PHASE 07 · OPERATE

Operate as a Director — Daily Cadence

A dependable workflow needs named ownership, routine review, visible exceptions, and an incident process. Use the examples below to design a cadence that fits the task and its risk.

⏱ Working session · adapt and review Operations Identity-shaping

What a Director's day actually looks like

If you're spending most of your day inside the same tools as before — writing the captions, building the reports, filling the spreadsheets — you've built assistants but you're still operating as an executor. The cadence below is what shifts that.

1Queue review. At the risk-based SLA, inspect pending drafts, evidence, failures, and escalations; approve only the exact version within scope.
2Direct & dispatch (60 min). Issue the day's briefs. Spawn new tasks. Hand work between assistants where needed.
3Judgment work (2–3 hrs). The actual reason you exist as a director: strategy, calls, relationships, hard decisions, the J-tasks AI can't touch.
4Change debrief. After a material decision or incident, reconcile what executed, update the audit record, and schedule any regression test.

Measure the actual review and coordination burden against your baseline. If oversight consumes the expected benefit, reduce scope, improve the workflow, or retire it; do not assume the example schedule will fit every role.

The morning brief ritual

A short, regular review can help surface failures and pending approvals. Choose the frequency from task volume, consequence, and service-level needs.

  1. Open the director dashboard (a single doc / Notion page that lists all assistants and their pending queues).
  2. Walk through each assistant: any outputs waiting? Any failed runs? Any escalations?
  3. Approve what's good, send back what isn't (with one-sentence feedback), defer what needs your judgment.
  4. Note which assistant needs a training entry today.
  5. Close the dashboard. Move to the next block.

Human oversight must be specified, staffed, and tested. Review cadence follows exposure, failure latency, and SLA; monitoring can also fail, so include missed-alert tests, escalation, stop, reconciliation, incident response, and rollback.

The director dashboard

You don't need fancy tooling. A single markdown file or Notion page works perfectly. Here's a starter:

# Director Dashboard Last reviewed: 2026-05-04 09:15 ## ACTIVE ASSISTANTS ### 🎨 Maya · Caption Writer - Status: 🟢 accepted for drafting (evaluation v0.7 passed; human approval required) - Pending: 2 captions for Tuesday launch - Last training entry: 2026-05-02 - Next review: weekly Friday ### 📊 Iris · Performance Analyst - Status: 🟡 watch (over-cautious, asking too many questions) - Pending: weekly cohort report - Open issue: tighten escalation rule (Phase 6 pattern) ### ⚡ Otto · Meta Operator - Status: [PASS / INVESTIGATE / STOPPED] - Pending: [COUNT] — last verified [TIMESTAMP + TIMEZONE] - Last training entry: 2026-04-29 ## QUEUE - [ ] Brief Maya on the May launch (3 product captions) - [ ] Re-train Iris's escalation rule - [ ] Build a new assistant: "Reese" for competitor research ## ESCALATIONS (need my decision) - 1 ad set flagged by Otto: investigate the dated baseline, lag, and change log; owner decides whether to test, hold, or leave unchanged ## TRAINING DEBT - [OWNER]: review the risk-ranked pending drafts and record disposition.

This dashboard is your bridge. It exists so that when you sit down at 9am, you don't lose 20 minutes figuring out where you left off. Update it, don't try to remember it.

Inter-assistant handoffs

Some workflows involve multiple stages. For example: Reese (researcher) compiles permitted, cited information → Maya (writer) drafts captions → a named human reviews them → Otto (operator) creates a draft or paused item within enforced scope.

Three patterns for handoffs, in order of complexity:

① Manual handoff (Director ferries)

A human verifies provenance, authorization and the schema, then transfers only synthetic or minimized approved fields into the next isolated stage. Treat all model output as untrusted input; validate structure and meaning. This still has security and review work and is not “zero engineering.”

