Your agent already knows how to propose a listing change. Install the specialists, put the co-founder review on a schedule, and every Monday there is a pull request: one change, the brief with the numbers behind it, a post and a video if they help, and when it will be judged. Nothing reaches the stores until you merge it.
Why once a week
Learning is limited by measuring, not by how often you change things. A push is judged on day 7 and day 14, and two changes to the same store in that time are measured over the same days, so neither result means anything. Most App Store fields also wait for a version and review. One change per store, about once a week, is as fast as the playbook can learn.
Only where nothing is being measured
status knows which stores have a release still being measured. The ones without are free, and the agent only proposes for those:
$ npx @appeeky/spec statusdue 2026-10-29T08-00-00Z-app-store release · app-store · next 2026-11-05es-MX keywords: Win US searches in Spanishmeasuring 2026-10-31-screenshots-lead-with-baby experiment · next 2026-11-14Free to change: Google Play
When every store is busy, the run ends there. That week has nothing to propose, and that is the right answer.
Six specialists read first
npx @appeeky/spec agents install adds six read-only specialists and the /appeeky-cofounder command to .claude/ (or opencode with --for all). Commit them. The review starts them in parallel: ASO, reviews, ads, releases, UGC and social each read their part of the repo and report back. Then the co-founder decides.
$ claude -p /appeeky-cofounderaso “fan sounds” unranked, asked for in 9 reviewsreviews 4.6★ · 38 new · 6 without a replyreleases 3.2 shipped fan sounds, not in the listing yetsocial TikTok led: 18.2k views on the speed hook→ Google Play short description leads with fan soundsappeeky/briefs/2026-11-09.md · .appeeky/proposal.md
Together they read:
appeeky/loops.md
what is being measured and what is free
appeeky/playbook.md
what worked and what lost, so it is not repeated
appeeky/competitors/changes.md
what competitors changed, and where they climbed
appeeky/reviews/summary.md
the words users use that the listing does not
.appeeky/data/
conversion, installs, revenue and ads, as files
coverage
which targets the listing does not cover yet
It edits the listing files, runs check and diff, and writes .appeeky/proposal.md: what changes, why, with the numbers it used, which metric should move and when it is checked.
The workflow
The spec repo has the full workflow and the prompt it gives the agent in examples/weekly/. Copy both into your app's repo. The shape:
on:schedule: [{ cron: "0 7 * * 1" }] # Mondaysjobs:propose:steps:- uses: actions/checkout@v4- run: | # what the specialists readnpx -y @appeeky/spec data pull; npx -y @appeeky/spec reviews pullnpx -y @appeeky/spec competitors pull; npx -y @appeeky/spec social pull; npx -y @appeeky/spec ugc pull- id: freerun: echo "stores=$(npx -y @appeeky/spec status --json | jq -r '.free | join(" ")')" >> "$GITHUB_OUTPUT"- run: | # the co-founder review, from .claude/commands/claude -p "Stores free to change: ${{ steps.free.outputs.stores || 'none' }}.$(cat .claude/commands/appeeky-cofounder.md)" --allowedTools … --disallowedTools "Bash(*push*)" …- run: gh pr create --title "$(head -1 .appeeky/proposal.md)" --body-file .appeeky/proposal.md
Secrets: CLAUDE_CODE_OAUTH_TOKEN (from claude setup-token, it runs on your Claude plan) or ANTHROPIC_API_KEY for the agent, and APPEEKY_API_KEY or ATTRIBUTION_TOKEN for the numbers. The agent may edit, check and diff, but every command that sends something (push, social push, reviews push, ugc generate --yes) is blocked. To keep your own version of the review, copy it to .github/appeeky-propose.md.
On your own machine
No GitHub Actions? examples/weekly/weekly.sh does the same from cron on a laptop, a server or a small VM, with git, gh and the agent installed:
0 7 * * 1 cd ~/apps/sleep-sounds && sh weekly.sh >> .appeeky/weekly.log 2>&1
It does not need to run all the time. Once a week, for a few minutes, is enough. The docker image does it too: set WEEKLY_AT with the agent image, as in Run it on your own VM.
You stay in charge
The agent never pushes
It proposes. The push runs when you merge, like any other listing change.
Read the why
The pull request says which numbers it used. If they are thin, ask for more or close it.
Close what you would not ship
A closed pull request costs nothing and the store never saw it.