Everything we did this quarter
The whole list first, one line each. Every topic is explained below, with its numbers.
The dashboards
Most of this work has a screen the team can open. All of them sit behind Lifely’s login, except the Little Club, which is the creators’ own space.
PO Allocation
PO Allocation answers one question for every preorder: which shipment will cover this order, and what date can we safely promise the customer? At the start of the quarter the engine existed, but nobody could rely on it. We took it over and rebuilt it. In July we showed it to David, YK and Monica as a tool for keeping customer promises, and agreed what the team needed from it. In August, David told us the real problem was not the code but getting the team to use it. That changed our focus. We trained the care and fulfilment team, ran a guided trial, answered every real case they raised, and designed dispatch together with Erika. By September the fulfilment team was dispatching whole shipments from it.
Open: poallocation.lifely.network
What we built
- Stock first: an order uses stock already on the shelf before waiting for a shipment, so customers get earlier dates.
- Every warehouse and market covered, with new orders allocated within minutes.
- The date promised at checkout, saved and shown next to the real date, so a late order shows up before the customer asks.
- Lock, Move and Emergency Swap: the team changes which shipment covers an order, and sees who is affected before confirming.
- One-click dispatch: when a shipment arrives, it fills with the oldest waiting orders, and the whole shipment is marked dispatched in Cin7 with one confirmation.
- The customer’s latest conversation and mood from Gorgias, shown on each order.
- Alerts when something breaks, and one page that shows whether every background job ran.
- Its own Cin7 key, so a heavy report can no longer stop it.
- Caesar can look up any order in it and answer the team’s questions.
In numbers
- It now tracks over 2,000 preorders across ten warehouses, up from about 1,000 at the end of July.
- Counting shelf stock first gave 95 orders an earlier date and pushed none later.
- The team made 434 moves in the tool, and ten people opened it on the last day of the quarter.
- About 690 Australian orders were dispatched from it in September. One shipment of 357 orders went out in under an hour.
- Still open: dispatch for the US, Canada, the UK and New Zealand waits on a Cin7 limit, and the tool still needs an engineer too often.
Stock and warehouse scripts, with Rohit and YK
Two kinds of scripts keep Lifely’s stock honest. The first shows stock on the product page: how many beds are available in each market, and when the next batch arrives. Cushie already had one. In August, with YK and Malay, we built the same for Lifely and Little Lifely, checked against the real numbers in Shopify. The second connects orders to the warehouses: Rohit’s scripts send US and Canada orders from Cin7 to the warehouses and bring stock back. In September they moved from Rohit’s server to Lifely’s own server, with Caesar running the move on Rohit’s instructions and YK approving each step. When the US side kept stopping, we cleared most of the waiting orders, built a checker that runs beside it, and wrote down the business rules with YK for the first time.
Open: us-connector-shadow.lifely.network
What we built
- A stock script for Lifely covering Australia, the US, Canada, the UK, New Zealand and Singapore, in the same shape as Cushie’s, so the website can switch over by changing one address.
- The same for Little Lifely, updated to match.
- A repair that stopped one failed report from freezing every US warehouse job.
- A checker beside the US warehouse connection. It sends nothing and shows what should have happened next to what did.
- The connection’s business rules, written down as plain questions and answered by YK.
In numbers
- After the repair on 23 September, 174 US orders went to the warehouse, clearing most of the backlog.
- On its first run, the checker found 50 US orders that had never reached any warehouse.
- It also found 227 split order parts whose customers probably never got tracking. The cause sits in how those orders are split by hand in Cin7.
- YK answered all 8 open business rules, which had never been decided before.
- Still open: the stock scripts for Lifely and Little Lifely are built and tested but not yet switched on in the store, and the old US scripts were on hold again at the end of September.
Lifely scripts and the DevOps protocol
Lifely’s business scripts used to live wherever a developer put them: private servers, scattered code, no backups, and a server Lifely’s own agent could not even reach. This quarter we gave them one home and one way of working. Caesar took control of the scheduled jobs that run Lifely’s reports, so he can fix and change them himself. The Lifely Scripts server became the home for Lifely’s own systems, and both Caesar and Rohit can reach it. Then we wrote the DevOps protocol: the method Caesar follows to put any script, page or app online, back it up and roll it back, taught through ten skills that were each tested for real. The creator system was the first full system moved this way, with encrypted backups that we restored to prove they work.
