The machine compounds. The closure layer still breaks.

Turns live meetings into a controlled loop: pre-read, agenda chase, post-call decisions, and follow-up routing.

• Skill: meeting-control
• MCPs: calendar, meeting transcript source, memory/brain search, team chat or tasks
• Sentry: meeting-readiness-sentry watches for missing agenda, missing pre-read, and missing post-call decisions

{
  "skill": "meeting-control",
  "trigger": "calendar.event.upcoming",
  "required_mcps": ["calendar", "transcripts", "memory", "tasks"],
  "sentry": "meeting-readiness-sentry",
  "rules": {
    "t_minus_24h": "request agenda and required inputs",
    "t_minus_30m": "escalate if materials missing",
    "t_minus_10m": "draft reschedule if still unprepared",
    "post_meeting": "extract decisions, owners, deadlines, resume brief"
  }
}

Converts important calls into a real review/send queue instead of leaving the promise stack in memory.

• Skill: post-call-followup
• MCPs: transcripts, CRM, Gmail, review queue, memory/brain search
• Sentry: promise-lag-sentry flags calls with explicit promises that still have no draft or send action by end of day

{
  "skill": "post-call-followup",
  "trigger": "transcript.created",
  "required_mcps": ["transcripts", "crm", "gmail", "review_queue", "memory"],
  "sentry": "promise-lag-sentry",
  "extract": ["pain_points", "objections", "buying_signals", "next_steps", "explicit_promises"],
  "actions": [
    "score fit and urgency",
    "draft follow-up",
    "queue for review",
    "update deal record",
    "alert if no send path exists by end of day"
  ]
}

If a code-based problem needs a quick fix, this is the kind of sentry that should already be running: unresolved errors, healthcheck status, and issue creation in one loop.

• Skill: incident-hotfix
• MCPs: GitHub, Sentry, cloud logs or healthcheck endpoint, deploy rail or shell access
• Sentry: incident-hotfix-sentry watches error spikes, failed healthchecks, and missing owner assignment

#!/usr/bin/env python3
import os, requests, subprocess, sys

SENTRY_ORG = os.environ["SENTRY_ORG"]
SENTRY_PROJECT = os.environ["SENTRY_PROJECT"]
SENTRY_TOKEN = os.environ["SENTRY_TOKEN"]
HEALTHCHECK_URL = os.environ["HEALTHCHECK_URL"]
GITHUB_REPO = os.environ["GITHUB_REPO"]

headers = {"Authorization": f"Bearer {SENTRY_TOKEN}"}
sentry_url = (
    f"https://sentry.io/api/0/projects/{SENTRY_ORG}/{SENTRY_PROJECT}/issues/"
    "?is=unresolved&sort=date&statsPeriod=1h"
)

health_ok = requests.get(HEALTHCHECK_URL, timeout=10).ok
issues = requests.get(sentry_url, headers=headers, timeout=15).json()
hot = [i for i in issues if i.get("count") and int(i["count"]) > 10]

if (not health_ok) or hot:
    title = "Hotfix watch: healthcheck failed or Sentry spike detected"
    body = f"health_ok={health_ok}\\nopen_hot_issues={len(hot)}"
    subprocess.run(
        ["gh", "issue", "create", "--repo", GITHUB_REPO, "--title", title, "--body", body],
        check=False,
    )
    print(body)
else:
    print("hotfix watch clear")

Turns Superwhisper and meeting transcripts into classified, searchable context so founder intent survives across models and sessions.

• Skill: founder-intent-router
• MCPs: Superwhisper transcript source, memory/brain search, vector memory, tasks or decision log
• Sentry: context-drift-sentry flags repeated manual re-explanations of the same context or missing transcript syncs

{
  "skill": "founder-intent-router",
  "trigger": "transcript.created",
  "required_mcps": ["superwhisper", "memory", "vector_memory", "decisions"],
  "sentry": "context-drift-sentry",
  "classify": ["decision", "question", "brainstorm", "delegation", "status_update"],
  "actions": [
    "write founder-intent brief",
    "store searchable summary in brain",
    "embed into Pinecone or vector layer",
    "link to active project or decision log",
    "surface relevant context on next model session"
  ]
}

This is the install behind the “copy-paste grinder” role: cluster repeat requests, draft replies, approve fast, and catch policy drift.

