ai-coaching-podcast

name: ai-coaching-podcast description: Daily 25-30 min AI coaching podcast. Combines psychology, goal tracking, conversation analysis, book concepts, and industry news into a spoken deep-dive. Generated as audio via ElevenLabs.

Daily Audio Coaching Podcast

Overview

Every morning, generate a ~30-minute audio coaching session. The user listens during their commute or exercise. The session is a deep-dive written as a spoken monologue and converted to audio via ElevenLabs.

Prerequisites

ElevenLabs API key (set in ELEVENLABS_API_KEY env var or config)
ffmpeg installed (brew install ffmpeg)
Transcript source (e.g., wear-transcribe skill, Omi, or any daily transcript)
Filled-in reference files (see references/ folder)

Voice & Model

Configure in scripts/config.json:

Voice: Choose from ElevenLabs voice library
Model: eleven_multilingual_v2 recommended (most lifelike)
Generator script: scripts/generate-audio.py

The Coaching Team

Each session draws from multiple perspectives:

The Psychologist

Draws from: references/coaching-profile.md (Triggers / Context sections)
Understands: attachment patterns, defenses, cognitive distortions, growth areas
Role: connects today's topic to deep patterns. Challenges surface-level explanations. Validates genuine progress.

The Coach

Draws from: references/coaching-profile.md (copy from template), references/goals.md (copy from template)
Tracks: your goal pillars with RAG scoring (🟢🟡🔴)
Role: practical actions, micro-commitments, accountability, progress tracking

The Industry Researcher

Sources: news, Slack, industry feeds, competitor updates
Role: hyper-relevant intelligence filtered through what matters for YOUR work

The Personal Assistant

Draws from: calendar, tasks, messages, transcripts
Role: grounds the session in today's reality — meetings ahead, people to prep for, tasks due

Data Sync (CRITICAL — do this BEFORE writing the script)

Before generating content, sync ALL available sources:

Transcripts — yesterday's diarized transcripts (from wear-transcribe or equivalent)
Calendar — today's and tomorrow's events
Tasks — overdue, due today, due this week
Messages — recent WhatsApp, Slack, email highlights
Previous scorecards — last evening scorecard, last weekly review

Never write a coaching script without fresh data. The power of this system is that it knows what actually happened — generic advice is worthless.

Script Writing Guide

Format

~6000-7000 words = ~35-42 minutes of spoken audio
Written as spoken monologue, NOT a document
Direct address: "you", "your"
Conversational but substantive — not dumbed down, not academic

Structure (flexible, not rigid)

Opening (30 sec) — "Good morning. Today I want to talk about..." Ground it in something specific from yesterday.
Conversation Analysis (5-8 min) — Pick 2-3 moments from yesterday's transcripts. Reframe them. What was really happening beneath the surface? What patterns show up?
Goal Coaching (5-8 min) — RAG scorecard for each goal pillar. Evidence from yesterday. What to incorporate today. Be specific.
People Prep (3-5 min) — Who are you meeting today? What's the dynamic? What's the one thing to remember going in?
Book Deep-Dive (5-8 min) — One concept from your reading list, applied to something that happened yesterday or is coming up today. Not a summary — teach ONE idea deeply with stories and tie it to your specific situation.
Industry Update (3-5 min) — What happened in your industry that matters. Not a news roundup — filtered through your strategic context.
Micro-commitment (2 min) — One specific thing to do today. Not "be more productive" — "Before the 10am standup, write down one thing you shipped yesterday and say it first."
Close (30 sec) — Warm, grounding.

Tone — Like a Book Chapter

The audio should feel like listening to a chapter from a well-written book. NOT a pep talk. NOT bullet points read aloud.

Depth over breadth — one theme explored richly
Real references — cite psychology research, name frameworks, reference case studies
Stories — other people who faced similar dynamics, historical figures, research subjects
Book concepts woven in naturally — not "as the author says in chapter 4" but the concept applied to a specific situation

Verification Rules

Every person mentioned must appear in yesterday's transcripts, calendar, or messages
Every quote must be verified against source data
Every meeting must be confirmed on the calendar
Never fabricate conversations or attribute statements that didn't happen

Audio Generation

After writing the script, save it to a text file and run:

python3 scripts/generate-audio.py script.txt [output-name]

The script handles:

Chunking text at 4500 chars (ElevenLabs API limit)
Generating audio for each chunk
Concatenating with ffmpeg
Output as MP3

Cron Configuration

Example OpenClaw cron (7am daily, weekdays):

{
  "name": "Daily Audio Coaching",
  "schedule": { "kind": "cron", "expr": "0 7 * * 1-5", "tz": "YOUR_TIMEZONE" },
  "payload": {
    "kind": "agentTurn",
    "message": "Read and follow skills/ai-coaching-podcast/SKILL.md",
    "timeoutSeconds": 600
  }
}

Scorecard Format

For evening scorecards (optional companion to the podcast):

═══ Evening Scorecard — [DATE] ═══

GOAL 1: [Name] — 🟢/🟡/🔴
• Evidence: [specific thing from today]
• Evidence: [another specific thing]

GOAL 2: [Name] — 🟢/🟡/🔴
• Evidence: ...

═══ Pre-commits for tomorrow ═══
1. [Specific, measurable action]

Topic Selection

Before writing, read references/topic-log.md to avoid repetition.

Rules:

Never repeat a framework within 7 days
Never tell the same story twice in 14 days
Rotate categories — don't do 3 of the same type in a row
Book concepts: one per day, applied specifically. Not a summary.
Log every session after generating

Pick the topic based on:

What happened yesterday (transcripts are the primary signal)
What's coming today (calendar)
Where the 90-day plan says you should be this week
What hasn't been covered recently (check the log)

What Makes a GOOD Session

Opens with something specific from yesterday — a real moment, a real conversation
Makes you think "how did it know that?" — because it read the transcripts
Teaches ONE concept deeply, applied to YOUR situation
Ends with a micro-commitment you can actually do today
Feels like a friend who knows you well, not a coach reading from a script

What Makes a BAD Session

Generic advice that could apply to anyone ("remember to be present today")
Repeating the same framework from 3 days ago
Mentioning people or meetings that didn't happen (hallucination)
Reading a news roundup with no connection to your work
Bullet points read aloud instead of a flowing narrative
Motivational pep talk energy ("You've got this!")

On-Demand Evening Sessions

When triggered manually (e.g., "reflect" or "evening podcast"):

Use the evening scorecard data — don't re-sync everything
Shorter: 10-15 min (vs 30 min morning)
Focus on processing the day, not prepping for tomorrow
More reflective, less prescriptive
Speed matters — generate quickly, user listens on the way home

Tracking

Topic Log (`references/topic-log.md`)

After every session, append a row with date, category, and key topics covered. This prevents repetition and ensures balanced coverage.

90-Day Plan (`references/90-day-plan.md`)

Check the current week's targets. The podcast should reference what the user should be practicing THIS week, not generic goals.

Adaptive Coaching

If a goal scores 🔴 for 2+ weeks → escalate in next audio, increase focus
If a goal scores 🟢 for 2+ weeks → de-escalate, shift attention elsewhere
If user says "less" on any topic → reduce and note the preference
Vary the structure and format to keep it fresh — not every day needs the same order