A few friends asked how I work with agents, so I had one go through a month of my own sessions (Oct 7 to Nov 11, 2026) and pull out the uses that seemed worth sharing. Routine bug fixes are left out. Client and personal details are removed.
My default is Claude Opus 5.5 on high effort. I had extra Fable credits that were about to reset, so from Nov 8 to Nov 11 I put them on the biggest creative jobs I had.
For the release of our in-house meeting recorder, I asked for a demo video in a new style and wanted to see the feel of it before it was built. After one round of notes (hit the benefits and the UI harder, add voiceover, say it runs on Mac and PC), it rendered a 4K master with music and a 1080p web cut. Then it wrote the release page, the YouTube upload, an essay with better visuals, and the LinkedIn and X posts.
I asked for the feel first, so the notes cost one round instead of five.
For a Google Ads landing page, a separate session wrote a script for me to read to camera, fixed the lines that were hard to read cold, and reworked the ending around owning your own software and avoiding rising license costs. I recorded one take with retakes in it and dropped it in Downloads. It cut the retakes, added motion graphics, music and footage of our demo, re-centered my framing, uploaded it to YouTube, and gave the website session what it needed to build a video version of the page for an A/B test.
The only part I did was stand in front of the camera.
I asked for a conversion audit of the ad landing page and research on ad landing-page best practices. It implemented the changes, simplified a hero that had three competing calls to action, moved the phone number to the header, set up a session-recording project so we can watch real visits, and emailed my business partner a full list of changes to make on the ads side. Later, from two teammates' Slack feedback, it built three new features into the demo app so the page could speak to them, did a mobile review, and posted an update in the team channel.
One session covered the page, the demo product behind it, and the note to the team.
I gave it a PDF pitch deck and asked for something judges could click through on an iPad or the web, with no backend. It told me its plan first, I asked it to embed the app experience too, and it built and hosted the demo behind a password.
From a deck to a hosted demo in one evening.
Before the credits reset, I pointed it at the operating system my CNC shop runs on and asked for a thorough review: simplifications, speed, data quality. It ran sub-agents on separate areas (one only checked that the financial numbers were correct), made the improvements, then emailed the shop manager what changed and what it held back for him to decide. When he replied, it adjusted.
Use-it-or-lose-it capacity went into maintenance work that usually never gets scheduled.
We have a browser-based 3D model of a house we're remodeling that you can walk through like a video game. I recorded my notes as I walked it: furniture shouldn't block you, only walls; Shift to run; lofted ceilings; pocket doors; a rolling island; art on the walls; a hidden door from the powder room. When the context got full I asked for a fresh agent. A worker made six layout changes and added three tools, checked them in a headless browser and deployed. The next day I asked it to critique the model as an architect would, fix every issue, keep a visible change log in the sidebar, and check the design against the local building code.
One agent was the 3D modeler, the architect reviewer and the code checker. A separate lead agent checked the work.
For a campaign aimed at companies spending six figures a year on a large field-service software platform, it read a month of my calls on the topic and sent parallel research agents to find untried channels, including how to identify those companies. It emailed the plan to me and my partner. Then we talked through a small, high-end print magazine for about 100 owners, sent with a handwritten note and a QR code. It made the PDF. When I said it was hard to read, it rewrote it for clarity, aimed at a heating-and-cooling owner who has done well.
It reused a format that already worked for us (our print magazine) for a narrow audience.
It read my recorded calls and wrote a quarterly magazine for executives, aimed at a thoughtful long-form tone. Output was a print-ready 7x10 PDF with a cover, collage art, generated photos, charts and a mock coupon ad. It set up the printer order in the browser and answered the printer's color-proof email. The magazine site takes orders through Stripe, paid orders print shipping labels on my label printer automatically, and each article has an audio version that also goes out as a weekly podcast.
One person acted as editor, art director, e-commerce developer and print buyer for a physical product.
Each new feature of our internal system gets a short silent film. The agent builds the scenes as animated HTML with made-up data, renders to 60 fps video, generates music with ElevenLabs (I pick from takes), uploads to YouTube with thumbnails, and schedules LinkedIn. When a teammate sent a Loom critique, I had it fix the film-making bot itself, so later films improved too.
It wrote a word-for-word script and a B-roll shot list for my woodshop, put it on a one-page PDF, then took over my mirrored iPhone to load the script into a teleprompter app. After I shot it, it edited the footage.
It made contact sheets of cap designs (six at a time, then six variations of the one I picked), got quotes from four vendors in the browser, and handed the order to my assistant bot. A second model compared the vendor's mockup against the original design.
