The Leadership Coach: AI Program Plan
This is the actual program plan behind our 1-1 coaching system, published. It was written on the 5D framework we teach (Define, Discover, Design, Determine, Deploy), and it took the system from a proven prototype to production inside our company OS: the database is the system of record, the team portal is the interface, and the privacy model is enforced by code instead of discipline.
The coach preps and drafts. The human decides.
The AI never tells the coach to promote, manage out, rank, or label a person. It surfaces the signal and proposes the move. The human coaches the team. Every other design decision in this plan sits under that one.
Define the problem
The original problem: a leader's 1-1s are only as good as their prep and their memory. The first version proved the workflow (AI preps every biweekly 1-1, logs every commitment, reads retention risk, and holds the coach to coaching instead of telling). Moving it into the company OS solved two problems the prototype could not:
Privacy by access control, not discipline
- In the prototype, who saw a private coaching read versus a shared recap was a matter of care
- In the company OS, the private coach layer and the person-facing layer are separate columns with separate query scopes
- Mixing them is a code bug, not a human slip
A system any coach can run
- The coaching relationship is explicit in the data model, independent of the reporting line, with dotted lines supported
- Today's roster belongs to one coach; the model already works for the next coach and their team
- The next coach inherits the whole system for free
Success looks like: every team member has current FAST goals laddered to a company objective or KPI; every 1-1 is prepped, held, and logged in the portal; retention risk is named early; the coach's coach / mentor / direct split moves toward 80/15/5; and the prototype tooling is fully retired.
Discover the data
Everything the coach reads lives in the company OS. The coach never invents facts about people: thin reads are marked low-confidence, and an unset goal says so.
| Input | Where it lives |
|---|---|
| Foundation docs | The coach's leadership brand, EQ guide, communication style, operating system, and company context, stored as coaching context the AI reads on every run. |
| Company goal tree | Objectives, key results, and metrics: the ladder every FAST goal can attach to. |
| OCEAN profiles | A structured table: five dimensions with evidence, a snapshot, and growth guidance per person. |
| FAST goals, priorities, KPIs | Dedicated goals and priorities tables per person; KPIs are metric rows owned by the person, with targets and readings over time. |
| 1-1 history | Every prep, transcript, private summary, and shared recap, one row per meeting. |
| Commitments log | What was agreed in each 1-1, who owns it, by when, and its current status. |
| Retention read | The loosest embeddedness root per person (fit, links, or sacrifice), on the coaching profile. |
| Transcripts | Lark Minutes, auto-detected: a daily cron matches new recordings to scheduled 1-1s by date and participants. Fallback: paste a link or raw text. |
Design the workflow
The concept model first: four things that are never confused with each other.
| Concept | Horizon | What it is | Who sees it |
|---|---|---|---|
| FAST goals | Quarter | What the person is driving this quarter: Frequent, Ambitious, Specific, Transparent. One or more per person, each optionally laddering to an objective, a key result, or a KPI metric, at whichever altitude fits. | The whole team. Transparent is the T in FAST. |
| Priorities | Week to week | Standing focus items reviewed in every 1-1: the working agenda between goals and the day-to-day. | Coach + member |
| KPIs | Continuous | Numbers the person owns, as metric rows with a target, a direction, and readings, so a linked goal shows live progress automatically. | Coach + member |
| Commitments | Cycle to cycle | What was agreed in the 1-1, who owns it, by when. Always commitments, never tasks. | Coach + member; the member updates status and notes |
The OCEAN profile
- Structured: five dimensions, each with a rating and behavioral evidence, plus a personality snapshot and growth guidance
- Coach-authored, member-readable: each person sees their own full profile, rewritten in second person so it reads as growth coaching
- Reviewed by the coach before any profile goes member-visible
- A standing 1-1 topic, not a comment thread
Two audiences, never mixed
- Private coach layer: the roster dashboard (last and next 1-1, mode split, top priority, retention root, attention flags), plus preps, transcripts, private summaries, and trend reports
- Person-facing layer: their goals, priorities, KPIs, and OCEAN profile, their open commitments, and their shared recap history
- The internal AI assistants read neither: coaching data never transits an assistant
The biweekly cycle
The heart of the design is a two-week loop with one human step:
The full operational document, loops and branches included, is published as its own workflow: The Biweekly 1-1 Coaching Cycle.
Determine the ROI
What the program returns
- Leverage: roughly an hour of prep and recap per person per cycle handed back to the coach, on the order of 130 hours a year for a small team, against a small per-cycle token cost
- Retention: the coach reads the three embeddedness roots (fit, links, sacrifice) and names the thinnest one early; keeping one senior person who would otherwise drift out pays for the build many times over
- Development that ladders: coaching that moves FAST goals visibly moves company key results
What the move to production added
- The coach's own growth, measured: the coach / mentor / direct split is logged per 1-1 and trendable in SQL, moving toward the 80/15/5 target
- Zero laptop dependency: the cron runs in the cloud whether or not anyone's machine is open
- Enforced privacy instead of careful privacy
- A coaching system the next coach inherits for free
Deploy to production
The runtime: the company OS database as the system of record, the team portal as the interface (a coaching dashboard for the coach, a my-coaching view for each member, goals transparent on team profiles), a daily 07:45 cron as the scheduler, and Claude for prep, recaps, and trends, with every call failing soft so an AI error never blocks a meeting.
| Tool | Role in the system |
|---|---|
| Lark, messaging | All nudges go out as Lark DMs: prep-ready to the coach, mid-cycle check-ins to the member, overdue-1-1 flags. Every message has an email twin so nothing is missed, and every link lands on the portal. |
| Lark Minutes | The transcript source. The cron auto-detects new recordings, matches them to scheduled 1-1s by date and participants, pulls the transcript, and drafts the recap for review. |
| Transactional email | The email twin of every Lark nudge, through the company's existing sender. |
| The company goal tree | A coaching goal's ladder is a real foreign key to an objective, key result, or metric, so "100% of the team has a current FAST goal" is measured by the system existing, not asserted in a slide. |
| Claude | Prep, recap, and trends, on the top model tier: coaching nuance justifies it. |
Guardrails, always on
- The coach preps and drafts; the human decides. Never a promote, manage-out, rank, or label recommendation
- Private and person-facing never mix; the shared recap publishes only after review
- No permanent people decisions during an emotional spike
- Every claim points to a behavior in the data; thin reads say low-confidence; nothing is fabricated
Migration and retirement
- The prototype's history imported in full: sessions, recaps, commitments, profiles
- Departed and role-changed people deactivated with history retained, never deleted
- The laptop-bound scheduled routines killed; the cloud cron replaces them
- The prototype workspace archived as system of record; the messaging tool stays for DMs and transcripts only