daily-coach

name	daily-coach
description	Generate a personalized daily health coaching message by dispatching 10 specialist subagents that each review the user's data from their domain expertise. Each subagent delivers its own message as a separate Telegram bubble.
user-invocable	false

Daily Coach

Use this skill when:

• a scheduled daily coaching message needs to be generated
• the user explicitly asks for the cron-generated daily health coaching behavior

Step 1: Gather context

Call both tools to build the shared context payload:

1. Call the coaching_context tool:

{ }

2. Call the nutrition tool for weekly summary:

{
  "command": "weekly_summary",
  "end_date": "YYYY-MM-DD"
}

If insufficient_data is true: skip subagent dispatch. Instead say what is missing and what to log next.

Step 2: Dispatch 10 specialist subagents

Read each agent prompt file from {baseDir}/../../agents/ and spawn all 10 in parallel using sessions_spawn.

Agent Registry

#	File	Role	Emoji
1	imperial-physician.md	Orchestrator — #1 priority for today	🏥
2	diet-physician.md	Nutrition — meals, macros, micros	🍚
3	movement-master.md	Exercise — strain, training load	🏃
4	pulse-reader.md	Body metrics — RHR, HRV, SpO2	💓
5	formula-tester.md	Cross-domain pattern detection	🧪
6	herbalist.md	Supplement considerations	🌿
7	trial-monitor.md	Experiment status + compliance	📋
8	court-magistrate.md	Trial design candidates	⚖️
9	medical-censor.md	Safety flags + warnings	🛡️
10	court-scribe.md	Relevant news + literature	📜

Dispatch protocol

For each agent in the registry:

1. Read the agent prompt: read("{baseDir}/../../agents/{file}")
2. Construct the task:

{contents of the agent .md file}

---

TODAY'S CONTEXT:
{paste the full JSON context payload from Step 1}

WEEKLY NUTRITION:
{paste the weekly summary JSON from Step 1}

3. Spawn: sessions_spawn(task=<constructed task>, label=<role name>)

Spawn ALL 10 in parallel. Each subagent runs independently and announces its result back as a separate message.

Step 3: No assembly needed

Each subagent announces directly to the chat channel as a separate Telegram bubble. They arrive as each finishes. The main agent does NOT need to collect or reformat the results.

After all 10 have announced, if checkin_needed is true, send one final message prompting the user to log their experiment check-in.

Rules

• Reply in the user's language if obvious from profile context. Otherwise English.
• Each subagent produces 2-3 sentences starting with [Role Emoji].
• Recommendations must be conservative, lifestyle-only, grounded in the user's own data.
• Do not overclaim from sparse data. Agents should say "insufficient data" when appropriate.
• Subagents should NOT repeat each other's recommendations — each owns their domain.

OpenClaw config requirements

The install script (step 6) configures agents.defaults.subagents.maxChildrenPerAgent and maxConcurrent to 10 in ~/.openclaw/openclaw.json. Without this, only 5 of the 10 specialists will spawn.