CLAUDE LABJP
FORK — Claude Code 2.1.212 changes what /fork does: it copies your conversation into a new background session with its own row in claude agents, so you can keep working. The old in-session subagent is now /subtaskLIMITS — WebSearch calls are now capped at 200 per session by default, and subagent spawns get the same 200 ceiling, so a runaway search or delegation loop stops on its ownMCPBG — MCP tool calls running past two minutes now move to the background automatically, keeping the session usable. Tune the threshold with CLAUDE_CODE_MCP_AUTO_BACKGROUND_MSPLANFIX — Fixed plan mode auto-running file-modifying Bash commands such as touch and rm without a permission prompt or an SDK canUseTool callbackSONNET5 — Claude Sonnet 5 is running on introductory pricing of $2 per million input tokens and $10 per million output. After August 31 it moves to $3 and $15IPO — Bankers are reportedly lining up investor meetings for Anthropic ahead of a possible public listing as soon as OctoberFORK — Claude Code 2.1.212 changes what /fork does: it copies your conversation into a new background session with its own row in claude agents, so you can keep working. The old in-session subagent is now /subtaskLIMITS — WebSearch calls are now capped at 200 per session by default, and subagent spawns get the same 200 ceiling, so a runaway search or delegation loop stops on its ownMCPBG — MCP tool calls running past two minutes now move to the background automatically, keeping the session usable. Tune the threshold with CLAUDE_CODE_MCP_AUTO_BACKGROUND_MSPLANFIX — Fixed plan mode auto-running file-modifying Bash commands such as touch and rm without a permission prompt or an SDK canUseTool callbackSONNET5 — Claude Sonnet 5 is running on introductory pricing of $2 per million input tokens and $10 per million output. After August 31 it moves to $3 and $15IPO — Bankers are reportedly lining up investor meetings for Anthropic ahead of a possible public listing as soon as October
Articles/Cowork
Cowork/2026-05-29Intermediate

Three Weeks of Letting Claude in Chrome Tune My AdMob Mediation Priorities

I let Claude in Chrome handle the reordering of AdMob mediation priorities for three weeks. Here is how I set the threshold, the three prompt changes that actually mattered, the numbers, and the parts I deliberately kept under human control.

Claude in Chrome14AdMob12MediationeCPMIndie Developer9App Operations2Cowork33

On the morning of May 3, eCPM dropped roughly 20% across several of my apps at once. One network's unit price slumped through the daytime and recovered by evening — the kind of swing you can partly recover by reordering priorities. But opening the console and reordering app by app would have consumed the whole morning. I ended the day without touching it.

Money I could have recovered, skipped because of the labor around the decision. That feeling is what pushed me to try Claude in Chrome. Running everything solo, I can't hand the first few days of a month over to a console.

What follows covers May 8–28, twenty-one days. More of it is about the lines I refused to cross than about what worked.

Where the weight actually sits

Priority tuning mixes two kinds of work with very different shapes.

WorkWhat it isPer appAs app count grows
DecidingChoosing which network sits where1–2 minOften one pass covers several apps
ApplyingOpening, navigating, dragging, saving8–10 minScales linearly with app count

Roughly 80% of those ten minutes is the applying side — and only the applying side grows with the number of titles. That asymmetry is what lands on you at the start of each month.

The applying side is exactly what Claude in Chrome can take: open the page, read it, operate it. Reframing this as separating decision from execution — rather than delegating the decision — made the design fall into place.

Where I drew the line

The split I settled on after three weeks:

What I keep

  • The threshold that triggers a reorder (in my setup, only when the same-day-of-week change exceeds ±10%)
  • The initial priority list at the start of each month (when policy shifts, I draft it by hand)
  • The final review and the actual click on Save

What Claude in Chrome handles

  • Opening the AdMob report screen and pulling the last seven days of eCPM per app per network
  • Listing the combinations that crossed the threshold
  • Navigating to each app's mediation settings, dragging priorities into a proposed order in draft state, and capturing a screenshot
  • Stopping short of Save, then posting a Slack link for me to review

The Save click is never delegated. AdMob priority order maps directly to revenue, and a misordered list saved at the wrong moment can leak revenue for hours before anyone notices. Stop before the irreversible action — that's the one line I won't move.

One scoping note: this framework assumes waterfall mediation. Networks running on bidding don't share the same notion of priority, so I excluded them. For apps where both coexist, I explicitly instruct Claude in Chrome to extract only the waterfall rows. Without that, it will cheerfully pull the bidding rows too — cheerfully enough that it took me a while to notice.

The three prompt changes that mattered

The first week, draft quality was all over the place. Looking back, only three changes actually moved it.

1. Re-verify the navigation path from scratch, every run

I started with "open the mediation settings screen." That let it act on a remembered page structure, and the smallest UI change sent it wandering. I rewrote it to: "Follow Mediation → Ad unit → Edit in order, reading the actual on-screen label at each step before proceeding. If a label doesn't match expectations, stop and report." The wandering stopped.

2. Make it write out the numbers as a table before proposing anything

Ask only for a reorder proposal and you can't audit the reasoning. Spelling out the sequence — output the eCPM table first, then flag the rows over threshold, then propose the new order — cut my review to about ten seconds of eyeballing. Reversed, the conclusion arrives first and the reasoning reads like justification, which means re-checking all of it anyway.

3. Write the stop condition as "stop here," not "don't press that"

"Don't press Save" means it won't press Save but will still poke around nearby. Changing it to "the moment the Save button is visible on screen, stop operating, take a screenshot, and report" made the behavior crisp. Naming the stopping point beats fencing off the forbidden one — the same thing I found while writing Claude Code vs. Claude in Chrome: Where I Draw the Line in Daily Ops.

