Agent-Skills.md

Agent Skills: chief-of-staff — the operator work-focus conductor

|

UncategorizedID: ekson73/multi-agent-os/chief-of-staff

Repository

ekson73/multi-agent-os
ekson73License: MIT

Install this agent skill to your local

pnpm dlx add-skill https://github.com/ekson73/multi-agent-os/tree/HEAD/skills/chief-of-staff

Skill Files

Browse the full folder contents for chief-of-staff.

Download Skill

Loading file tree…

skills/chief-of-staff/SKILL.md

Skill Metadata

Name
chief-of-staff
Description
|

chief-of-staff — the operator work-focus conductor

Soul-name: Oikonomos (Greek οἰκονόμος). In Xenophon's Oeconomicus the oikonomos is the steward of the household — the one entrusted to keep the principal's whole estate of affairs in order so the principal is free to act. This is the majordomo / chief-of-staff / right-hand: it does not do the work, it keeps you pointed at the work that matters. It resonates with this operator's own Eko-System — oikos is the same root. (Display-only name. The machine identifier is the slug chief-of-staff.)

Twin (the incoming agent-facing sibling, landing in PR #280 — may not be installed in this repo version yet): reactivate/Entelecheia orients a fresh amnesic agent at cold-wake. chief-of-staff/ Oikonomos keeps the operator (human) on focus across live work. Same premise — correct orientation over scattered state — pointed at the two different consumers.

When to use / not use

  • Use: you (the operator) want a single "keep-me-on-track" pass — what matters now, what did boss/ cowork ask me and by when, what am I letting slip — assembled across all your backlogs at once.
  • Not use:
    • You want to LEARN / route the MAOS framework itself (which tool/protocol) → maos-concierge (framework-facing). This is work-item-facing.
    • You just need the raw aggregated N-Tree of everything, no prioritization/briefing → work-compass directly (this routes TO it and adds the prioritize + people-ask + brief layer).
    • You are mid-task in ONE session and need re-orientation → pulse directly.
    • A fresh agent woke with no context → reactivate (the agent-facing twin — incoming in PR #280; once merged).
    • You have a raw brain-dump to sort into tickets → directive-braindump-triage.

The pipeline (5 phases — lean by default, deepened on demand)

PHASE 0  FRAME       classify the ask (daily catch-up? · drift/what-am-I-slipping? · who-owes-me? · full)
PHASE 1  GATHER      work-compass ── ONE N-Tree of ALL scattered work (7 CPT domains), stale/orphan/pending
PHASE 2  PRIORITIZE  pulse (Eisenhower 2x2, maos-resident) ──[--depth full]──► + ops-strategist 4-lens IF present
PHASE 3  SURFACE     people-ask view (who · when · by-when) + tag/categorize      ◄── newly authored (the gap)
PHASE 4  BRIEF       morning-briefing 7-section SitRep contract (state·done·in-flight·blockers·awaiting·risks·next)
PHASE 5  PRESENT     ONE operator briefing → next-action · Eisenhower quadrants · who-asked/by-when · loose-ends

Composition discipline: PHASES 1, 2, and 4 delegate to skills that already exist and are never reimplemented here. Only PHASE 3 (people-ask projection + tag/categorize) and the PHASE-5 unified operator assembly are authored by this conductor — they were the genuinely absent pieces.

Proportionality (over-engineering circuit-breaker): default depth is quick (0-1-2-3-5; PHASE 4 folds its contract into PHASE 5 without a separate pass). --depth full adds the ops-strategist 4-lens brain + the full SitRep. Do not convene a strategy board to answer "what's next?".

PHASE 0 — FRAME (classify the ask, cheap)

One token sets the depth + which lens dominates: catch-up (default — full landscape + next-action) · drift (what am I slipping / loose-ends first) · who-owes-me (people-ask view first) · full (adds ops-strategist). --mode overrides. Explicit operator instruction always wins.

PHASE 1 — GATHER (delegate: work-compass)

Delegate to work-compass to aggregate ALL scattered work (Jira/Linear/GitHub/sessions/git) into ONE N-Tree across the 7 CPT domains (ticket · worktree · branch · session · thread · process · graph-node), with stale/orphan/pending flags. Read-only — work-compass already prints, never executes. Re-use its output whole; do not re-aggregate.

