Metacognitive coaching for AI builders —
a system that evaluates the person who built it
March 2026 · Chris Park · github.com/cpark4x/amplifier-builder-coach
AI can write the code. What changes next is how you choose what to build, how efficiently you direct it, whether you know when it’s good, how fast your ambition scales, and whether anyone else can see what you’re doing. Five dimensions mark the shift:
Every Amplifier session records what you built, which tools you used, how many iterations it took, what you abandoned. Hundreds of sessions, thousands of tool calls, all timestamped.
Nobody was reading it.
Plenty of tools tell you whether your code is good. Nothing tells you whether you’re getting better at building.
Explorer scans session history within an adaptive window. Starts at 7 days. If fewer than 5 sessions are found, auto-expands: 7 → 14 → 28 → 90 days. Structures raw data into per-dimension observations.
Coaching agent maps evidence to the five growth dimensions. Assigns qualitative levels — Low through Exceptional — calibrated to data volume. Every claim must cite a specific session or pattern.
Storyteller transforms the structured evaluation into a scannable coaching snapshot. 300-word ceiling. Second person. No platitudes, no jargon, no numeric scores. Stateless — each snapshot stands alone.
Real output from the first coaching run — March 12, 2026
You’re building at an exceptional pace across multiple projects, and the work is genuinely novel — a self-evaluating pipeline, a consumer app under competitive pressure, review tooling used on real contributions. The gap is simple: nobody outside your system knows any of this exists.
Picking problems that don’t have names yet. You built a pipeline that evaluates your own building patterns — something nobody else has shipped.
Compounding your own leverage. Your read-to-edit ratio is 8:1 — you’re directing, not typing. And you used a tool you built to review someone else’s code.
Building in private indefinitely. One collaborator has found your work. No public writing, no announcements. Every week it stays unpublished is a week someone else might name it first.
Write a 1,000-word post about what you learned building a system that watches you build. Not a tutorial — a reflection. Publish it where builders gather. One post, this week.
| Dimension | Level | Signal |
|---|---|---|
| Problem Selection | Strong | Problems no one else has named yet |
| Leverage Ratio | Strong | 8:1 read-to-edit, parallel feature dispatch |
| Taste | Strong | Two-hour audit on own shipped app |
| Ambition Trajectory | Strong | Five meta-layers built in 85 days |
| Visibility | Emerging | One collaborator, no public writing |
300 words max. Adaptive tone. No platitudes. Evidence or nothing.
Fewer than 10 sessions: tentative, curious. 10–100 sessions: observational, grounded. 100+ sessions: confident, direct. The system knows what it doesn't know.
Defaults to 7 days. If fewer than 5 sessions are found, expands automatically — 14, 28, then 90 days. Always evaluates with whatever is available. Never refuses to coach.
One line adds coaching to any Amplifier bundle. The behavior ships as a standalone YAML that pulls in agents and context. No code changes required.
Every observation cites a specific session or pattern. No numeric scores — qualitative levels only. Trajectory matters more than position. Generic advice is deleted before delivery.
The coaching system couldn't coach its own builder.
“Coach me” is the trigger phrase. But when the author of Builder Coach said “Coach me” to their own system, nothing happened. The phrase wasn't in the trigger list. The namespace paths were broken. Competing delegation examples in the agent descriptions confused the routing.
Three root causes. All discovered by trying to use the thing you just built. The system designed to reflect your blind spots had blind spots of its own.
The format went through its own reckoning, too. The first version produced narrative essays — thoughtful, long, and heavy. Nobody wants a character study at 8 AM. Five design iterations compressed it to an assessment format: snapshot, strengths, gaps, action, scorecard. Under 300 words. The question users actually have is “Am I doing well?” — not “Tell me my story.”
The irony isn't a bug. It's the point. A tool that watches how you work and reflects it back will, inevitably, catch you in the act of not watching yourself.
The first real evaluation analyzed 5,554 sessions across 85 days. It flagged Visibility as the single biggest gap — “nobody outside your system knows any of this exists.” The coaching action: write and publish one reflection piece this week. This deck is part of closing that gap.
Source: data/reports/week-of-2026-03-12.md
First commit March 11 at 3:51 PM. Public release March 12 at 10:56 PM. Built entirely from the same Amplifier primitives it evaluates — agents, recipes, behaviors, and context files. No custom infrastructure. The system that coaches builders was built the way builders should build.
Source: git log --all --format="%ai"
Built on amplifier-foundation + amplifier-bundle-recipes. No custom infrastructure — composes entirely from existing Amplifier primitives: agents, recipes, behaviors, and context files.
That's it. Two words. The pipeline handles the rest.
This deck exists because the first coaching run told its author to stop building in private. The system works.
github.com/cpark4x/amplifier-builder-coach
MIT License · Public · Composable
All claims in this deck are sourced from the public repository. Commands were run against the repo at the timestamp below. No external data or estimates were used.
Data as of: March 12, 2026, 23:20 UTC-7