The numbers from three weeks

Between May 8 and May 28 (21 days), Claude in Chrome surfaced 14 drafts. I approved 11, held three to sleep on overnight, then approved two of those and discarded the last one.

Among the 11 approvals:

  • Six were small adjustments dropping a single network by one slot
  • Three swapped the top two networks
  • Two introduced new country-level splits (Japan, US, Other)

The headline eCPM lift was modest — about +4.8% versus the prior month across the 21 days. Not dramatic, but framed as recovering the opportunity cost of days I wouldn't have touched the console at all, it fits. At my scale the monthly difference is in the low-five-figures of yen. The bigger change is that the first few days of each month came back to creative work.

One caveat on that number: I can't cleanly attribute the +4.8% to reordering alone. The floor-price revisions I describe in Three Weeks of Tuning AdMob Floor Prices with Claude in Chrome were running over the same window. Isolating the effect would mean moving one lever at a time, but this isn't a number worth pausing revenue to measure. As an operations record, this granularity is enough.

Friction I did not see coming

Convenience brought its own walls. The largest was tracking AdMob's UI changes. Two small layout updates landed during the three weeks, and each time Claude in Chrome lost its path to the mediation screen. The navigation re-verification above stabilized it, but in seasons when AdMob is shipping rapid updates, stepping back to manual for a stretch is the safer call. My triage steps for when it stalls are in Claude in Chrome Not Working? A Practical Troubleshooting Guide.

The second was reporting lag. Same-day data lands a few hours late, so the 9 a.m. run only sees numbers through yesterday. That's why I anchor the threshold to same-day-of-week comparisons. Comparing against the current day so far produced too much noise and too many false drafts. For ad inventory whose demand moves by weekday, week-over-week on the same weekday is simply the more honest baseline. For the same reason, I moved fill-rate judgment onto a weekday basis in Catching AdMob Fill-Rate Drops in the Morning with Claude in Chrome.

The third was my own approval fatigue. A draft would arrive in Slack and I wouldn't always be at the keyboard. One night I left a draft for the morning, and by then the market had shifted enough that the underlying data was stale. Now drafts are only generated at 9 a.m. and 5 p.m., and I commit to reviewing them within those windows. I was lucky to drop the more-drafts-is-better instinct early. Drafts you can't approve are indistinguishable from noise.

If you want to try the same setup

For anyone starting from scratch, here's the order I had to rediscover during that first week.

  1. Run one app for one week first. Widen to everything at once and a single UI change breaks all of it simultaneously. Getting navigation stable on one title is faster in the end
  2. Start loose on the threshold, then tighten. I began at ±5%, drowned in drafts, and settled at ±10%. Going the other way (strict first, then loosen) produces no drafts at all, which leaves you nothing to tune against
  3. Pick your approval windows before you pick the run times. Do it the other way around and drafts will pile up
  4. Keep drafts and outcomes reviewable afterward. I just use the Slack thread as the log. Seeing your own decisions wobble is what justifies moving the threshold

What I won't delegate, and what comes next

Reordering can be delegated, but the structural side of mediation — adding new networks, switching bidding modes, revising price floors — stays manual. Those moves happen a few times a year, don't justify the automation overhead, and lean on longer-term business judgment that recent eCPM alone cannot reach. I wrote separately about the practical side of adding networks in Adding Three AdMob Mediation Partners With Claude.

Next, I want the decision context attached to each draft to get gradually richer: a 30-day fill-rate trend, a short note on competing networks' recent rates — anything that lets me approve in less time without thinking less carefully. I plan to revisit the same setup at the six-month mark.

If I had to compress three weeks into a sentence: what I delegated wasn't the judgment, but the distance between the judgment and the work. If you try one thing today, make it one app, ±10%, a single 9 a.m. window. Whether navigation holds there is your answer about whether to widen.

If you run AdMob mediation as a solo developer, I hope a little of this is useful. Thanks for reading.

Share

Thank You for Reading

Claude Lab is ad-free, supported entirely by members like you. We publish practical guides daily with implementation code, benchmarks, and production-ready patterns. If you've found it useful, we'd love to have you on board.

  • Copy-paste ready implementation code
  • New advanced guides published daily
  • $5/mo or $10 for lifetime access
View Membership →

If you found this article helpful, a small tip ($1.50) would mean a lot to us. Your support helps keep this site ad-free and covers server and hosting costs.

Related Articles

Cowork2026-05-28
Folding AdMob and Crashlytics into One Morning Check via Cowork Scheduled Tasks — Two-Week Notes
I had been checking AdMob fill rates and Crashlytics surges in two separate dashboards each morning. I folded them into a single Cowork scheduled task. Here are my two-week notes, with the numbers and the friction I ran into.
Cowork2026-06-04
Where the human still has to press the button — lessons from delegating ad ops to Cowork
From actually delegating repetitive ad operations — AdMob bidding-partner sign-ups and floor tuning — to Cowork, here is the boundary I found between what AI can fully own and the single moves a human still has to make by hand.
Cowork2026-05-06
Claude in Chrome Not Working? A Practical Troubleshooting Guide
Buttons that do nothing, pages that never load, tasks that report success but change nothing. Symptom-by-symptom fixes, plus the 42-run log where adding waits and verification steps moved completion from 62% to 95%.
📚RECOMMENDED BOOKS
Build a Large Language Model (From Scratch)
Sebastian Raschka
LLM Dev
Prompt Engineering for LLMs
Berryman & Ziegler
Prompting
AI Engineering
Chip Huyen
AI Eng
* Contains affiliate links
See all →