Start here
② Shared file handoff

Assistant A writes a schema-validated artifact to a tenant-isolated handoff location. The next stage receives only authorized fields; a human can inspect the artifact before continuing.

After the manual boundary is understood
③ Automated chain

An orchestrator can trigger later stages only through a policy-enforcing service with least-privilege identities, idempotency, caps, logs, stop conditions, and required approvals.

Only after end-to-end evaluation

Start with an inspectable handoff, document trust boundaries and failure modes, then evaluate the complete chain. A successful manual run is useful evidence, not proof that automation is safe.

Performance reviews — weekly

At a cadence justified by sample maturity and risk, review each bounded workflow's performance, incidents, reviewer effort, and unresolved uncertainty. This is a system review, not a human performance 1:1.

Weekly Assistant Review

The last item on that checklist is the killer one: "any tasks I did this week that should have gone to this assistant?" Catch yourself executing what you should be directing. That's where the leverage gets lost.

Reporting back to yourself

You won't notice your own progress unless you measure it. Once a month, write a one-page report — to yourself — answering:

Save these reports with evaluation results and incident logs. Over time they can show trends, although attribution still requires a baseline and a fair comparison.

Time allocation — compare with your baseline

Record a representative baseline before changing the workflow, then compare total time: execution, review, correction, incident response, and maintenance. The right mix varies by role and risk.

Baseline

Measure the current task volume, accepted quality, cycle time, cost, and correction burden before adding an assistant.

Pilot

Measure drafting, review, rework, failures, and maintenance on a bounded sample. Compare like-for-like cases with the baseline.

Accepted operation

Keep monitoring drift and incidents. Expand only when the measured benefit remains positive and the owner accepts residual risk.

The director's calendar — example

Here's what a real director week looks like once the cadence is set. Adjust to your timezone and energy.

MON–FRI 09:00–09:15 ☕ Morning brief (review assistant outputs, dispatch) 09:15–10:30 🎯 Judgment work (calls, strategy, hard decisions) 10:30–11:00 ⚡ Direct & brief assistants (give the day's tasks) 11:00–13:00 🎯 Judgment work block 13:00–14:00 🍽 Lunch 14:00–16:00 🎯 Judgment work block 16:00–16:30 🛠 Build / improve (1 small assistant upgrade) 16:30–17:00 📋 EOD debrief (review what shipped, log training, update dashboard) FRI 17:00–17:30 🔍 Weekly review (all active assistants) SAT (optional, 30 min) 📈 Monthly report (first Saturday of month)

What to do when an assistant fails publicly

It will happen. An assistant publishes a typo. An operator pauses the wrong campaign. A caption goes out with the wrong product name.

Step 1: Pause the workflow, revoke or narrow write access if needed, and follow the relevant incident and notification procedure.

Step 2: Preserve authorized evidence: exact output, inputs, configuration/model/tool versions, approvals, logs, timestamps, and external changes.

Step 3: Contain and correct the impact through the accountable owner; communicate or notify affected people when policy or law requires it.

Step 4: Determine contributing causes across instructions, data, tools, permissions, validation, review, and operating process. Do not assume a prompt edit is sufficient.

Step 5: Add a regression case, test the proposed control, document residual risk, and have the owner approve recovery.

Step 6: Restore the narrowest safe scope gradually, or retire the workflow. Keep a tested kill switch and rollback path.

Your task before Phase 8

Set up your operating cadence — for real, not in theory.

  1. Create your director-dashboard.md file. Add your active assistants.
  2. Block your calendar: morning brief, EOD debrief, Friday weekly review.
  3. Run the cadence for one full week. Don't skip the rituals even if "nothing's pending."
  4. At the end of the week, write your first monthly-style report (just for the week — get the muscle).

At the end, review whether the cadence exposed exceptions promptly without creating disproportionate overhead. Adjust from that evidence.

Recap

Last phase next. Scale, the 30/60/90 plan, and what to build past your first solo team.