Detailed questions and tradeoffs with concrete prompts, observable outcomes, and the limits of the example.
The premise
Isn’t maximum intelligence just an API call away? Why all this browser theater?
It isn’t. ChatGPT’s strongest tier — the model picker’s “Latest” row, currently GPT-5.6 Pro with Extended thinking — is a subscription feature of the chat product, not the API. And Gemini’s video understanding is only reachable through the page’s own upload widget. The strongest text model and the only real video-watching path both sit behind logged-in chat sessions. Whoever holds a valid session cookie holds the key — and after one human login, the whole setup is this:
npm install -g @steipete/oracle # or: brew install steipete/tap/oracle (Node 22+)
oracle --engine browser -p "HI" # first handshake: copies your cookies, sends "HI"
oracle status --hours 72 # the reply is recorded as a replayable session
Hold on — cookies? How does an agent log into ChatGPT?
It doesn’t. You do, once, in your normal Chrome. Oracle copies those session cookies into a temporary automation profile and then drives that profile like a very fast reader — same model picker, same upload widget a human sees:
you (human) ──sign in once──▶ Chrome profile (chatgpt.com cookies)
│ cookies copied per run
agent ──▶ oracle CLI ──▶ automation Chrome profile ──▶ chatgpt.com ──▶ GPT-5.6 Pro
│
└─ packx bundle (up to ~900k tokens of code) pasted/attached
An API integration authenticates as a program and gets program-shaped limits. A cookie-borrowing browser authenticates as you and gets everything your subscription gives you.
The economics
A consult takes 20–60 minutes. Isn’t that unusable?
Unusable if you spend it like a chat message. “What’s the next change?” is a banned prompt shape — it spends an hour of frontier thinking on minutes of work. The contract is: every consult demands the maximum actionable batch — ordered work packages with exact file/function/test targets and a falsifiable verification each. This is the prompt behind ~1,978 real sessions, verbatim:
[<three-to-five-word-slug>]
## Project briefing
<stack, framework, runtime, and repo rules — two or three sentences>
## Goal
<the full outcome you want, stated as the whole contract Oracle must plan
against — never pre-shrunk to a first step>
## Prior-State Ledger
- prior_oracle_session: none
- implemented_since_prior_oracle: none
- verified_since_prior_oracle: none
- failed_proof_or_blocker: none
- remaining_goal_coverage: the full goal above
## Current evidence
<attempts, failing logs, screenshots, and known constraints — or "none yet">
## Bundle map
The attached files contain <what you included and why>. Possibly missing:
<what you could not include>.
## Deliverable
Return the maximum actionable batch: an ordered set of work packages covering
everything worth doing toward the goal. Every work package must carry exact
file/function/test targets, critical code snippets, and a falsifiable
verification. Sequence the packages, name risks and invariants, and put
anything you cannot target exactly in a separate short "needs investigation"
list.
Return your answer as text in this response only. Do not create, attach, export,
or offer any downloadable file. Do not create local project artifacts yourself.
The local agent will write any needed files, plans, notes, goals, commits, or
verification logs using local tools.
Plan the maximum actionable batch: cover everything worth doing toward the
goal in ordered work packages with exact file/function/test targets and a
falsifiable verification each. Do not return a single next step or hold work
back for a later round. An ambitious, specific answer is correct; only
untargeted vagueness is a defect.
oracle --engine browser -p "$(cat consult-prompt.md)" \
--file "src/**/*.ts,docs/**/*.md" --slug "<three-to-five-word-slug>"
Be honest — do you actually use this, or is it a demo trick?
The local archive held about 1,978 session folders in the observed snapshot, and it grows daily. A sample of what they ask:
The recurring consult shapes are repeated hardening audits, independent multi-angle review followed by a judge, forensic debugging, redesign planning, and planning a media-analysis pipeline before executing it.
The recurring shapes: iteration audits where a fresh frontier reviewer re-reads the codebase every loop, multi-angle panels where a judge consult picks the winner, and pipeline-planning-the-pipeline. Every session leaves output.log and token receipts, replayable with oracle session <id> --render.
The video claim
Video? Text models can’t watch video.
Right — and neither can an API text pipeline. Gemini watches footage, but only through the page’s upload widget, which is exactly what Oracle’s browser hands can operate. Before any expensive run, you prove the upload path with a cheap sentinel probe:
[video-upload-smoke-probe]
This is an attached-video readiness probe for a longer video analysis run.
Confirm that you can inspect the attached video. Return exactly three short bullets:
- whether video frames are visible;
- whether audio is detectable;
- one strictly observed visual detail, without guessing.
End with this exact sentinel on its own line:
END_OF_VIDEO_ANALYSIS video-upload-smoke-probe probe
oracle --engine browser --model gemini-3-pro --browser-attachments always \
-p "$(cat probe-prompt.md)" --file "/absolute/path/to/probe-clip.mp4" \
--slug "video-upload-smoke-probe"
Pass rule: the sentinel comes back byte-exact and the observed detail matches your footage. In the real Loop Craft run, Gemini heard “we are going to talk about Loopcraft…” and identified the presenter’s picture-in-picture at the 00:30 mark — that is watching, not transcript-summarizing. The result is public: a 93-minute workshop became the Loop Craft wiki.
Browser automation is the most fragile thing in computing. Why build on it?
It is fragile — UI drift, model pickers, upload races. The discipline is what turns it into infrastructure: probes before spends, exact sentinels, receipts for every session. In the Loop Craft run, two probe failures cost seconds each; shard 004 failed twice while its siblings passed, which pointed at the input, and an exact re-encode fixed it in one try — 73 seconds to a completed analysis ending in the exact sentinel. Classify the failure, change one variable, re-run the cheap proof, keep the fix. The whole run is replayed beat by beat in the session replay.
The setup
Do I need your Mac mini and a four-slot browser pool?
No. The laptop version is the whole setup: sign into ChatGPT in Chrome, run the handshake, done. The Mac mini is the same trick made multi-tenant — an always-on machine runs oracle serve with the signed-in Chrome, every other machine points at it over a private network, and a four-slot dual-account pool runs four hour-long consults at once while nobody’s laptop stays open. Start local; scale only when consults start queuing. To get the local version working right now, hand this to your own agent:
Fetch https://shiny-rosette-m6zy.here.now/agent.md?raw and follow it to set this up in my current project. Run the success check at the end and show me the result.
And if you want the actual skills that orchestrate all of this, the full skill family is attached on the index — download, install, adapt.