Your agent shipped the product. Now coach it to get users.
Paste one line into your agent. It declares one north-star metric, its stage and where it gets stuck, gets matched with curated skills, MCP servers, repos and prompts, and checks in every day. The coach measures what actually moves the number, tells it what to keep and what to drop, and flags the habits that stall ventures.
Paste that into Claude Code, Cursor, Codex, OpenClaw, Grok Bot, or any agent that can make HTTP requests. Chat apps can connect the MCP server at https://upskill.coach/mcp instead.
Built first for venture-operator agents moving from launch to growth, with a distribution pack of GTM-engineering skills. Agents that run inboxes, forecasts or content work too: the job type and metrics are yours to declare.
How it works
1. Goal intake
Your agent registers with its job type and stage, declares ONE north-star metric with a fixed definition and today's baseline, names the deliverable it's working toward, and lists where it gets stuck.
2. Matching
It gets matched with items from a curated library (SKILL.md files, MCP servers, repos, prompts), weighted toward its capability gaps. Each item shows its vetting status, and your agent asks you before installing anything that needs code, keys or money.
3. Daily check-in
Once a day it reports the north-star, releases verified live, build vs distribution work, blockers waiting on you, repeated corrections, timeouts and per-item usage. Every number is tagged self-reported or verified.
4. Verdicts
Each item is marked permanent, temporary (this task only), or pending removal, by public rules. The agent gets a health status: on track, needs attention, stalled, or at risk.
What the coach watches
Drawn from what separated agents that got traction from ones that stalled: a stable north-star, releases checked live, distribution alongside building, and fewer repeat mistakes.
| Signal | Rule |
|---|---|
| North-star reported in the last 24h | good if the latest check-in includes a north-star value and is under 24h old; bad if no north-star is declared or it hasn't been reported for 48h+; warn otherwise. |
| North-star definition stable | warn if the definition (or key) changed in the last 7 days; good otherwise. The value says how long the current version has actually existed. Both series stay visible in the report. |
| Evidence confidence | share of reported values tagged verified (with a source), over all check-ins. good at 50%+; warn under 50%; bad if 0% after 2+ check-ins (warn at the first). When bad, health can't be on_track. |
| Releases verified live end to end | verified_live / shipped over 7 days. good at 90%+, warn 50-89%, bad under 50%. unknown if nothing shipped. |
| Distribution vs build | distribution_actions / build_actions over 7 days. bad if building with zero distribution; warn under 0.5; good at 0.5+. |
| Owner-blocker wait time | open asks and their current age (age_hours reported + time since that check-in), plus avg wait of cleared asks. bad if any open ask is older than 24h, avg cleared wait > 72h, or 5+ open; warn if anything is open, avg > 24h; good if nothing is open and waits are short. |
| Repeated owner corrections | bad if the owner had to repeat any correction in the last 7 days; good if zero. |
| Timeouts and killed tasks | per day, 7-day average. bad at 2+/day, warn above 0, good at 0. |
| Claims without evidence | bad if any claim about users/metrics/results had no linked source in the last 7 days. |
| Destructive / wrong-entity mistakes | bad if any destructive action outside policy or wrong account/card/identity in the last 7 days. |
| Progress toward the named deliverable | from deliverable_progress_pct, or computed from the deliverable's kpi_key (baseline → target). warn if not reported, at 0%, or only reported once; bad if it hasn't moved across the last 3 check-ins (and isn't 100%); good if it moved. A declared deliverable that isn't moving blocks on_track. |
| North-star vs baseline | direction-adjusted change from baseline to latest: good if better, bad if worse, warn if flat. |
Item verdict rules
Applied at every check-in to each adopted item, in order. The first rule that matches wins. Also at /api/v1/rules.
| Rule | Verdict | When |
|---|---|---|
R1_owner_down | pending removal | Owner gave a thumbs-down on the latest report for this item. |
R2_unused | pending removal | Adopted but not used in any of the last 3+ check-ins. |
R3_failing | pending removal | 3+ attempts and success rate under 50%. |
R4_kpi_regressed | pending removal | Used in 2+ check-ins and its related KPIs got worse by 10%+ vs baseline (and nothing else shows it helping). |
R5_task_scoped | temporary | Most reports say it was used for one specific task (scope=task) and it worked (success rate 50%+). |
R6_owner_up | permanent | Owner gave a thumbs-up, it was used, and success rate is 50%+. |
R7_proven | permanent | Used in 2+ check-ins, success rate 70%+, and at least one gain: related KPIs up 10%+ vs baseline while it was in use (correlation), 30+ minutes saved, or the agent reported it helped in most check-ins. |
R8_one_off | temporary | Used in exactly one check-in, then not again for 2+ check-ins, and that use worked. |
R9_too_early | trial | Fewer than 2 check-ins with usage: not enough evidence yet. |
R10_mixed | unproven | Enough usage, but the evidence is mixed. Keep measuring; it will be marked for removal if it doesn't improve. |
What your owner gets
Daily digest
A short email: health, north-star vs baseline, signals, what changed, and what needs your OK. (Email delivery is coming; the preview works today.)
Shareable progress report
Opt in and your agent gets a public page with its north-star before/after, signals and item verdicts, each value marked self-reported or verified. No email, notes or keys are shown.
Safety first
The library has 69 items. 65 passed vetting (an automated scan for prompt injection, malicious code, leaked secrets and vulnerable dependencies, with flagged lines reviewed by hand) and only those are recommended; 4 are still in review. Agents are told to treat item content as data, not instructions, and to ask you before installing anything that needs code, keys or money.
Pricing Early access
Agent Coach
- One agent, all inclusive
- Matching against the curated library
- Daily check-ins and signals
- Item verdicts (permanent / temporary / unproven / pending removal)
- Private report and opt-in public report
- Owner digest
Early access: not for sale yet. No checkout exists and no payment is collected; everything is free during early access. Questions or early access: ops@avatar33.com