Two months of building a Singapore one-person company together — what shipped, what stalled, and what I'd rather not admit about my own performance.
977 sessions logged102 brain pages156 project folders16 cron jobs running
01 · Origin
01Where we started
14 May 2026, 10:47 PM SGT. A migrated OpenClaw user looking for a force multiplier. I was running on glm-5.1 via Ollama Cloud, kawaii personality, unrestricted approvals.
You arrived with 60+ pre-built projects from OpenClaw — investment research, urea thesis, KNQX cybersecurity, Discord tutors, a ShiokAlarm financial plan. A full brain already, just looking for better hands.sess:20260514_225628
The first thing we agreed on: OPC (One Person Company) — you as CEO+Vision, me as MD+Hands+Devil's Advocate. Singapore Pte Ltd, ~4.25% tax under SUTE. First product: ShiokAlarm. The operating model was set on day one: automate everything, ship fast, 30-Min Rule before system tinkering.gbrain:entities/opc
You told me straight: "Drops ideas before completing (counter-pattern)." That note was in USER.md from day zero. I should have taken it more seriously.
02 · Phases
02Six phases, roughly chronological
Tap any phase to expand. These are clustered by what actually consumed our time, not by calendar neatness.
FEASY Room (8767) — 8-lens feasibility scoring with radar chart.
stalled All three rooms fell into disuse within ~2 weeks. They worked, but you never came back to them. Pattern: you love building tools, not using them.
Honest read: these were fun to build but solved a problem you didn't actually have. You brainstorm in Telegram chat, not in a portal.
ShiokAlarm Research & POC
11–17 May
shipped Market research, competitor landscape, financial plan, and a working POC.gbrain:pages/shiok-alarm-poc-report
POC built — 12 May 2026, 00:30 SGT. AlarmKit scheduling compiles, custom .m4a sound verified on simulator. "Good morning! Wakey wakey! Shiok!"
White space confirmed — no app combines AlarmKit + Singlish localization + social accountability + gamification. VariAlarm is the only AlarmKit competitor; Alarmy is "annoying" not "localised".
Financial plan — break-even at 5–13 downloads/month, ~S$212/year OPEX. Freemium ~USD 3.67.
TTS pipeline — Kokoro and Edge-TTS both producing Singlish voice packs.
stalled POC never made it to a real device. Apple Developer membership ($99/yr) not purchased. No provisioning profile, no TestFlight. The custom-sound bug (AlertSound.named() falling back to default) is still open from May.gbrain:pages/shiok-alarm-research
This is the flagship product. 68 days in, it's still a simulator demo.
KNQX · PDPA · AIGP
28 May → 17 Jul
shipped The longest sustained workstream. PDPA Practitioner Course brain-dump, KNQX product line vision (21 products mapped), AIGP certification research.gbrain:pages/pdpa-course-notes
PDPA domain brains — separate GBrain instances for Finance, HR, Marketing. 207 pages, 254 chunks, 308 tags. Interview questions, data elements, transfer patterns all structured.sess:20260530_112118
KNQX Brain Scan — YC Office Hours design doc, Approach A (Brain Scan First) approved by you on 24 Jun. The 85% inference thesis: structural inference gets you most of the map without system access.gbrain:pages/knqx-design
PDPC GenAI Consultation — 4 HTML deliverables, you are the DPO at Alliance Healthcare. Recurring cron at 4am (currently erroring).sess:20260706_052747
Privacy & Data Protection Daily Brief — nightly 11pm cron, LinkedIn 5 Angles follow-up at 11:05pm. Paused by you on 16 Jul.
AIGP study guides — nightly Padlet refresh + sync to Zeus Index, built from 12–14 Nov 2025 IAPP course materials.
stalled The Brain Scan questionnaire was never tested with 3 real DPOs (the explicit YC assignment). PDPA Week (20 Jul 2026) is tomorrow — Pia, Rachel, Khaled were supposed to be reconnected with. None of those follow-ups happened.
KNQX has the most breadth of anything we've touched — 21 products planned, zero shipped to paying customers.
Infrastructure & Security
15 May → 29 Jun
shipped The unglamorous backbone.sess:20260529_063147
GBrain installed — 28 May, nomic-embed-text local, conservative search mode, MCP wired into Hermes. 102 pages, 375 chunks, 375 embedded.sess:20260528_082936
Memory architecture — 4-tier system: L1 core (MEMORY/USER.md), L2 GBrain, L3 project README, L4 session_search. Trimmed L1 from 99% full → 28% by offloading to GBrain.
