If there is no mechanism to route a recommendation after it is made, it will be raised repeatedly without being implemented. An observation does not automatically become an action; a queue needs to absorb it.
Raymond’s explanation
A two-AI review ritual in Week 22 of 2026 identified a friction pattern Raymond called “you told me, but it never got implemented”: the same recommendation appeared repeatedly in daily logs, but because there was no mechanism to move it from “suggestion” to “done,” it was forgotten after being raised and then raised again. The observed numbers were poor: the same tool setting that should have been written into a Skill was brought up four times within two weeks (during which the same busywork was repeated); another infrastructure item was brought up three times but still not implemented.
The problem was not that the AI lacked knowledge or that Raymond did not want to do the work. It was that a “suggestion” has no natural home: it is not a task, a project, or knowledge, but just an observation. The solution is to give it a home: absorb suggestions into an actionable backlog, automatically raise priority based on the number of times an item appears, and leave completed items in Done rather than deleting them (so they are not mistaken for new recommendations next time). The value of a review is not “knowing what happened,” but whether its suggestions have somewhere to go.
Current implementation
- A work queue that must be read before the weekly review: the rule is that the same suggestion appearing two or more times automatically enters the queue as P1; at three or more appearances it rises to P0; at four or more it is marked RED ALERT and must be handled that week. A P0 that has been untouched for more than 30 days must be downgraded or broken into smaller parts rather than left to rot.
- Backlog downgrade for the TODO publishing list: content candidates that have not been selected as featured and have gone unchosen for three consecutive weeks are automatically downgraded—the same principle in reverse, so unwanted recommendations naturally exit instead of occupying a slot forever.
When to use it
- When a review, meeting, or AI conversation produces a recommendation that “X should be done,” first ask who will read it again and when.
- When the same person (or AI) says the same thing for a second time, treat it as a signal that there is no routing mechanism. Do not say it again; create a queue first.
- When designing any automated inventory (daily logs, weekly reports, health-check scripts), connect its output to an entry point that will be processed; otherwise, the inventory is only a ritual.
Counterexamples and boundaries
- Routing does not mean everything has to be done. The other function of a queue is to say no gracefully—downgrade and exit rules matter as much as escalation rules.
- Do not create a separate queue for every type of suggestion. Keep one main entry point for the same work; multiple queues themselves become new “suggestions no one reads.”
- One-time decisions do not need a routing mechanism; a daily-log entry is enough. Only “actionable suggestions that recur” are worth placing in a queue.
Origin
A two-AI review ritual in 2026-W22 first named the friction “you told me, but it never got implemented” and proposed a Work Queue mechanism; it was implemented as a standing action queue on 2026-05-25.
Where it has been discussed
Related concept pages
- Retrospectives — upgrades review output from “an observation report” to “an action with somewhere to go”; this card is the final mile of the review loop.
- AI Tool Applications — automated AI inventories (daily logs, weekly insights) generate more recommendations, making a routing mechanism even more necessary.
Related concepts
Externalizing Judgment, The Problem Analysis Method, Automation