AI chair runs GTD weekly-review 13 times, gaps and all

The article's starting point is a common failure: the author accepted the six-category GTD method but could never sustain the weekly review that requires walking all six categories every week. The diagnosis is not weak willpower but a structural one — deciding where to start and checking for gaps both fall on the same person, and when both jam at once the review never begins.

The proposed fix moves those two jobs to an AI. A Claude Code skill called /todo ships a weekly-review command that runs six steps in dialogue form, including automatic detection of projects missing a next action and of projects stalled for a long time. The author stresses that the real dividing line is not Todoist versus GitHub Issues, but whether tasks live somewhere the AI can read and write directly; the practical barrier is that this still assumes comfort with GitHub operations.

The record itself is mixed by design. Thirteen reviews over roughly five months include a stretch of about two months with only two completions, and the restart on August 11 was followed by eight reviews in about seven weeks in which no gap exceeded 10 days, though six of the eight finished one to three days late. The mistakes are catalogued as carefully as the successes: a miscount, an off-target alert traced to an inconsistent label left by a half-finished earlier update, and a dropped step in the author's own procedure that stalled review dates for 52 tasks. The reading offered is that the AI is not accurate, but errors do not stay unnoticed — a structure, not a guarantee. Where the bottleneck went is also tracked: project cleanup improved from 22 to 15 missing next actions and from 10 to 5 stalled projects, while the pressure shifted to Someday reviews, which were bulk-reset to 0 and had climbed back to 51 by September 27. Whether that shift is acceptable likely depends on whether seeing where the congestion sits is itself worth the residual errors.