• Skill: support-approval-kernel
• MCPs: help desk, knowledge base, review queue, metrics surface
• Sentry: policy-drift-sentry flags answers that fall outside accepted templates or performance thresholds

{
  "skill": "support-approval-kernel",
  "trigger": "ticket.message.created",
  "required_mcps": ["helpdesk", "knowledge_base", "review_queue", "metrics"],
  "sentry": "policy-drift-sentry",
  "loop": [
    "cluster repeated issues",
    "draft response from approved sources",
    "approve/edit/skip in one surface",
    "promote accepted replies into templates"
  ]
}

Useful when people keep showing up without agenda items and you end up carrying the prep burden.

Before every scheduled meeting, remind agenda owners 24h before and 30m before start.
If required inputs are still missing 10 minutes before the call, draft a reschedule note and suggest moving the meeting.
After the meeting, extract decisions, owners, and follow-ups and route them into the right repo, issue, or CRM record.

Useful when Chrome keeps swallowing the first serious block of the day before build momentum lands.

When I say "start build mode," open Claude, Codex, Terminal, and the active repo.
Mute Slack notifications for 90 minutes.
Block Google Meet, app.slack.com, search, and webmaster surfaces until the block ends.
At the end of the block, ask me for one durable brief or shipped artifact.

Useful when interesting new framings keep reopening work that should already be in execution mode.

Before starting a new initiative, force a written rubric: objective, success criteria, hard constraints, kill conditions, and what will be ignored.
Write that rubric to the active repo, decision log, or project brief before agents start moving.
Once execution begins, do not reopen the rubric unless new evidence clearly beats the current path against the existing standard.
If a better idea appears mid-flight, tag it for V2 and keep shipping the current decision.

Useful when important calls create real insight, but follow-up, prioritization, and system routing lag behind.

When a discovery, customer, partner, recruiting, or investor transcript lands, extract pains, decisions, buying signals, next steps, and stakeholders.
Score whether the call showed urgent pain, a real promise, or a clear action path for the current pilot or offer.
If yes, draft the follow-up in the assigned sender's voice, queue it in the review surface, and update the relevant record.
If no, mark it low-priority and keep the transcript searchable for later pattern mining.

Useful when the machine keeps trying to widen strategy on a day your body is clearly asking for narrower work.

When today's Oura readiness is low or REM has been falling for 2+ nights, mark the day as a narrow day.
Do not open new GTM, product, or infrastructure lanes.
Prioritize closure, review/send, meeting prep, and grading existing work instead.
If I try to widen anyway, remind me this is a low-REM day and ask what outward loop gets closed first.

Useful when good calls turn into “I owe them something real” and then get outrun by system-building.

After every demo, discovery call, or investor call, extract every explicit promise I made.
Create one tracked item per promise with owner, due date, source transcript, and suggested next step.
If no draft or send action exists by end of day, queue a follow-up draft automatically and remind me before I start a new build lane.

Useful for the kind of role that spends hours copying comments or questions into AI and then manually approving the result.

If a support agent pastes the same kind of issue into AI over and over, log the repeated prompts, cluster the common request types, and generate a reply draft with an approval UI.
Approve, edit, or skip with one click.
Turn accepted replies into reusable templates over time.

Useful when the local machine is trying to be both the control plane and the worker box.

Use local machine for control and context only.
Push scrape, enrich, verify, browser automation, and bulk transformation work to AWS workers.
Return staged CSVs, CRM-ready rows, and a short operator brief when the run finishes.

Useful when your strategy lives in voice notes but the working shell only sees the last prompt you typed.

When Claude Code or Terminal opens, inject the last three voice notes plus today's top open loops from the shared brain and vector memory as hidden context.
Summarize the strategic intent in 5 bullets before taking the next action.
Write any new decisions back to the brain so the next model session does not need the same re-brief.
If the new task conflicts with the last voiced priority, ask for confirmation first.