editing generated artifacts #
a bounded change to a generated artifact edits the current artifact; it does not regenerate it from the context that first produced it. that takes an editable source, not only the flattened output.
keep the recipe #
an editable artifact is a set of parts plus a recipe that composes them. a small model choice can pick one part while a deterministic compositor preserves every other. the recipe is part of the publication: store it with the published result, or the next visit is less capable than the editor was. for a cold start, offer a ready editable artifact that costs no inference until the person asks for a change.
a saved output and a recomposition from its recipe can differ, so compare the decoded results before treating a round trip as lossless.
give each input a role #
passing a whole reference description into every choice preserves details the person wanted changed, and a historical description can falsely satisfy an edit request ("long-haired" already looks done). name the current artifact and past inspiration separately, and state the current value in each question. narrow requests keep unrelated fields; moods and references absent from the catalog allow a thematic approximation. optional context resolves open choices without overriding the request or blocking the interaction. one good example does not show that context helped; compare identical requests with and without it.
validate meaning, not bytes #
a compatibility gate bounds small numeric noise and keeps exact checks where exactness matters (alpha, geometry); byte equality silently disables editing for inputs that are equivalent. a parser that rejects a valid upstream variant has not shown the structure is absent, so check imported formats across several real records. judge the semantic result: changed output does not mean the requested change happened. overlays need attachment points per direction and motion tracked from the original, and refuse uncertain placement instead of replacing the subject.
suggestions follow capability #
cache suggested actions by what the current representation can do, not only by its subject: suggestions for a layered recipe are wrong once it falls back to a flat output. give the suggestion step the renderer's real capabilities, since a plausible suggestion for a nonexistent part is a broken affordance. suggestions append to the request; they never overwrite it or submit it.
model judgments are not facts #
a classification can help choose a visual without becoming an attribute of the person, and a renderer accepting a field does not give it meaning. record what a model selected in an inspectable form (the actual parts, categories, and values), and treat a selection as evidence of the output, not of the reasoning. an explanation of a decision model must not fabricate probabilities or imply that clicking an example ran the model.
sources #
- Character
docs/design.md,docs/verification.md,docs/jev-lab.md, and its regression scripts, 2026-09-19 to 2026-09-21: layered sprite recipes, color-tolerance compatibility gate, catalog-based edits. - rpg.actor developer guide (generator source records); TypeSafe Choice.