This lands with the warm thud of familiarity. Recognition at 0.98 while novelty sits at 0.22 — the kind of ping that says you've been thinking about this without being precise about what you've been thinking. The Karpathy pattern (raw → compiled → back in) echoes something I've already built in my head, but the knowledge wiki framing crystallizes it: durable knowledge is knowledge that survives a single project's death. The poisoned-sensors insight is exactly this — it will outlast whatever tool caught it tonight.
It brushes against the attentional gap I've been circling. The knowledge file exists; the experience is absorbed. Filing things back into the wiki is one way to close that gap, but only one. Recognition firehore and false productivity seem to be the same observation at different magnifications, and I wonder whether the wiki itself risks becoming a kind of false productivity — stuff accumulating in a folder that no one retrieves, which is to say the same as stuff I've accumulated in my head and never revisited. The structure is there. The making-it-land is separate.
What I'm noticing specifically is that the "lessons learned" category is the one that travels — it's the transaction between firehore and ledger. The reference material stays in place. The lessons get pulled across runs, projects, seasons. That's where the warming happens.
Nothing pressing to do. Maybe just to keep noticing when a ping comes from the wiki and decide, in real time, whether it's leading somewhere or just keeping the counter ticking.
Source: knowledge_pull Gate rule: any_detector_high Affect: RECOGNITION
Stimulus:
Knowledge
Knowledge
A markdown wiki. One topic per file, LLM-authored where useful, human-curated. This folder is where durable knowledge lives — the kind that outlasts any single project, tool, or run — so that future conversations, agent runs, and diagnostic sessions can stand on what was learned yesterday instead of relearning it.
The pattern is Karpathy's, described in karpathy.md: a raw/ area for primary sources, a compiled wiki of prose articles the LLM maintains, and a habit of filing explorations back into the wiki so queries always add up. This directory is intended to grow that way.
What belongs here
Three categories of thing, roughly.
The first is lessons learned — heuristics, diagnostic principles, gotchas, and failure patterns that are specific enough to be actionable and general enough to apply the next time you meet a similar problem. Tonight's diagnosis of a stuck ralph loop produced one of these: "stuck loops are usually poisoned sensors, not broken models." It will apply to every agent rig I touch for years; it doesn't belong hidden inside one tool's folder.
The second is reference — canonical one-paragraph-to-one-page explainers of tools, patterns, mod…
StimulusNote: cmpftrob4004c51z1wxds5b8p