Coach Daily Operating System

🔵 Accountability 🟢 Habit tracking 🟡 Next-day planning 🟣 Projects & priorities

The intent

Turn the coach from something that logs your day into something that runs it — a closed daily loop of phone calls that holds you to revenue-driving work, tracks habits without the perfectionism spiral, plans tomorrow, and always knows your single #1 lever.

The core pattern it targets: you build and optimize systems as a comfortable avoidance of the scarier revenue work — because building never gives a verdict and selling does. Every workstream below bends toward closing that gap, humanely.

The daily loop

One persistent call-loop is the spine — it delivers accountability, habit logging, and planning in the same four conversations.

A day, end to end

Morning launch→ Midday accountability→ Lunch reset→ Evening plan-tomorrow

Calls at your stated wake-time and dynamic callbacks; relentless within humane rails (quiet 07–23, escalating retries, opt-out, daily caps).

Morning — win the first hour

live shower companion→ MacBook = task #1 only→ meditation timer→ sets midday callback

"How long will the first tasks take?" → that's when the coach calls back.

Six workstreams W0–W5

Built in order — protect the live system first, then layer capability. Live-path changes ship behind the E2E gate; markdown skills sync with no deploy.

W0 · Voice-credit resilience (prerequisite)

Kill the "calls go silent when the Anthropic balance hits $0" outage — a /health balance check + alert and a non-thinking Gemini auto-failover. The loop is moot if the line can die.

W1 · Persistent humane call-loop 🔵

Dynamic callbacks (schedule_callback, set_wake_time), new midday/lunch/evening_close kinds, escalating retry ladder, hard humane caps, atomic+locked state, sparring tone, live shower companion.

W2 · Habit tracking + scoring 🟢

Keep all 33 fields (additive). Code day-score: 60%+ = WIN, 1-min = full credit, never-miss-twice. Overlay 3 keystone pillars + auto-detect the objective ones; low-friction one-question-at-a-time logging.

W3 · Next-day planning + rituals 🟡

Evening plan-tomorrow (MITs, frog-first, close loops + tabs, if-then, wake-time, next-session brief); morning launch; afternoon NSDR reset. All markdown-skill edits.

W4 · Projects & priorities overview 🟣

Registry schema upgrade; coach computes a leverage score and proposes your #1, you ratify weekly; WIP-limit 3 + an explicit "parked — avoid" list; progressive-disclosure surfacing.

W5 · Memory · capabilities · security

Three editable memory files (AGENT/MATT_PROFILE/MEMORY); state-diffing heartbeat; "how I work" explainer; Notion safe-write; sanitize security baseline; per-utterance transcript + voice-matched summary.

The 22 locked decisions 2 interview rounds

Each was web-grounded; these are settled — Instance 2 does not relitigate them.

#	Area	Locked choice
Q1	Build order	W0 voice-credit resilience first, then W1 → W5
Q2	Call cadence	Escalating retry ladder (+5/+10/+15/+30), Telegram fallback after ~5 tries — not fixed 5-min
Q3	Coach tone	Supportive by default; firm "tough-love" only when you're clearly avoiding
Q4	Morning shower	Coach stays live on the line (body-doubling)
Q5 · Q15	Memory	Three editable files — AGENT (soul) / MATT_PROFILE (you) / MEMORY (decisions + active projects), loaded every prompt, promoted nightly
Q6	#1 lever	Coach computes a leverage score, proposes #1, you ratify weekly
Q7	WIP limit	3, soft — names violations + the "parked — avoid" list
Q8	Transcript	Fix per-utterance capture; send summary + a separate transcript link
Q9 · Q22	Notion + autonomy	Read + safe-write behind a confirm gate; Advisor→Partner (low-risk auto-actions freely, ask before external/irreversible)
Q10	Self-knowledge	Capabilities summary + a plain-language "how I work / where your data lives" explainer
Q11	Timing	Lunch proactive; evening 18:00 soft → 20:00 retries, 21:00 hard anchor
Q12	Scoring	In code (testable): 60%+ = win, 1-min = full credit, never-miss-twice; celebration tone in the skill
Q13	Encoding	Behavior = hard-rails in AGENT.md/SKILLs (no deploy); code only for mechanics — not the directives store
Q14	Definition of done	Text E2E green + one live phone-call smoke test per new call-kind
Q16	Heartbeat	Data-driven state-diffing — off-schedule nudge when the delta warrants, stay silent when nothing changed
Q17	Habit pillars	Overlay 3 keystone pillars + auto-detect objective ones; never auto-check personal (kept additive)
Q18	Persona guard	AGENT.md edits require an explicit confirmation (anti soul-drift)
Q19	Security	Sanitize + trust-boundary now; add the guardrail-agent the moment external/3rd-party content is read
Q20	State robustness	Atomic writes + file-locking on state.json (cron + webhook write it concurrently)
Q21	Voice	Voice-match summaries/drafts via RAG over your past messages/journal

Area

Locked choice

Build order

W0 voice-credit resilience first, then W1 → W5

Call cadence

Escalating retry ladder (+5/+10/+15/+30), Telegram fallback after ~5 tries — not fixed 5-min

Coach tone

Supportive by default; firm "tough-love" only when you're clearly avoiding

Morning shower

Coach stays live on the line (body-doubling)

