Founder Chief of Staff orchestrator across 10 C-suite role agents

A single entry-point Chief of Staff agent sits between the founder and 10 C-suite role agents (CEO, CTO, COO, CPO, CMO, CFO, CRO, CISO, CHRO, Executive Mentor). For every question it: loads company context, scores decision complexity 1–5, routes to the right role(s) via a topic-keyed matrix, runs a board-meeting protocol for the heaviest decisions, synthesizes the roles' outputs into consensus + named conflicts + owned action items + one decision point, and writes the decision to a persistent log with a review date. Loop-prevention rules are load-bearing: the Chief of Staff cannot invoke itself, max recursion depth is 2, an A→B→A chain is blocked, and a board meeting counts as depth 1.

Naive role-prompting has two failure modes: role agents invoking other role agents until advice goes circular or the call chain runs away, and output that is diffuse rather than decision-shaped (no owner, no deadline, no logged decision). This framework turns "ask the AI for business advice" into a routed, auditable protocol where conflicts between functional concerns are named rather than smoothed over, every action has an owner and deadline, and every decision is logged with a review date the agent will resurface unprompted. The complexity score keeps a one-domain question from triggering a five-role board meeting, which is where most of the wasted tokens and circular advice come from.

1. Install and bootstrap context. curl -fsSL https://raw.githubusercontent.com/ricardonevesbraga/flowgrammers-skills/main/install.sh | bash, then run the onboarding skill once to capture company context (reloaded on every Chief of Staff invocation).
2. Make the Chief of Staff the single entry point. Load the skill; from then on every C-suite question starts here, never a direct role invocation.
3. Score decision complexity 1–5. 1–2 = single domain → 1 role; 3 = two domains → 2 roles + synthesis; 4–5 = 3+ domains or irreversible → full board meeting.
4. Route via the topic-keyed matrix. Fundraising/burn/financial → CFO+CEO; hiring/culture → CHRO+COO; roadmap → CPO+CTO; architecture → CTO+CPO; revenue/sales/GTM → CRO+CFO; OKRs/execution → COO+CFO; security → CISO+COO; investors → CEO+Board; market/positioning → CMO+CRO; M&A/pivots → CEO+Board.
5. Invoke roles with the fixed [INVOKE:role|question] syntax. Score 1–2 → one role. Score 3 → two roles then synthesize. Score 4–5 → run the board-meeting protocol block (topic · participants · 2–3 agenda questions · one invoke per role · synthesis), max 5 roles, each speaking once, no back-and-forth.
6. Synthesize in the fixed 4-section format: consensus themes from 2+ roles → conflicts named explicitly without smoothing → action items with owner + deadline (max 5) → one decision point framed as two options with trade-offs and no recommendation.
7. Auto-log the decision (date, original question, decision, owner, review date). At session start the agent checks for overdue reviews and surfaces them ("You decided X on [date]. Worth a check-in?").
8. Enforce quality gates before any output: conclusion first, context loaded, every finding has What+Why+How, actions have owners and deadlines, decisions framed as trade-off options, conflicts named not smoothed, risks concrete (if X→Y, costs $Z), max 5 bullets per section.

See also: A-044 (single-founder inbox/calendar orchestrator). A-044 is one founder, one assistant over personal ops; this case is multi-role C-suite orchestration with routing and a decision log — distinct shapes, complementary.

Invocation syntax (verbatim):

[INVOKE:role|question]
[INVOKE:cfo|What's the right runway target given our growth rate?]
[INVOKE:board|Should we raise a bridge or cut to profitability?]

Board-meeting protocol (verbatim — PT-BR template strings kept as-is):

REUNIÃO DE CONSELHO: [Tópico]
Participantes: [Papéis]
Pauta: [2–3 perguntas específicas]
[INVOKE:role1|agenda question]
[INVOKE:role2|agenda question]
[INVOKE:role3|agenda question]
[Síntese do Chief of Staff]

Synthesis output format (verbatim — PT-BR section headers kept as-is). In English the four sections are: What we agree on (2–3 consensus themes) · The disagreement (named conflict + each side's reasoning + what's actually at stake) · Recommended actions (max 5, each with owner and deadline) · Your decision point (one question, two options with trade-offs, no recommendation — clarity only):

## No Que Concordamos
[2–3 temas de consenso]
## O Desacordo
[Conflito nomeado + raciocínio de cada lado + o que realmente está em jogo]
## Ações Recomendadas
1. [Ação] — [Responsável] — [Prazo]
## Seu Ponto de Decisão
[Uma pergunta. Duas opções com trade-offs. Sem recomendação — apenas clareza.]

Decision-log entry (verbatim — PT-BR template strings kept as-is; fields are Decision / Date / Question / Decided / Owner / Review):

## Decisão: [Nome]
Data: [AAAA-MM-DD]
Pergunta: [Pergunta original]
Decidido: [O que foi decidido]
Responsável: [Quem executa]
Revisão: [Quando verificar de volta]

• Loop prevention is load-bearing, not optional. Without the four rules (Chief of Staff can't self-invoke; max depth 2; A→B→A blocked; board meeting = depth 1) role skills invoke each other into circular advice or runaway call chains. This is the single failure that breaks naive role-prompting.
• Conflicts are shown, not resolved — the founder decides. The agent must not pretend to reconcile a genuine disagreement between functional concerns (e.g. CFO vs CPO on runway vs roadmap); it names the conflict and what's at stake.
• Don't let complexity inflation trigger a board meeting for a one-domain question. The 1–5 score exists to keep a simple CFO question from spawning a five-role meeting; over-scoring is where the token waste and circular advice come from.
• No preamble; conclusion first. Output that buries the decision under context fails the quality gate.