrete work from chat, terminal, browser, files, web, and scheduled jobslive▶
Tool-backed actionsCommands are executed and verified instead of described as advice.
Scheduled checksRecurring jobs can collect data, watch services, and report only when useful.
✎SkillsReusable procedures capture how recurring work should be donereuse▶
PlaybooksStable steps for coding, research, GitHub, vault work, shopping, media, and operations.
PitfallsKnown edge cases and verification steps are documented where they are reused.
Cumulative improvementWhen a workflow changes, the skill gets patched instead of rediscovered next time.
⚐Delegated agentsIndependent workers handle research, implementation, and review in separate contextsparallel▶
Parallel workstreamsIndependent subtasks can run side-by-side without flooding the main context.
Fresh reviewA separate reviewer can inspect the diff without sharing the implementer's assumptions.
Verifiable handoffWorkers return concrete paths, URLs, statuses, or findings that can be checked.
☷Personal Knowledge System contextThe vault-backed knowledge system stays the durable source for tasks, references, people, health, and reflectionvault▶
Vault-first memoryDurable context is written back as notes instead of disappearing into chat history.
Typed contextTasks, contacts, purchases, references, health logs, and reflections keep their own structure.
ValidationSchema checks and explicit changed paths keep automated updates auditable.
Operating modes
/Chat to actionShort Telegram prompts become concrete tool-backed work with compact checkpointsact▶
Direct executionThe default path is to inspect, change, test, and report rather than ask for step-by-step permission.
Safe escalationQuestions are reserved for missing credentials, destructive actions, or materially different choices.
⏱Review loopDiffs, checks, and independent review keep output measurablecheck▶
Local gatesRun project checks, inspect diffs, and avoid publishing brittle or stale claims.
Devil reviewA separate pass looks for slop: vague copy, unverifiable numbers, missing verification, and stale assumptions.