Q5 · Q15

Memory

Three editable files — AGENT (soul) / MATT_PROFILE (you) / MEMORY (decisions + active projects), loaded every prompt, promoted nightly

#1 lever

Coach computes a leverage score, proposes #1, you ratify weekly

WIP limit

3, soft — names violations + the "parked — avoid" list

Transcript

Fix per-utterance capture; send summary + a separate transcript link

Q9 · Q22

Notion + autonomy

Read + safe-write behind a confirm gate; Advisor→Partner (low-risk auto-actions freely, ask before external/irreversible)

Q10

Self-knowledge

Capabilities summary + a plain-language "how I work / where your data lives" explainer

Q11

Timing

Lunch proactive; evening 18:00 soft → 20:00 retries, 21:00 hard anchor

Q12

Scoring

In code (testable): 60%+ = win, 1-min = full credit, never-miss-twice; celebration tone in the skill

Q13

Encoding

Behavior = hard-rails in AGENT.md/SKILLs (no deploy); code only for mechanics — not the directives store

Q14

Definition of done

Text E2E green + one live phone-call smoke test per new call-kind

Q16

Heartbeat

Data-driven state-diffing — off-schedule nudge when the delta warrants, stay silent when nothing changed

Q17

Habit pillars

Overlay 3 keystone pillars + auto-detect objective ones; never auto-check personal (kept additive)

Q18

Persona guard

AGENT.md edits require an explicit confirmation (anti soul-drift)

Q19

Security

Sanitize + trust-boundary now; add the guardrail-agent the moment external/3rd-party content is read

Q20

State robustness

Atomic writes + file-locking on state.json (cron + webhook write it concurrently)

Q21

Voice

Voice-match summaries/drafts via RAG over your past messages/journal

Borrowed from the video + starter repo 7 patterns

Read coleam00/second-brain-starter in full. The coach already beats it on memory, live calls, and its sim harness — so these are adaptations, not a rewrite.

Pattern	How we're using it
Editable memory split (SOUL/USER/MEMORY)	AGENT / MATT_PROFILE / MEMORY files, loaded every prompt, promoted nightly
SOUL write-protection	AGENT.md edits require confirmation — stops the coach silently rewriting its own personality
Data-driven heartbeat + state-diffing	Off-schedule nudges + "stay silent when nothing changed" — the real anti-nag mechanism
HABITS pillars + auto-detection	3 keystone pillars overlaid on the field tracker; auto-check only objective wins
Proactivity dial (Observer→Partner)	Locked at Advisor→Partner — low-risk auto-actions, ask before external
Lethal-trifecta security	Sanitize + XML trust-boundary baseline; pre-flight guardrail agent when external content lands
Shared-utils robustness	Atomic state writes + file-locking on the call-loop's state.json

Pattern

How we're using it

Editable memory split (SOUL/USER/MEMORY)

AGENT / MATT_PROFILE / MEMORY files, loaded every prompt, promoted nightly

SOUL write-protection

AGENT.md edits require confirmation — stops the coach silently rewriting its own personality

Data-driven heartbeat + state-diffing

Off-schedule nudges + "stay silent when nothing changed" — the real anti-nag mechanism

HABITS pillars + auto-detection

3 keystone pillars overlaid on the field tracker; auto-check only objective wins

Proactivity dial (Observer→Partner)

Locked at Advisor→Partner — low-risk auto-actions, ask before external

Lethal-trifecta security

Sanitize + XML trust-boundary baseline; pre-flight guardrail agent when external content lands

Shared-utils robustness

Atomic state writes + file-locking on the call-loop's state.json

Still open — for Instance 2 5 technical

The product decisions are settled; these need code-level research before the build checklist.

ID Question to resolve

D2 Register the new tools (schedule_callback, set_wake_time, Notion) in the Pipecat voice path vs the dormant Gemini Live manifest

D4 Confirm every reader of project_registry.json before changing its schema (non-breaking)

D7 Which objective signals can actually auto-detect the revenue/focus pillars without false positives

D8 Make sure MATT_PROFILE/MEMORY injection doesn't double-inject with the live compact_recall layer

D9 Per-utterance transcript: fix _reconcile vs store the raw Deepgram transcript

ID	Question to resolve
D2	Register the new tools (`schedule_callback`, `set_wake_time`, Notion) in the Pipecat voice path vs the dormant Gemini Live manifest
D4	Confirm every reader of `project_registry.json` before changing its schema (non-breaking)
D7	Which objective signals can actually auto-detect the revenue/focus pillars without false positives
D8	Make sure MATT_PROFILE/MEMORY injection doesn't double-inject with the live `compact_recall` layer
D9	Per-utterance transcript: fix `_reconcile` vs store the raw Deepgram transcript

Next step: open a fresh Claude Code chat and paste plans/coach-daily-os/PROMPT_to_techspec.md. Instance 2 resolves D2/D4/D7/D8/D9, optionally runs /user-journeys for the multi-state call-loop, then /plan-techspec turns this into checklist.md. Nothing has been built or deployed — this is planning only.

Generated 2026-06-21 · source of truth: plans/coach-daily-os/HANDOFF.md (+ manifest, Instance-2 prompt) · telegram-coach · planning only, nothing deployed or committed.