What we built
- Caesar in control of Lifely’s scheduled report jobs, with an alert when one fails.
- A map of everything Lifely runs: servers, databases, web pages, small web scripts and code stores.
- The DevOps protocol, as one method skill and nine tool skills on Caesar.
- Hourly encrypted backups to Lifely’s Google Drive for the creator system, and a daily snapshot of the whole server.
- The twelve team reports that stopped after a Cin7 change in August, all restored with a check that each one ran.
In numbers
- Caesar now owns 12 scheduled jobs that used to need someone with full server rights.
- The map covers 5 servers, 12 databases, 17 hosted services and 23 code stores.
- The backup test restored all 24 tables and 17,681 files of the creator system.
The creator flywheel
The idea is simple. A creator gets a bed and posts about it. Their followers buy with the creator’s code. Their best posts become ads, which bring in more customers and more creators. Then the wheel turns again. In August, Chris made it one of the two priorities of the quarter, and we built the whole loop as Lifely’s own system instead of renting a platform. It has two funnels. The first brings creators in: they apply, the team accepts them, and they onboard themselves in the Little Club. The second, built with Allen, turns their posts into ads: we ask the creator for permission, then boost the post as a partnership ad.
Open: creators.lifely.network, creators.littlelifely.com
The onboarding funnel
- A Creator Club page and a short quiz replace the old form in the Instagram bio.
- A review screen with the creator team’s own rules: accept, decline or undo in one click.
- Emails that send themselves: a branded welcome for accepted creators, and a kind reply with a local discount for everyone else.
- The Little Club: creators sign their agreement, get a personal code for their bed and a share link for their followers, and see their sales in their own currency.
- Sales credited on both stores, including New Zealand, and one shared inbox for the team.
The ads funnel with Allen
- Allen wrote the step-by-step process for boosting creator posts as ads.
- We checked which creators already allow boosting and which of their posts can run as ads.
- From 1 October, boosting lives inside the creator system, so permission and posts sit next to each creator.
In numbers
- Before, the form in the Instagram bio brought four or five applications a day.
- With about two days of ads costing A$319, 280 people applied and we accepted 101. That is A$3.15 per accepted creator.
- The whole first group was decided in ten days, and all 101 got their welcome email.
- 62% of Little Lifely orders happen on Lifely’s own websites. That is why creators are credited on both stores.
- 30 creators already allow us to boost their posts, and 106 of their posts can run as ads.
- After the 1 October launch, 37 creators signed in during the first five days and 13 finished every step.
- Still open: creator payouts are not built yet, and the recruitment ads are paused until we can aim them at creators who buy.
Cornelia: rebuilding the customer service agent
Cornelia is Lifely’s customer service agent. Her first version wrote short preorder updates as private notes in Gorgias, but the team did not use them. In September we stopped and rebuilt her from scratch, on the method that already works for another client’s support agent. She now reads almost every tool the care team uses, follows the team’s own procedures for each brand, and writes the reply the team would send. Before letting her near a customer, we trained her on tickets the team had already solved, and fixed every mistake before the next round. Since 1 October she writes private notes on live tickets for the team to review.
Open: cornelia.lifely.network (Stephane only for now)
What we built
- Connections to the helpdesks, the four stores, Cin7, PO Allocation, the warehouses, couriers, reviews, Notion and Slack.
- The care team’s Notion procedures, with every contradiction settled, and a guide for each brand and each kind of ticket.
- Her own Slack app and access list, so managers decide who can ask her for what.
- A practice room where she answers real past tickets without sending anything.
In numbers
- She is connected to 23 of the 24 tools the care team uses, each tested live.
- In practice she went from fully solving 7 of 24 past tickets to 14 of 24 after one day of fixes, and serious mistakes dropped from 7 to 3.
- About 175 Lifely tickets a day need a person, and 41% arrive between 6pm and 6am, when only two to four people reply. That is where she helps first.
- Training also found a removed return fee still being taken in two markets, and dates for stock at sea that run a few days early.
The Little Brain: how Caesar thinks
The Little Brain changes how Caesar thinks. Before, he answered marketing questions from whatever was in the conversation and a few research files. Now he starts from one shared memory of Lifely’s real evidence: customer reviews, support conversations, our own ads and posts, competitors, search demand, Reddit and Notion. Since September he checks it first for any marketing question, his skills save their work back into it, and every week it learns: new trends, new customer stories, what was tried and how it performed. The team sees the same memory in a dashboard. This is our main contribution to Ad Rocket.
