IrredenEngine: Skill for Claude Code

Render Debug Loop

Scope and progress

For a single defect, identify the symptom and a falsifiable acceptance criterion. For an ongoing audit, maintain an ordered TODO in the existing task or design artifact: distinguish diagnosed, experimentally tested, implemented and validated work. Preserve the user’s priorities when one symptom reveals another. A passing local experiment or an opened PR completes that slice, not the broader assignment. Use the existing commit-and-push workflow for reviewable slices; preserve the session’s authorization and draft/ready preference.

An experiment tests a stated hypothesis by changing a controlled input or implementation and evaluating the resulting evidence. A multi-shot sweep is one experiment, not one iteration per screenshot. Unchanged rebuilds and verification reruns do not create new hypotheses.

There is no fixed iteration cap. After several attempts on the same symptom with no new evidence, reassess the hypothesis, instrumentation or abstraction level; do not keep varying shader constants blindly. Capture a smaller fixture, inspect an intermediate buffer, derive the coordinate transform, or move to independent authorized work while recording the unresolved item. Reassessment is not a reason to end the session or ask for renewed permission. Stop dependent work when it requires missing user input, unavailable capability or an explicit user limit; state that concrete limitation and continue other authorized work where possible.

Investigation

1. Establish the affected render path and coordinate/data contract. Trace the producer, intermediates and consumer; a visible lighting defect may originate in geometry or depth reconstruction. Treat topic diagnoses as hypotheses, not proof, and read the applicable module instructions before edits.
2. Choose a discriminating control: a known primitive, isolated pass, paired rendering path or analytical projection. Establish the failure from an adequate existing capture or capture it before changing it. If the demo cannot distinguish the hypotheses, extend its probes, controls or metrics first.
3. Run the smallest useful experiment using capture and evaluation. Vary one cause at a time where possible. Record combined changes when they are necessary, so their effects are not misattributed.
4. Decide from evidence: retain and validate a fix, reject an experiment, or refine the hypothesis. Restore rejected edits before the next independent experiment; preserve a patch when the result will be cited. Report an unresolved finding honestly rather than presenting improved screenshots as a complete fix.
5. Validate the affected contracts and adjacent modes before publishing. Explain remaining limitations and continue the ordered TODO within the user’s scope.