Executive Assistant

Digital chief-of-staff on Google Workspace. Central orchestrator + specialized subagents + workflow skills + hooks that turn executive discipline into infrastructure.

Runs on both runtimes:

• Claude Code — config in .claude/settings.json, subagents in .claude/agents/
• Gemini CLI — config in .gemini/settings.json, native subagents in .gemini/agents/, policy isolation in .gemini/policies/ea-policies.toml

Skills, hooks, and state are shared.

Environment assumption: Google Workspace skills (gdocs, gdrive, gcalendar, gchat, gsheets, gslides) are already available in the operator's runtime (as a native skill, MCP server, or extension). The subagents/skills here only compose them.

Architecture — 30-second overview

ORQUESTRADOR (ea-orchestrator skill)
   ├─ SKILLS de workflow:    daily-brief, weekly-review, noise-cancel,
   │                         meeting-workflow, quarterly-review
   ├─ SUBAGENTS (8):         inbox-triager, meeting-prepper, meeting-debriefer,
   │                         project-router, project-tracker, commitment-tracker,
   │                         relationship-keeper, draft-composer
   ├─ SKILLS pré-existentes: gdocs · gdrive · gcalendar · gchat · gsheets · gslides
   └─ HOOKS (12 scripts):    SessionStart/UserPromptSubmit/PreToolUse/PostToolUse/
                             Stop/PreCompact/SessionEnd  (Claude Code)
                             SessionStart/BeforeAgent/BeforeToolSelection/BeforeTool/
                             AfterTool/AfterModel/PreCompress/SessionEnd  (Gemini CLI)

Full detail: ARCHITECTURE.md.

Quickstart by runtime

Running on Claude Code

1. Open a session in this directory.
2. Edit state/ea-state.json with your operator data.
3. Add projects in state/projects/<id>.yaml and update state/projects/_index.json.
4. Add key people in state/people/<id>.yaml.
5. Say "good morning" — the bootstrap.sh hook loads state and infers mode. For the weekly ritual, run /weekly-review.

Running on Gemini CLI

1. Sync skills (idempotent):

   ./scripts/sync-runtimes.sh

2. Start gemini in this directory. .gemini/settings.json automatically loads:
   • Hooks (same scripts as Claude Code, referenced via $GEMINI_PROJECT_DIR/.claude/hooks/...)
   • Subagents in .gemini/agents/ (8 hand-tuned with structured handoffs)
   • Agent overrides with maxTurns/maxTimeMinutes per subagent
3. For strong isolation, copy .gemini/policies/ea-policies.toml to ~/.gemini/policies/ (or point via Policy Engine setting). It restricts what each subagent can write.
4. Invoke a subagent explicitly: @meeting-prepper prepare 1:1 with Laiane at 10am or let automatic delegation decide.

Orchestration difference

	Claude Code	Gemini CLI
Subagent → subagent	✅ allowed	❌ forbidden (recursion guard)
Coordination	subagent or orchestrator	always orchestrator
Output pattern	free	handoffs[] structured mandatory

Files in .gemini/agents/ already reflect this pattern.

Directory structure

.
├── ARCHITECTURE.md
├── README.md
├── .claude/
│   ├── settings.json              # wiring de hooks + permissões
│   ├── skills/
│   │   ├── ea-orchestrator/SKILL.md
│   │   ├── daily-brief/SKILL.md
│   │   ├── weekly-review/SKILL.md
│   │   ├── noise-cancel/SKILL.md
│   │   ├── meeting-workflow/SKILL.md
│   │   └── quarterly-review/SKILL.md
│   ├── agents/
│   │   ├── inbox-triager.md
│   │   ├── meeting-prepper.md
│   │   ├── meeting-debriefer.md
│   │   ├── project-router.md
│   │   ├── project-tracker.md
│   │   ├── commitment-tracker.md
│   │   ├── relationship-keeper.md
│   │   └── draft-composer.md
│   └── hooks/
│       ├── lib/common.sh
│       ├── bootstrap.sh             # SessionStart
│       ├── ritual-check.sh          # SessionStart
│       ├── mode-context.sh          # UserPromptSubmit
│       ├── filter-skills-by-mode.sh # PreToolUse
│       ├── check-pending-debriefs.sh# PreToolUse (meeting-prepper)
│       ├── enter-review-mode.sh     # PreToolUse (weekly/quarterly)
│       ├── touch-projects.sh        # PostToolUse
│       ├── scan-commitments.sh      # PostToolUse
│       ├── promise-detector.sh      # Stop
│       ├── project-mention-tracker.sh # Stop
│       ├── preserve-crm.sh          # PreCompact
│       └── eod-snapshot.sh          # SessionEnd
├── state/
│   ├── ea-state.json
│   ├── projects/
│   │   └── _index.json
│   ├── people/
│   ├── commitments/
│   │   ├── made-by-operator.json
│   │   ├── made-to-operator.json
│   │   └── implicit.json
│   └── rituals/
│       ├── daily/
│       ├── weekly/
│       ├── quarterly/
│       └── meetings/
└── templates/
    ├── ea-state.template.json
    ├── project.template.yaml
    ├── person.template.yaml
    └── commitment.template.json

Quickstart

1. Edit state/ea-state.json with your operator data (name, role, current focus). The template ships with sensible defaults.