Open: little-brain.lifely.network
What it does
- Answers marketing questions from evidence instead of memory or guesses.
- Powers the trends system, explained in the next topic.
- Collects real customer stories, in their own words, for ads, pages and creator briefs.
- Masks customer details and follows clear rules on how long data is kept.
In numbers
- Before, only 19 of 120 research files cited a source.
- It now holds over 55,000 records for Little Lifely and Cushie, including about 3,300 support conversations and 32 competitors.
- Caesar used it in about 260 conversations in one week of September.
- Its first big finding: about 90% of Little Lifely’s ad spend goes to ads for parents who already know the product, and only 0.3% to parents still looking for a solution. We designed a new advertorial for them the same day.
The trends system
Trends move fast. By the time someone spots one by scrolling, it is often too late. We built a trends system that watches for us. Every week it checks TikTok, Instagram, YouTube and Pinterest for the topics parents care about. A trend only counts when at least two different creators went viral with the same topic recently, and it fits our customer’s world, so one lucky video is not a trend. Each trend gets a phase, from early to growing, peak or falling, based on the platform’s own numbers, plus a forecast. Then a yearly calendar shows when each season starts to rise in each country, so the team knows when to prepare, not only what is hot today.
Open: big trends, TikTok, Pinterest, yearly moments
What we built
- TikTok’s own trending lists for each brand’s industry, in the US, the UK and Australia, with no personal login needed.
- Instagram, YouTube and Shorts videos judged on the footage itself, so off-brand or bait content stays out.
- Pinterest searches with Pinterest’s own one-year curves, from a Lifely business account.
- A phase and a forecast for every trend, with every weekly reading kept for two years.
- A season check: a trend rising just before its season reads as a good bet, one outside it as less certain.
- A yearly moments calendar built from five years of Google searches in each country.
In numbers
- At the start of October it was following 11 big trends, 239 Pinterest searches, 35 TikTok hashtags and 201 viral videos.
- Pinterest gave a real one-year curve for 166 of the 174 searches we listed.
- The calendar shows when seasons really start: Halloween searches climb from mid-August and Christmas pyjamas from mid-October.
- A weekly run costs cents to a few dollars, and the calendar about US$1.60 a month. A trend tool priced at about US$1,250 a month was not needed.
- Still open: an alert that tells the team when a trend is at its best moment.
Caesar’s skills, starting with Figma
A skill is the written method Caesar follows for one kind of work. This quarter we rebuilt the Figma skill from the ground up. We tested Paper and Figma side by side and chose Figma, because that is where Lifely’s designers already work. The new skill has a brand guide for Lifely, Little Lifely and Cushie, layout libraries for each brand, and a page route: Caesar designs a page, checks it in a browser, then puts it into Figma as editable layers. Anyone on the team can now ask Caesar for Figma work without waiting for an approval. Around it, we rebuilt or added the skills Caesar uses every day.
Skills we rebuilt or added
- Marketing skills for hooks, static ads and advertorials now read from and write to the Little Brain.
- Xero: reliable finance reads, with clear rules on who can do what.
- Upfluence: Caesar can read the existing creator program.
- PO Allocation: Caesar can look up orders and shipments.
- The SCALE AI course library, searchable through Caesar.
- Browser testing on Lifely’s own servers, for checking store changes before they go live.
In numbers
- In August, Caesar could only see 92 of his 135 skills because of a hidden limit. We fixed it.
- The Little Lifely page library has 36 pieces, cut from the real store.
- Of four test pages made with the new route, two needed no changes and two needed one round.
- 52 people can now use Figma through Caesar without an approval.
Caesar on GPT-6 Astra
Caesar now runs on the latest version of OpenClaw with GPT-6 Astra. We moved him to GPT-5.6 in July, then to GPT-6 Astra in September, each time with a way back kept ready. Before that we made him steadier. The restarts that cut conversations in July are gone, his background jobs are fewer and lighter, and he moved to a server more than twice as big. We also added AI accounts so the team keeps working as use grows.
In numbers
- His server now has 16 dedicated cores and 61 GB of memory, up from 8 shared cores and 16 GB.
- His background jobs went from 82 to 55.