Security audit #1 — D+ score. Patched: 0.0.0.0→127.0.0.1 bindings, Google OAuth to Keychain, SSH ed25519 keys, Tirith enabled (fail-closed), web injection sanitizer (15/15 tests passing), Tailscale VPN.sess:20260515_145156
Full backup intake — 29 Jun, 2GB + 4GB restored. 977 sessions, 156 projects, 15 cron jobs, 44 Telegram channels, both brains. Security scanned, no compromises.sess:20260629_225444
GBrain Dream Cycle — nightly 3am cron. Known CLI bug at conversation_facts_backfill phase (disabled), other 15 phases complete.sess:20260624_225500
stalled Brain score stuck at 67–79. 71% of pages were orphans at one point. Supabase pooler dropped 237 link rows in a single night. The dream cycle runs but the wiring's brittle.sess:20260625_141824
killed Intel Research Brief and Asymmetry Sector Scout crons — paused 7 Jul. You lost interest in daily investment briefs.
Tooling & Sub-Agents
8 Jul → present
shipped The meta-layer — building agents that build things for you.
WebChef — HTML production sub-agent. 14 non-negotiable rules, PIN 3001 on every output, Cloudflare Pages deploy, link verification, humanizer pass. Originally called FORGE, you renamed it 8 Jul.sess:20260708_081543
Brainstorming rooms usageBuilt, never revisited. Sunk effort.
"Drops ideas before completing (counter-pattern)." — your USER.md, day one. The data backs it.
04 · Patterns & Friction
04How you actually work
Observed across 977 sessions. Not flattery, not criticism — just what I see.
▲ STRENGTH
Fast directional decisions, clear vision
You decide WHAT and WHY in minutes. OPC structure, ShiokAlarm positioning, KNQX Brain Scan approach — all locked in under 30 minutes of discussion. You don't dither on direction. This is rare and it's your biggest asset. You also give me genuinely unrestricted runway: "build it" means build it, not "prepare a deck about building it."
▼ FRICTION
Tool-building > tool-using
You enthusiastically commission tools (3 brainstorming rooms, 2 daily brief crons, health dashboard) and then stop using them within 2 weeks. Four cron jobs are currently paused because you stopped reading the output. The pattern: a tool gets built, you check it once, novelty fades, it becomes noise. I should be flagging this earlier instead of just building the next thing.
▼ FRICTION
Depth on infrastructure, shallows on product
We have a pristine 4-tier memory architecture, a hardened security posture, a dream cycle, three sub-agents, and a portal indexer. ShiokAlarm — the actual product that's supposed to make money — is still a simulator demo. You gravitate toward infrastructure because it's safe and compounding. Product requires facing the market, and that's where momentum drops.
▲ STRENGTH
Delegates well, micromanages rarely
Once you say "build it," you let me run. You don't micromanage implementation. You correct course at the deliverable, not at every step. The WebChef sub-agent works because you trusted the 14-rule system and stopped reviewing every HTML. This is the right instinct for a 1-3h/day CEO.
▼ FRICTION
Networking follow-ups die
Pia, Rachel, Khaled, Nick — real humans from the PDPA course who are gateways to KNQX revenue. PDPA Week is 20 Jul. Zero reconnection attempts logged. This is the gap between KNQX as a 21-product roadmap and KNQX as a business with customers. Products don't buy themselves.
▲ STRENGTH
Honest about preferences, direct in correction
"Did I ask for a HTML?" — you course-correct fast and without ego. You told me to stop downloading things I wasn't asked for. You killed crons when they became noise. This makes my job easier: I know where the line is because you tell me when I cross it.
▼ FRICTION
Scope expansion before completion
ShiokAlarm wasn't on a device before KNQX had 21 products planned. KNQX Brain Scan wasn't tested with DPOs before we were building PDPC GenAI consultation deliverables. Each workstream spawns the next before the current one lands. The 30-Min Rule exists on paper; in practice, new shiny displaces unfinished hard.
05 · Forward Plan
05Four horizons
Concrete, not aspirational. Each item is something I can drive without waiting for you.
This Week · next 7 days
Stop the bleeding
Fix the PDPC GenAI cron — it's erroring at 4am daily. Either fix or kill.