From three phone videos of a house and its floor plans, Opus 5.5, GPT-6 Astra and GPT-6.1 Terra each built the same 3D model and animated remodel from one shared brief. I kept the best one and sent it back for furniture, a walk-through mode and realistic lighting.
The agent drove Cricut's design app, found a 3D-printable pen adapter and printed it, and plotted handwritten cards. That was too slow, so it moved to printed 5x7 cards, each with its own QR code tracked in our system. It researched each recipient, checked any compliment against what they had actually said on podcasts, and wrote a short personal note.
I had Codex install an open-source Mac call recorder, then fork it into our own app. It has sign-in through our system, automatic upload, import of old calls, speaker separation and naming, and summaries in the format I like, with the transcription and analysis models swappable. For every build, a Codex computer-use agent clicked through the real app under strict rules (never record, never delete, never quote call content) and reported pass/fail with screenshots. Four chained Codex reviews checked the security of the changes.
I click elements on a page with a browser annotation tool, write a note on each, and paste the batch to the agent. I did several dozen rounds on a team capacity planner I built in a day.
For one landing page, three workers each followed a different design process from the same brief, and I picked the winner. Separately, the agent wrote our design system as JSON tokens plus an HTML style guide, so any future agent can be pointed at one link to match it.
When I need to make a lot of small calls (which hosting projects to delete, who gets a magazine), the agent builds a one-off HTML page with everything switched on by default. I switch off what I don't want and paste the result back.
A coordinator hands narrow tasks to workers in Orca with hard rules: don't push, don't write to the database, never print secrets. The coordinator re-runs the tests itself (about 2,300 in one case) before it trusts a worker's report.
A headless job reads all of yesterday's prompts every morning, finds lasting preferences and recurring friction, updates my skill and instruction files, and writes a short report.
A small model reads each AI session the team runs and assigns it to a client, internal work or personal.
When an agent figures something out (how to open a new chat in another tool, how to keep a repo's state between sessions, how to build a client brief site), I have it write that up as a skill so every agent can reuse it.
For a client's engineering hire, parallel agents screened a talent network, parsed 573 applicant emails, researched candidates in batches and read profiles, and returned one HTML shortlist.
I moved my standing bots onto my own Mac in OpenMausBot with Opus 5.5 as the default, a shared 1Password vault just for bots, and a phone-call skill any bot can use. Each one has a job and its own routines:
A scheduled agent reads every call, email, Slack message and repo change for each active deal and writes a report. I moved the report into our Mac app, and any comment I leave sends an agent to make the change. By this week it sends one worker per deal to apply fixes itself and only flags what needs me.
Each client gets a site with goals, a delivery map, who's who and proposed meetings. A daily sync pulls in new calls, email and Slack. Staffing someone on a project in our system now gives them the repo, the site, the Slack channel and the shared vault, and takes them away at roll-off.
For the shop's software, I paste a screenshot or email link to a feedback ticket. The agent fixes it, deploys, closes the ticket and emails the shop manager.
It built a tracked sales line that rings two people in turn and turns missed calls into CRM leads. When test calls hung, I had a voice agent call the line an hour later and report back.
Codex read one day of my Claude and Codex sessions and drafted an anonymized post about everything one person got done with agents. A second model rewrote it, and the agent charted the work against my meetings.
I saw a new browser-agent tool on X. The agent checked it for security problems, ran the same task four ways (Claude and Codex, each with and without it), wrote an HTML report, then turned that into an X article and replied in the original thread.
Weekly automations work on how our sites show up in Google and in AI assistants' answers. They show me a before and after first, then publish, add internal links from Search Console data, and track leads that come from AI referrals. For the shop it also built a blog and writes a post a week.
A Thursday automation arrives with sourced story ideas for Friday's issue. Drafts get pushed toward a plainer, more executive register, the list is checked, the send is scheduled, and an agent watches the live send.
A date night bot keeps track of what we have coming up, and my assistant bot puts a leave-by time on the calendar so the night doesn't get eaten by a late call.
| What | Used for |
|---|---|
| Claude Code, Opus 5.5 on high | Default for building, writing and orchestrating |
| Claude Fable 5.x | Big creative jobs: films, ads, demos, architecture critique, audits |
| Codex (GPT-6 Astra, GPT-6.1 Terra) | Computer use in desktop apps and admin consoles, reviews of Claude's work, some lead-agent work |
| Claude Haiku 5.5 | Cheap classification, like assigning sessions to clients for time sheets |
| Orca, Superset | Running parallel agents with a lead and workers |
| OpenMausBot | About 20 always-on bots with their own routines |
| 1Password service account | A vault just for bots, so they can sign in without asking me |
| ElevenLabs, image models | Music, voice and images for films and print |
| Retell, Twilio | Phone calls and tracked phone numbers |