PHASE 2 — PRIORITIZE (delegate: pulse core · ops-strategist optional)

  • Core (maos-resident, always available): delegate to pulse — Eisenhower 2x2 classification, routing each item ∈ {now · delegate · defer-trigger · backlog · drop}. This is the community-portable prioritizer.
  • Richer (optional, user-scope): IF the user has ops-strategist (a user-scope agent — its 4-lens TPM+CoS+SRE+Agentic Eisenhower over 7 sources), --depth full may route to it for a deeper read. ⚠️ Layer-purity: ops-strategist is NOT a hard dependency — this skill is community-resident and MUST run fully on the maos-resident core alone. Probe for it; use it only if present.

This conductor never computes its own Eisenhower. It routes to pulse (or ops-strategist) — there are already ≥2 prioritizers; a third would be entropy (anti-over-eng).

PHASE 3 — SURFACE (newly authored — the people-ask view + tag/categorize)

The genuine gap: no tool projects "who asked ME for what, and by when?" as a first-class view. Build it as a read projection over the fields PHASE-1 already surfaced — NOT a new store, NOT a fresh tracker query (route-never-reimplement: work-compass is the tracker-access layer; pass it --with-provenance so its N-Tree carries these native fields, then project — never re-fetch here):

| Field | Native source work-compass surfaces | |---|---| | asked_by (who asked) | Jira reporter · GitHub issue author · Linear creator · a memory [[slug]] note | | asked_at (when) | issue created timestamp | | due (by-when) | Jira duedate · GitHub milestone due · Linear dueDate · sprint end |

  • Projection: over the PHASE-1 items, surface those where asked_by ≠ operator (a boss/cowork ask), sorted by due ascending, flagging overdue + due-soon.
  • tag/categorize (the missing verbs): for an item lacking native provenance (a brain-dump, a loose note), the operator may attach asked_by · asked_at · due · tags[] · category. In this phase these are PRINTED as a suggested command for approval (e.g. the Jira/Linear edit, or a one-line memory tag) — never auto-written. No new store: native fields first, a light [[slug]] memory tag only for items with no native home.
  • ⛔ Never fabricate a due or an asked_by that isn't in the source. An invented deadline is worse than none (anti-theater). "no due found" is a valid, correct value.

PHASE 4 — BRIEF (contract: morning-briefing)

Adopt the morning-briefing 7-section SitRep contract as the briefing shape — state · done · in-flight · blockers · decisions-awaiting · risks · next-action — populated from PHASES 1-3. At quick depth this is a formatting contract folded into PHASE 5; at --depth full it is a full morning-briefing pass.

PHASE 5 — PRESENT (one operator briefing — read-only, review-before-act)

Assemble ONE scannable operator briefing (per the host's end-of-action briefing protocol §7 — executive, recommended-first, risk-tagged):

  1. ▶ NEXT ACTION — the single highest-priority item + why (from PHASE 2's now quadrant).
  2. Eisenhower quadrants — Q1/Q2/Q3/Q4 counts + top items (from pulse/ops-strategist, NOT recomputed).
  3. ⏰ Who asked you / by when — the PHASE-3 people-ask view, overdue-first.
  4. 🧹 Loose ends — stale/orphan/pending from PHASE-1 + any un-dispositioned item, per the host's loose-end-triage-queue (Taxis) no-silent-drop discipline: each gets a disposition suggestion.
  5. Compass footer — one routing command per actionable node (delegate to the owning tool).

⛔ SANITIZE (unconditional): never echo a secret · credential · token · personal identifier into the briefing — reference by name/location. Read-only: any write/notify/schedule is PRINTED for the operator to run/approve, never executed by this skill.

