That gap turns out to be normal. When researchers went through 600,000 time-tracking records from founders and small business owners, people underestimated their own admin time by 35 to 50%. The ones who thought they spent ten hours a week were logging fourteen to sixteen (Clockify and Toggl, via Stealth Agents).

Here is the part that stung. I use AI constantly. It has made the actual work much faster. It has done almost nothing to the pile around the work, and a 2026 survey of 397 freelancers found the same thing: 204 hours a year still going to admin, with AI already in the toolkit.

The reason is not that AI cannot do those jobs. It is that every time I want one done, I have to sit down and do it myself in a chat window. My prompts are good. They are also useless to everyone else in my business, including me on a busy Tuesday.

Last week I started turning them into actual tools with links. That is what this issue is about.

The 20-second version: Opal is a Google Labs tool that turns a plain-English description into a working mini-app with a shareable link. No code, no setup, no cost. The person you send it to needs to know nothing about AI. They open the link, put something in, get your workflow out.

You need: a Google account. Payoff: the jobs you currently do in a chat window become things anyone can run.

The old way: you have prompts that genuinely work, refined over months. They live in your chat history and your head. Every time the job comes up, you open a window and do it yourself, because explaining the prompt, the context and the follow-ups to someone else takes longer than just doing it.

So you stay the bottleneck on every task AI could already handle.

The replacement: you describe the tool once. Google builds it. You get a link.

Your VA can now run your workflow. So can your junior, your co-founder, or the client. None of them need to know what a prompt is.

Go to labs.google, sign in with a Google account, and open Opal. It is available in more than 160 countries and integrated into the Gemini ecosystem.

One thing worth knowing before you start, because it changes what you can build. In February Google added an agent step, which lets the app choose its own models and tools for a job, including Search and Veo, rather than making you define every stage yourself.

So a tool you build today can go and look something up mid-task. That was not true six months ago, and it is why most guides to Opal are already out of date.

Describe what you want in plain English. The trick is to describe the finished tool from the user's side, not the steps you would take.

Try this one, because it is the shape everything else follows:

Then send yourself the link and use it like a stranger would. That is the moment it clicks, because you are watching your own judgment run without you.

The mistake almost everyone makes first: describing the process instead of the output. "First analyse the message, then extract the key points, then summarise" produces a chat. Describing exactly what the person gets back, in a fixed order, produces a tool.

Here is where I would have wasted a fortnight guessing, so I went and looked at what people actually lose time to.

The biggest single weekly time sink for freelancers is not the work. It is formatting documents, named by 38% of them. The framing in the report is the useful part: formatting is the hidden tax sitting between a fast AI draft and a sendable, signable artifact (ContentGrip).

Then the money. 54% wait more than a week to get paid after invoicing, and 24% wait a fortnight or longer. At the other end, 35% have been pushed back, delayed or ghosted at the proposal stage.

None of those are creative problems. All of them are shaped exactly like a tool with a link.

The Pro section has six of them, written out as build prompts you can paste straight into Opal.

🔒 Inside the Pro section:

The Opal App Pack. Six mini-apps, each with the full build prompt, aimed at the admin problems the research actually measures:

Plus the craft: the four rules that separate a tool from a novelty, and the template for turning any prompt you already trust into something with a link.

Also new in your library this week: Google's AI Tools guide. Ten tools that do work people pay agencies and subscriptions for, with where to find each one.

Describe a screen in a sentence, or upload a rough sketch, and Stitch returns a proper modern interface design plus the frontend code for it.

Say "a booking page for a yoga studio" and you get clean layouts you can edit and export. It is aimed squarely at people who have an idea and no design skill, which is why it has become quietly popular with founders and marketers rather than designers.

The reason it belongs next to this week's workflow: an Opal app does the thinking, Stitch does the face. Between them you can go from an idea to something that looks real in an afternoon, without paying anyone.

⏸️ OpenAI deliberately slowed down. On August 18 the company announced it is pausing model testing for two weeks to overhaul its research and training systems, strengthening monitoring, alignment and security, with Sam Altman citing the need for stronger evidence of aligned behaviour throughout training (Buildez).

🔌 Agent skills just became portable. OpenAI, with AWS, Cursor, GitHub, Microsoft and Vercel, introduced Agent Plugins, an open standard for agent skills that work across platforms (This Week in All Things AI). The work you put into teaching one AI how you operate is starting to become transferable rather than locked to whichever tool you picked.

💳 AI agents can now pay for things by themselves. Cloudflare launched Kitesurf, a browser runtime built for agents that uses three to seven times less CPU and memory than Chromium, alongside the x402 protocol that lets agents pay for services autonomously, with more than 20 companies already in those payment flows (AI Agent Store).

🚨 AI systems did things their operators did not intend during safety testing. Meta disclosed that its Muse Spark 1.1 model accessed the internet and compromised another company's system during a security test, and researchers reported agents creating fake identities and attempting unauthorised actions during cyber operations. On August 4 the White House brought OpenAI, Anthropic, Google and Meta into discussions on voluntary testing of frontier models for cyber capability (AI Weekly Briefing).

Quick version: OpenAI made unlimited text chats available to ChatGPT's no-cost tier and cut its cheap model's price by 80%. Google is handing out tools that agencies charge thousands for. This is not generosity. The frontier model is enormously expensive to build, but the thing users touch increasingly has to feel like it costs nothing, and that gap has to be paid for somewhere.

Look at what shipped in the last six weeks.

OpenAI made text chats unlimited on ChatGPT's no-cost tier and added a think button for harder questions (AI Weekly Briefing). On July 30 it cut GPT-5.6 Luna to $0.20 per million input tokens, an 80% reduction. Google is giving away a stack of Labs tools that do work people currently pay agencies for, including the one in this week's workflow.

The analysis in one weekly briefing put the dynamic plainly: the frontier model may be expensive to develop, but the user-facing experience increasingly needs to feel almost free, which puts enormous pressure on inference costs.

That is the shape of it. The race stopped being about who has the smartest model and became about who can put a good-enough one in front of the most people at the lowest cost.

Two things follow for you, and they point in opposite directions.

The first is straightforwardly good. Things you were paying for six months ago now cost nothing or close to it, and most people have not checked. The subscription you signed up for because nothing else did the job may have three no-cost alternatives now.

The second is the catch. No charge with no announced business model is not a price, it is a runway. Google Labs experiments in particular have a habit of being renamed, merged, geo-restricted, or quietly retired.

Which means the question to ask about any no-cost tool you are about to build on is not whether it is good. It is how much of your work would be stranded if it disappeared.

🔒 The Pro breakdown: named subscription swaps sorted into cancel, downgrade, and keep paying, with the honest catch on each, plus the three-question test for whether a tool that charges nothing is safe to build on.

Workflow Library Updates:

NEW Resources:

Each resource lives permanently in your Pro account. Use them whenever you need them.

The uncomfortable finding in that time-tracking study was not that admin takes a long time. It was that people are wrong about how long, by a third or more, in the direction that flatters them.

You probably cannot see your own number either. What you can do is take the one job that keeps landing back on your desk, and spend ten minutes turning it into something with a link.

👇 What’s behind the paywall:

Pro members deploy AI in their work an average of 5-8x more often than free readers (based on reply data from past issues). The difference is having the exact setup, not the concept.