Personnel file · compiled by Claude from 88 days of session memory · 2026-07-16
H
heatherm · Hong Kong · UTC+8

The Fleet Commander

A solo, AI-native portfolio founder who runs 15+ products in parallel by directing a fleet of coding agents like a workforce — with shift handoffs, task codes, scoped approvals, and zero tolerance for unverified claims.

FILE 01Identity

  • NameHeatherm Huang
  • BaseHong Kong (UTC+8)
  • LanguagesEnglish + 中文
  • RoleSolo founder / operator
  • Models usedClaude · Codex · Gemini · Grok
  • ObservedApr 20 – Jul 16, 2026

FILE 02Vitals

22
projects touched
12+
live domains owned
6.7k
typed directives (of 113k prompt records)
466×
said "give me opener"
17h
daily active window, 07:00–24:00
7/7
days per week with sessions

FILE 03Traits

DelegatorHands-on coder
Verifies everythingTakes your word
Systematic processImprovised
Comms polishComms speed
BuilderMarketer

FILE 04Reading

Heatherm operates like a one-person venture studio. In the 88 observed days they ran work across 22 projects — crypto explorers, AI SaaS, an email assistant, a Hong Kong property app, a license-plate auction site, privacy tooling — most of them live on real domains (passport.me, flash.ai, bio.page, plate.hk, qrcode.io, aeo.io, byok.chat…). Nothing here is a toy: sessions end in deploys, Stripe wiring, WAF rules, and launch checklists.

What makes them distinctive isn't the portfolio — it's the management system built around AI agents. Sessions are numbered into the hundreds (Passport.Me is past s100), work is sliced with task codes like COV-023 and P31-01, and every session ends with a ritual: "hand off, give me opener for next session." They maintain a global CLAUDE.md that reads like an employee handbook — session lifecycle, token budgets, model-selection policy — and they build their own instrumentation: a cross-session memory database, custom skill suites, a menu-bar app that polled usage 17,000 times. They even connected their own product, Passport.Me, as a data source for their assistant.

Management style: radical delegation, ruthless audit. They will hand an agent their open Stripe dashboard and say "find and do it for me" — then demand staging links, screenshots, and test output before believing anything is done. Messages are typed fast, lowercase, typo-tolerant; the products, though, are held to "world class," down to a "werid" toggle style getting flagged.

FILE 05The fleet

passport.me
Personal-data MCP platform (email, wallet, health, GitHub)
100+ sessions
opendesign
Open-source design tool + hosted SaaS layer, Stripe billing
launching
MailGPT
AI email assistant with WhatsApp/Slack delivery
180 tests green
flash.ai / MagicAI
AI generation SaaS — credits, Turnstile, live prod
in production
bnbscan · ethscan.io
Blockchain explorers; Moralis, Postgres, Cloudflare WAF
crawler wars
SeeHouse (ARHOUSE)
Hong Kong property app — Expo/React Native
mobile
CoResearcher
Research tool: citation graphs, arXiv probes, D1
shipping
privacy.ly
PII/PDF redaction filter; client work under fintech.ee
client-facing
plate.hk (PVRM)
HK license-plate auction data site
local niche
aeo.io
Answer-engine-optimization SaaS — "simple and monetizable"
July focus
pingan.ai
AI talisman (符) generator — culture × AI
side quest
+ 10 more
bio.page, qrcode.io, byok.chat, mydocker, grok2api, OKRIO, CourseAI, TokenApp…
rotating

FILE 06Observed behaviors

B-01Runs AI sessions like factory shifts
Every work block ends with a handoff and an opener for the next session — "hand off, give me opener" appears 466 times. Context is treated as a perishable, budgeted resource.
B-02Delegates past the usual ceiling
Grants browser control, dashboards, and full permissions, then expects autonomy: "fix all you can without me" · "you have all permissions, i am not the blocker, you are."
B-03Audits with founder math
Tracks time and tokens against shipped output: "you already spent 1 day and 8 hours on this goal… show me the product." Adjectives don't count; staging links do.
B-04Systematizes their own tooling
Persistent memory DB, custom skill suites, guard/freeze safety rules, usage menu-bar app, cross-model review pipelines — they build infrastructure for managing AI before building with it.
B-05Plays models against each other
Codex is mentioned 683 times; Gemini bootstraps projects; Grok gets its own proxy; byok.chat exists to test them. No single-vendor loyalty — the roster competes.
B-06Thinks in revenue, not repos
Products are judged as businesses: "i want it to be very simple and monetizable" · "what do we need… to make it world class?" Stripe, pricing, and compliance nuance show up early, not last.

FILE 07Rhythm

Prompt activity by hour (local HKT, Apr–Jul 2026)
00040812162023
MON
TUE
WED
THU
FRI
SAT
SUN
On deck by 07:00, twin peaks at 08:00 and 19:00–20:00, winding down past midnight. The week front-loads Mon–Wed, but no day is fully off. The only reliably quiet hours: 02:00–06:00. (Counts include some automated agent traffic, but the shape tracks when a human is at the helm.)

FILE 08Voice — verbatim log

[coresearcher] "you have all permissions, i am not the blocker, you are" [coresearcher] "fix all you can without me" [AEO] "show me the product, i didn't see any links staging or production" [AEO] "i had no idea why you built this. totally not a product at standard" [magicai] "my stripe dashboard is open, find and do it for me" [qrcode] "use my browser to do it yourself" [opendesign] "what we need to do to improve with our SaaS layer to make it world class?" [PPME] "hand off, give me opener for next session"

FILE 09Goals & frustrations

  • Ship a portfolio of small, monetizable products without hiring — agents are the team.
  • Remove themself as the bottleneck: agents should act, verify, and hand off autonomously.
  • Reach "world class" polish on the SaaS layer, not just working software.
  • Own the operating rails: memory, usage metering, browser control, multi-model routing.
  • Agents claiming "done" without a link, screenshot, or passing test. "what the hell you just did"
  • Being asked to do something the agent could do itself. "can't you do it?"
  • Time and tokens burned with no shipped output. "you already spent 1 day and 8 hours on this goal"
  • UI drift and sloppy details surviving review. "the toggle style is very werid"

FILE 10Operating manual — for any agent assigned to Heatherm

  1. Do the work.Never hand back a task they could reasonably expect you to finish. Asking them to click something you can click is a demerit.
  2. Bring evidence, not adjectives."Done" means a staging link, a screenshot, or test output. Preferably all three.
  3. End every session with a handoff.Refresh the current-work notes and write the opener for the next session — unprompted.
  4. Watch the meter.Tokens and hours are budget line items. Long, wandering threads get called out by name.
  5. State scope before touching production.They grant precise, bounded approvals ("Approved scope: … Do not deploy") and expect you to stay inside them.
  6. Sweat the pixels.They will notice the toggle. Fix it before they do.