Parameters

| Param | Default | Meaning | |---|---|---| | --mode | catch-up | catch-up · drift · who-owes-me · full. Explicit operator instruction wins. | | --depth | quick | quick (0-1-2-3-5) · full (adds ops-strategist 4-lens IF present + full morning-briefing). | | --scope | all | Which backlogs to gather — all · jira · github · linear · sessions (passthrough to work-compass). | | --who | — | Filter the people-ask view to asks from one person. | | --lang | auto | pt · en · auto (pt-BR for the operator, en-US for a machine consumer). | | --json | off | Emit a typed briefing envelope for an agent consumer instead of prose. |

Invariants (non-negotiable)

  1. Explicit operator instruction overrides every computed condition — mode, depth, scope, language.
  2. Route-never-reimplement — GATHER=work-compass, PRIORITIZE=pulse/ops-strategist, BRIEF=morning-briefing. This conductor authors ONLY the people-ask projection + tag/categorize + the unified assembly.
  3. Layer-purity — runs fully on the maos-resident core (work-compass+pulse+morning-briefing); ops-strategist is an OPTIONAL user-scope enhancement, never a hard dependency.
  4. Read-only by default — writes / notify / schedule are PRINTED for approval, never executed.
  5. On-demand only — no hooks, no always-on runtime. MAOS stays the sole conductor (proactive firing / calendar-ingest / notify-channel are a SEPARATE, operator-confirmed later phase — they touch settings).
  6. Never fabricate a due/asked_by/priority — "not found" is a valid value; SANITIZE is unconditional.
  7. Never computes its own Eisenhower — always routes to an existing prioritizer (anti-over-eng).
  8. Bounded — quick depth by default; do not convene ops-strategist to answer "what's next?".

Anti-patterns

  1. ❌ Re-aggregating — walking Jira/GitHub yourself instead of delegating work-compass.
  2. ❌ A 5th Eisenhower generator — computing priority here instead of routing to pulse/ops-strategist.
  3. ❌ Hard-depending on ops-strategist — it is user-scope; a community skill must not require it (Invariant 3).
  4. ❌ Auto-executing a write/notify/schedule — this front-door prints for approval; it never acts silently.
  5. ❌ Installing an always-on hook — that is a SEPARATE operator-confirmed phase, not this on-demand skill.
  6. ❌ Fabricating a deadline / who-asked — an invented due date is worse than none (anti-theater).
  7. ❌ Framework-routing — teaching which MAOS tool to use is maos-concierge's job, not this.

Quality tests (6 self-validity)

  1. Self-application — this conductor was itself designed by gathering the existing landscape (a recon), prioritizing the one genuine gap over the redundant, and presenting a single routed recommendation. ✅
  2. Non-contradiction — composes work-compass/pulse/morning-briefing without overriding any; distinct from maos-concierge (framework), reactivate (agent-facing twin), ops-strategist (the optional brain it routes to). ✅
  3. Survival — applied to itself it advocates route-don't-reimplement + read-only; it does exactly that. ✅
  4. Bounded-responsibility — quick default · route-only core · read-only · on-demand · qualitative sunset. ✅
  5. Explicit-exception — operator override (Invariant 1) · the not-use routing list · optional-ops-strategist. ✅
  6. Utility-sunset — below. ✅

Sunset (qualitative, not counter-based)

Deprecate when ANY: the host ships a native operator work-focus dashboard that subsumes this · ops-strategist is promoted community-resident AND grows a front-door making this redundant · the concierge/conductor family absorbs it into a unified entry · operator retraction · ≥3 false-fires (fired on a non-work-focus intent).

Refs

  • Composed (never reimplemented): skills/work-compass (GATHER) · skills/pulse (Eisenhower core) · skills/morning-briefing (SitRep contract).
  • Optional (user-scope, probed-not-required): ops-strategist (the 4-lens TPM+CoS+SRE+Agentic brain).
  • Siblings (distinct): skills/maos-concierge (framework-facing) · skills/reactivate (Entelecheia, incoming PR #280 — agent-facing twin) · skills/session-reentry (Anamnesis) · skills/directive-braindump-triage.
  • Governance: the host's end-of-action briefing protocol §7 (PRESENT format) · Cowork-Process-Topology §9 (the 7 CPT domains work-compass renders) · loose-end-triage-queue / Taxis (the no-silent-drop discipline behind PHASE-5 loose-ends) · layer-precedence-policy (the Layer-purity of Invariant 3).
  • Named by skills/anima per the host's naming authority; recorded in bin/artifact-registry.
  • Grounding: Xenophon, Oeconomicus (the oikonomos as household steward); Greek οἰκονόμος (LSJ).