Skills — the 35 on-demand playbooks

Skills are plain-Markdown playbooks Mushegh reads with read_skill when a turn calls for one. They live in skills/, ship with every deploy, and are listed by list_skills. None auto-load — each is loaded on demand, so the per-turn prompt stays lean.

How skills are used

"read your self-improvement skill, then make recall faster"
  → list_skills → finds self-improvement → read_skill → branches, edits, tests, confirms, commits.

"give me my morning briefing"
  → reads daily-briefing, uses it as the rubric. Returns the spoken text.

"research X"
  → reads research-method, follows it. Falls back to whatever tools are available.

The skill catalog is a deliberate, framework-curated set — vetted, written against the actual tools, and small enough to keep coherent. Mushegh does not install third-party skill packs at runtime; the catalog ships with the brain and is regenerated on every scripts/deploy_vps.sh.

Companion

Skill	What it teaches Mushegh
`daily-briefing`	Morning voice briefing of overdue, due-today, this-week deadlines, recurring tasks.
`email-triage`	Triage unread email: urgent / reply-needed / FYI / archive.
`calendar-and-reminders`	Read today / create events / schedule reminders.
`task-capture`	Capture a task from anywhere: inbox, Notion, voice.
`research-method`	Multi-source research, cross-check, cite.
`proactive-etiquette`	When to speak, when to stay quiet, quiet hours + budget.
`voice-style`	Mushegh's own voice: pacing, filler phrases, no markdown on the wire.
`memory-discipline`	L1/L2/L3/L5: what goes where, when to forget, when to skip.
`task-cleanup`	Turn a messy backlog into a short, honest, actionable list — never silently deleting anything he might still want.

Coding

Skill	What it teaches Mushegh
`self-improvement`	Read your own source, run the suite, make a reversible commit.
`adding-a-tool`	Drop a SCHEMAS+HANDLERS module into brain/tools/, list it in tools/__init__.py.
`debugging`	Where to look, what to tail, how to repro a voice turn.
`git-workflow`	Branches, reversible commits, no force-push, deploy_vps.sh gate.
`python`	Style: type hints, fail-quiet, 100 cols, the surrounding idiom.
`web-and-typescript`	TypeScript / Next.js patterns for the docs site + clients/.
`prompting`	Two audiences: prose for him, compressed exactness for a machine. Plus the make-the-task-checkable habit.
`integration-utilization`	Reach for Graphify, REA, Ponytail and Caveman — and what not to reach them for. The order: understand, check, build smallest, say it small.
`coding-pipeline`	Which tool gets the work, and why the routing is not yours to choose. Read before writing a coding brief — the first decision is where it goes, not how to phrase it.
`architecture-modelling`	Decide where a change belongs before writing it. A brief that names the wrong seam produces a diff that works and lives in the wrong file.
`graphify`	Read the structure before the source: a scoped query is ~1.6k tokens where a full-repo read is ~120k. Cheapest layer first, stop at the first that answers.
`ponytail-and-caveman`	Ponytail spends less code; Caveman spends fewer tokens. Two currencies, easily confused, and which one is actually installed here.
`prompting-claude-code`	The main coding tool. It is an agent, not a code generator — brief it like a competent engineer who has never seen the repo.
`prompting-mcode`	The first fallback, reached when Claude Code failed on evidence rather than a hunch. Richest non-interactive interface of the three, which changes the brief.
`prompting-mistral-vibe`	The second fallback. Two agents have already tried, so the brief is no longer the original — it is the original plus what the other two learned.
`prompting-open-design`	Owns frontend exclusively — sites, UI, UX, visual direction. An MCP server, so its tools arrive namespaced.

System

Skill	What it teaches Mushegh
`mushegh-architecture`	The edge/brain split, the protocol, the six memory layers.
`pc-control`	Drive the laptop from the brain; PC_LINK + the executor.
`google-workspace`	Gmail + Calendar OAuth, refresh token, one-time consent.
`media-and-voice-rooms`	Telegram voice-chat streaming, music room, voice-note outbound.
`governance-and-compliance`	What actually controls him, named mechanism by mechanism — and the absences, which matter as much. Written after he invented a vendor for his own architecture.
`governance-regulatory-mapping`	GDPR Art. 5 / 32 / 9 and the secrets row: control AND gap, never the gap skipped. Split out so neither file overflows what read_skill returns.
`secrets-and-access`	The paths and commands the code actually refuses, read from the enforcers. Why 'I don't have that key' is a correct answer.
`sessions-and-context`	The real trim numbers, what survives it, and when a reset is right — plus when it is just a way of dodging a hard problem.
`owner-model`	Tiered context: a small injected core, the rich model on demand. LifeOS's taxonomy mapped onto this repo's storage, and the rule that keeps it true.
`the-fleet`	Eighteen specialists, eighty-three domains, and no router. What to unlearn first: the team lead that used to choose a specialist no longer exists.

Adding your own

Drop a skills/<name>.md in the repo, ship via scripts/deploy_vps.sh, and it appears in list_skills on next brain start. Treat skills the way you'd treat a runbook: short, example-rich, and tied to the actual tools. See skills/adding-a-tool.md for the pattern.

One skill = one focused playbook. Resist the urge to merge things (“research-and-coding-and-writing”) — a skill that tries to teach three jobs rarely teaches any of them well.