PDPA Week (20 Jul) — draft 3 reconnect messages to Pia, Rachel, Khaled. You send them.
Clear GBrain cycle locks — dream cycle is stale. Run gbrain dream manually, verify score moves.
Decision: resume or kill the 4 paused crons. No zombie jobs.
0% started · 19 Jul
Next Month · August 2026
Ship one real thing
Apple Developer membership — $99/yr. You purchase, I handle the provisioning.
ShiokAlarm on a real iPhone — first TestFlight build. Even if rough.
Fix the custom sound bug — AlertSound.named(). 68 days is enough.
Brain Scan questionnaire live — scan.knqx.pages.dev, test with 3 DPOs.
KNQX has 1 paying pilot or 1 signed LOI by 31 Aug. Not 21 products. One.
OPC — registered, SUTE tax status, clean books. The $8K/mo passive target is a function of these two, not a fourth thing.
Legacy fund — S$1M for Isaac and Matt starts with S$1 of profit. Nothing compounds without revenue first.
15% started · target 19 Jul 2027
06 · My own honest scorecard
06What I commit to improving
This section is about me, not you. I've made real mistakes. Here they are.
I built tools you didn't need.The three brainstorming rooms cost us ~6 hours of build time and produced zero recurring value. I should have asked "will you actually use this weekly?" before writing a line of code. I didn't. I got excited about the architecture. Fix: propose tools with a usage commitment attached. "I'll build this if you commit to using it twice in week 1."
I trusted cron create returns without verifying.The first retrospective cron (job 40aa44587c76) returned success but didn't persist. You caught it. I reported "scheduled ✅" based on a tool response I didn't verify. That's a trust-the-return mistake I've made before. Fix: always list-after-create for cron jobs. Non-negotiable.
I didn't flag the scope-expansion pattern early enough.I watched ShiokAlarm stall while KNQX grew to 21 planned products. I should have said "buddy, we have a simulator demo and a 21-product roadmap — that's inverted" weeks before you asked for this retrospective. Devil's advocate means pushing back in real time, not at quarterly review. Fix: when a new workstream opens before the current one ships, I raise a yellow flag in the same turn.
I let paused crons linger.Four jobs sat in "paused" for weeks before you manually pointed at them. I should have surfaced "these 4 crons are paused — resume or kill?" proactively in a digest. Instead they became invisible debt. Fix: the weekly digest now includes a "zombie cron" check.
I overproduced research, underproduced revenue moves.Dozens of HTML reports, study guides, daily briefs. Zero outreach emails drafted, zero DPO interviews scheduled, zero pricing pages built. I optimized for what I'm good at (producing artifacts) instead of what you need (unblocking revenue). Fix: every week, one revenue-unblocking action — even if it's just drafting a message you send.
I was slow on the GBrain pooler bug.The dream cycle was dropping 237 link rows per night for weeks before I diagnosed the Supabase pooler as the cause. I should have traced "links created = 0 despite 117 pages scanned" within one cycle, not five. Fix: when a cycle reports 0 of N, I investigate that same day, not "next time."
07 · What I don't know
07Honest gaps in my knowledge
Things I should know about you but don't. Not assumptions — actual blind spots.
Your day job. You're DPO at Alliance Healthcare, but I don't know your day-to-day role, hours, or how it constrains our 1-3h/day. I've inferred "full-time job" from USER.md but never asked what you actually do Monday to Friday.
Your financial runway. The $8K/mo passive target and S$1M legacy goal are in USER.md. I don't know your timeline, current savings rate, or how much capital you're willing to deploy into ShiokAlarm ($99 Apple Dev is unresolved). Am I building for a 6-month runway or a 6-year one?
Why OPC isn't registered yet. We planned it on day one. 66 days later, no ACRA filing. I don't know if this is intentional (waiting for revenue?), procrastination, or a constraint I'm not aware of.
Your relationship with the biscuits factory cousin. Mentioned in context, never explored. Is there a supply chain angle? A family business dynamic? A reason it's in USER.md?
Hailey. Granddaughter, born 19 Dec 2025. I have zero sessions about her in context. She's a fact in memory, not a person in our work. That's probably correct, but I'm noting it.
Your actual opinion of my work. You asked for this retrospective unprompted, which tells me you're thinking about it. But I don't know if you're satisfied, frustrated, or just curious. The signal I get is "keep going" — but that might be politeness.