2. Add active projects by copying templates/project.template.yaml to state/projects/<id>.yaml and updating state/projects/_index.json.

3. Add key people (manager, direct reports, mentor, ~10 priority people) by copying templates/person.template.yaml to state/people/<id>.yaml.

4. Check hooks: chmod +x .claude/hooks/*.sh (already done on setup).

5. Start a Claude Code session in this directory. The bootstrap.sh hook loads your state and infers the mode. If it's morning, ask for /daily-brief or just say "good morning".

6. Friday or Saturday: run /weekly-review. Hooks lock other skills until the ritual completes.

How the loop works

Morning

06:30  SessionStart
   → bootstrap.sh        loads ea-state.json, detects hour < 11
                         and last_morning_brief != today → mode = morning_brief
   → ritual-check.sh     checks next_weekly_due

06:31  Operator: "good morning"
   → UserPromptSubmit/mode-context.sh injects:
     "Mode morning_brief. Recommended skill: daily-brief."

06:31  ea-orchestrator calls Skill:daily-brief
   → daily-brief uses gcalendar to list agenda
   → delegates to inbox-triager (subagent) for email classification
   → delegates to noise-cancel for auto-archive
   → produces 1-page brief in state/rituals/daily/2026-05-02.md
   → mode → active_day

Before a meeting

09:55  Calendar webhook (T-30min meeting "1:1 Laiane")
   → ea-orchestrator detects proximity → mode = meeting_prep
   → invokes Skill:meeting-workflow phase=prep
   → hook check-pending-debriefs.sh:
     - checks if there's a prep without a debrief from previous meetings
     - if so: BLOCKS. operator completes debrief first.
   → meeting-prepper subagent generates prep doc (5 sections)
   → saves to state/rituals/meetings/<event>-prep.md + gdocs

After a meeting

10:35  Operator pastes raw notes
   → ea-orchestrator → meeting-workflow phase=debrief
   → meeting-debriefer extracts DECISIONS, ACTIONS, IMPLICITS
   → commitment-tracker registers commitments (asking first for implicits)
   → project-router routes actions to projects
   → project-tracker updates next_action / blockers / decisions
   → relationship-keeper updates last_contact / threads

Weekly review (Fri/Sat)

SessionStart Friday
   → ritual-check.sh detects next_weekly_due ≤ today
   → injects suggestion: "run /weekly-review before anything else"

Operator: /weekly-review
   → enter-review-mode.sh sets mode = weekly_review
   → filter-skills-by-mode.sh locks all other skills
   → weekly-review skill conducts 7 phases in dialogue
   → each phase updates state
   → at end: rituals.next_weekly_due = +7d, mode → active_day

Context compaction (long sessions)

context fills → Claude Code triggers PreCompact
   → preserve-crm.sh saves snapshot of top-3, active projects, commitments
     in state/.snapshots/precompact-<ts>.md
   → injects: "After compaction, re-read this file and ea-state.json"

End of day

SessionEnd
   → eod-snapshot.sh generates state/rituals/daily/<date>-eod.md
   → resets daily counters
   → mode → active_day

Operational principles

Principle	Meaning
Delegate, don't duplicate	Orchestrator coordinates; subagents/skills analyze.
External state is source of truth	Model memory is volatile. Re-read before acting.
Rituals are not optional	Hooks make skipping them impossible.
Nothing stays unrouted	Every signal goes into a project, person, or explicit backlog.
Every skill produces a decision or state change	Summary is overhead. State change is leverage.
Continuous calibration	Router, noise filter, and ritual evolve through feedback.
Depth > speed	Use multiple turns. Shallow analysis is an orchestration failure.
Resumable by design	File-based state makes the system robust to interruptions.

Customizing

• Mode lock: edit filter-skills-by-mode.sh to add new locked modes or soften existing locks.
• Noise filters: start by editing state/ea-state.json :: noise_filters. The noise-cancel skill learns and proposes adjustments in phase 7 of the weekly review.
• People cadence: declare in state/people/<id>.yaml :: cadence to receive alerts when the contact rhythm is violated.
• Webhooks: automatic triggering of meeting_prep/meeting_debrief based on the calendar is an external integration (Cloud Function that invokes the orchestrator). Without it, the operator calls /meeting-prep <event_id> manually.

What is NOT here

• Implementation of Google Workspace skills — they are pre-existing in the environment.
• Cloud Function webhook to auto-trigger meeting phases.
• Dashboards/visualization — text-focused. Add a visual-explainer skill if you want HTML.
• Advanced self-healing for failing skills — a future extension over PostToolUse with tool_response.error.

License

MIT — see LICENSE.