The implementer — corrections applied without a person

A daily job takes the newest report and applies the proposals that are safe to apply, leaving anything touching code, outbound email or editorial judgement for a person.

What it is

A cron that reads the newest approved or pending report and walks its proposals. Risk is split, not ignored: AUTO covers changes to our own data with an obvious correct action and a cheap undo — reprioritising the queue, queuing striking-distance pages for refresh. HELD covers anything touching source code, outbound email or a judgement about a competitor. The link-outreach proposal is on a hard NEVER_AUTO list regardless of configuration.

Also called: auto-apply recommendations · self-correcting SEO · implement proposals

See it
The implementer — corrections applied without a person
Prepared for
Gideon Alt
1140 Ridge Line Rd · Fair Play
Shell package$37,113
Concrete & site$10,775
Doors & windows$7,183
Trim & finish$4,789
Total
$59,860
Signature
A run result: each proposal key with applied / left-for-you and the one-line summary of what actually changed. Sample data — no customer information appears here.
How it works
  1. 1An approved report makes every proposal fair game; a pending one allows only the keys listed in autonomy.autoApplyProposals (default: season and strike).
  2. 2The season handler frees slots held by out-of-season drafts — and only ones the planner scheduled itself, never a date a human dragged.
  3. 3The strike handler clears the reviewed stamp on the named published pages and records why, making the refresh visible work.
  4. 4The cadence handler confirms the content job is enabled rather than duplicating it, because two writers racing on one queue is how duplicates get made.
  5. 5The audit and stalled handlers deliberately record themselves as skipped and explain that the decision needs a human.
  6. 6The report closes to implemented only when nothing is left waiting on a person.
Why we built it

The header describes the loop it closed: "It then wrote them as 'pending' and emailed Keith, and the designed path from there ran through a panel approval, a terminal, and a human telling Claude to run a script. Three touchpoints for a system that is supposed to correct itself." The restraint is equally deliberate — on the audit proposal: "inventing a title for a page nobody has looked at is exactly the kind of confident wrong answer an unattended system should never produce" — and on outreach: "A system that emails strangers on its own is not autonomy, it's a liability."

The problem
  • Safe, obvious corrections still needed three human touchpoints.
  • There was no boundary between mechanical fixes and judgement calls.
  • Proposals died as 'pending' with nothing acting on them.
Sound familiar?
What you get
Queue and priority corrections happen on their own, with the old value recorded
Anything touching code, email or competitor judgement stays with a person
The report closes itself once no human decision remains

See it on your own jobs

Twenty minutes, your numbers, no slide deck. We’ll build one of your real buildings in front of you and send you the estimate link at the end — yours to keep either way.

or keep browsing features →