A site that gets better without becoming unpredictable.
This plan turns a Cloudflare-hosted website into a cautious learning loop: observe what happened, propose one change, test it in the shadow, and only then promote it.
Planning surface · proposed flow, not live telemetry.
01 / The thesis
Self-improvement needs a shape.
The useful question is not “can the site change itself?” It is “can each change be small, measurable and reversible enough to trust?”
Design principle
Optimize the loop before optimizing the page.
The website remains the product. RSI is the operating rhythm around it: a daily habit for learning from real behavior without handing the keys to an unbounded agent.
01
Improve in small steps
A daily loop should make one legible change, not rewrite the site in one leap.
↗02
Keep the old world intact
Every release keeps a known-good snapshot and a fast path back to it.
↗03
Make evidence visible
The proposal, checks, decision and outcome all live in the same trail.
↗
02 / The loop
Five moves. One controlled orbit.
Select a phase to inspect its job, output and boundary. The center stays intentionally boring: one small step at a time.
Operating model
A daily release should leave evidence behind.
03 / The daily rhythm
Wake once. Learn all day.
A predictable daily window makes the system inspectable. The schedule is simple enough to explain in one sentence and strict enough to prevent drift.
Cloudflare Cron Trigger
0 3 * * *
Proposed schedule · one run per day · UTC. The trigger wakes the workflow; it never publishes by itself.
dailyUTCguarded
Wake the loop
Cloudflare Cron Trigger starts the daily run.
scheduler
Read signals
Pull the latest metrics, errors, feedback and last decision.
observe
Draft one bet
Write a bounded hypothesis and generate the smallest useful diff.
propose
Shadow test
Render the candidate, run checks and compare it with baseline.
verify
Decide
Publish, hold or roll back. A human gate stays available.
ship / hold
Write the log
Record the outcome and carry the learning into tomorrow.
learn
change budget1 / dayone bounded candidate
human gate100%veto stays available
rollback target< 5mreturn to baseline
decision trail∞every run is logged
04 / The guardrails
Autonomy with a handbrake.
A passing score is not permission to improvise. It is a narrow gate with a clear “stop” branch.
Gate
What it watches
Pass rule
On fail
Status
Data quality
Signals are fresh and complete
≥ 95% expected inputs
Block proposal
defined
Scope
The diff is small and bounded
≤ 1 page / 1 hypothesis
Ask for review
defined
Regression
Candidate is no worse than baseline
No critical check fails
Hold + preserve baseline
defined
Human override
A person can stop promotion
Always available
Veto / roll back
defined
05 / A bounded experiment
The unit of progress is a candidate.
A good proposal reads like a ticket, not a magic trick. It states the why, the smallest diff and the moment it should be discarded.
EXP-014 / candidate
Shorten the first-screen copy so the plan is understood in two scrolls.
needs review
signal
Visitors reach the loop section but skip the opening explanation.
hypothesis
A shorter framing block increases loop-section engagement.
diff budget
Copy only · no data model, route or runtime change.
evaluation
Compare scroll depth and completion against the last 7-day baseline.
06 / Runtime blueprint
A small surface area for a big idea.
Keep the first version legible: edge delivery, one scheduler, one evidence trail and one release guard. Expand only when the loop earns it.
CF
Site runtime
Serves the versioned site at the edge.
proposed boundary
CR
Cron Trigger
Wakes the loop once per day.
proposed boundary
LOG
Decision log
Stores the evidence and outcome trail.
proposed boundary
SAFE
Release guard
Protects baseline, review and rollback.
proposed boundary
07 / What happens next
Prove the loop before you widen it.
Start with a dry run, one human-reviewed candidate and a complete decision log. If the loop can explain itself, it is ready for the next level of autonomy.
01Connect the site’s real signals and define the first baseline.
02Run the daily schedule in shadow mode for one week.
03Review held, shipped and rolled-back candidates together.
04Only then decide which boundary can safely move outward.