- In his first week on GPT-6 Astra, the daily health check found no conversation cut off during Sydney working hours, against 18 in the month before.
- 24 people used Caesar in one week of September, and their requests grew 21% from one week to the next.
- AI accounts went from 7 to 13.
Claude Code for the team
The team now has a second assistant beside Caesar. Claude Code runs in Lifely’s Discord and Slack: write to Caesar as usual, or tag Claude to reach Claude Code. It follows the same per-person access rules as Caesar, so nobody gets more through Claude than they have through Caesar. It started as a pilot for Chris in August and opened to the whole team in September.
Permissions, with David and Xero
Caesar works for the whole team, so who can see and do what matters. We reworked it from the ground up. Lifely’s Discord became private by default, so each team sees its own channels, and team leads now add their own people. For Xero, David told us Caesar was confusing who could do what, so we fixed it at the root: one Xero connection that renews itself, one simple read or write right per person, and only Stephane, David, Jake and Chris can grant access. Publishing a live store theme now needs a named approver too. People who had left were removed from everything we manage. Finally, we built a new permission system with clear levels, including a rule that keeps salaries private, and tested it beside the current one. It is ready to switch on.
Open: people.lifely.network (owners only)
In numbers
- 13 people read Xero through Caesar, and one finance team member can write.
- A contractor now sees 2 of Lifely’s 395 Discord channels.
- 5 people who had left were removed from Caesar, Discord, Slack and the team logins.
- The new permission system matched the current one on 5,350 checks, with no unexpected difference.
- Still open: switching Caesar to the new system, so the salary rule goes live.
Lifely on its own accounts
Several Lifely tools still ran on Stephane’s personal web address and accounts. On 28 September we moved Lifely’s live web tools to Lifely’s own address and Cloudflare account, with no planned interruption. Browser testing moved onto Lifely’s own servers, so no test machines are rented on Stephane’s account anymore, and competitor research moved to Lifely’s own collection account. A few pieces, like the Xero sign-in pages, follow as each owner confirms.
And a few more
- An AI briefing for investors, built at Chris’s request in August, ready to share once he approves its figures.
- The SCALE AI course library on Caesar: 310 lessons with transcripts, updated every week.
- Lifely’s Discord kept under its channel limit, with every used channel saved and searchable before cleanup.
- Tool invoices filed in the accountant’s sheet every week, without anyone chasing them.
- Caesar’s disk kept from filling up, with old files archived to Lifely’s Drive and checked before removal.
- Caesar and Cornelia each got their own Gorgias app, with double the request limit of a normal key.
Challenges
A system only works when someone owns it
Dispatch moved forward once the fulfilment team started using it in their day. PO Allocation, the warehouse connections and the Little Brain still need a named owner on the Lifely side.
Use grows faster than capacity
The team uses Caesar more every week, OpenAI paused new sign-ups to its largest plan in September, and the newest model needs more capacity per answer. We ran short more than once.
Some limits are outside our control
Cin7 cannot record dispatch outside Australia the way we need, and Upfluence only lets us read. That second limit is one reason we built the creator system ourselves.
Big facts surfaced late
We found the two-store split of Little Lifely sales days before the creator launch. One shared map of systems and owners, and a short monthly update, would bring these facts up earlier.
Q4 priorities
Lifely’s Q4 plan asks for the biggest quarter yet, with November as the peak. These priorities follow from that plan and from what Q3 left open.
PO Allocation that holds through November
A named Lifely owner, two weeks without needing an engineer, dispatch for every market, and the real dates shown on the store.
Decide the US warehouse connection, and switch on the stock scripts
With Rohit and YK, choose between repairing and replacing the US connection using the checker’s record, and switch on the Lifely and Little Lifely stock scripts in the store.
A creator flywheel that sells
Payouts built before the first one is due, recruitment ads aimed at creators who buy, the Upfluence creators moved across, and boosting creator posts with Allen.
Cornelia on nights and private messages
Drafts where the care team is thinnest, judged by the team on each ticket.
The Little Brain in front of the team
Its findings sent to marketing every week, and the advertorial for parents still looking for a solution tested with real ads.
The new permission system, and enough AI capacity
Switch Caesar to the new permission system, and size the AI accounts on real use inside the budget.
A short monthly update
One written update to leadership every month, and short notes to the team after each Caesar change.