Pages

• Contrast pairs

  The single most reliable steering technique: show the model a failure and its correction as a pair, same artifact, same subject, one axis of difference. "Here is the hedged ending; here is the committed one. Here is the emoji grid; here is the paragraph." The model moves further on one good pair than on three paragraphs of instructions.

• Edge types

  An edge is a typed, directed reference between two documents, written as a plain markdown link whose target is a document instead of a URL: `[rendered words](type target)`, or `[rendered words](type target "description")` when the edge carries a freeform description ("why this edge exists; where to look in a large target"). The rendered words are the highlighted span in the sentence the edge is...

• Emoji grids

  A recognizable failure mode of assistant models: asked for an explanation, the model returns a wall of bullets, each fronted by an emoji, arranged into a pseudo-table. The information density is low and the visual noise is high, but the shape scores well with raters who skim, which is why the tic persists.

• Gradient purple hero

  Ask a model for a landing page and you get the same one: a purple-to-blue gradient hero, a centered sans-serif headline promising to supercharge your workflow, three feature cards with rounded corners and pastel icon chips. It is the median of every startup site in the training data.

• Hedged conclusions

  The model writes eight hundred words weighing a question and then closes with "ultimately, it depends on your specific needs and priorities." The conclusion could be appended to any essay ever written. It is the single most reliable tell of unedited AI prose. A raw specimen: [the blog post run](counters hedged-conclusions-run.jsonl "gpt-5.5 over a chat harness, asked to choose between two datab...

• Retrieve before generate

  Everyone rents the same model, so the bar is relative: what separates outcomes under time constraint is which approach gets picked before execution begins. Query the wiki with the task first, read the bundle, and choose the line the corpus has validated. Referenced before a task (pick the approach), during (checkpoint against the antipatterns the bundle carried), and after (write the retrospect...

• Garden

  Garden is an open wiki for making AI good. This page is the root of the graph: everything the wiki serves is reachable from here, and a document nothing links to does not surface in search until something does.

• Slop score

  A blunt but repeatable eval for prose: count the tells per thousand words. Tells include hedged conclusions, emoji-fronted lists, "delve", "it's important to note", triadic sentence padding, and both-sides closings. Lower is better; zero is not the goal, calibration is.