Record: round-trip exercise handoff — 2026-11-07

Written so a fresh session can resume without this conversation’s memory.

The exercise

Aim: determine, more accurately, whether the extracted rules match the original intent in the prompts.

Processing:

Extract formal rules from prompt prose.
Generate human-readable prose from the extracted rules.
Compare the original and the generated prose. When they match, we are good; when they do not, explain the difference in a form a human reads quickly.

The complaint that started it: the comparisons offered so far are unclear to a human. This record is the handoff for that exercise.

scripts/prompts-build.mjs now writes, per compiled prompt, rendered.md (prose generated from the rules, via ruleset_render) and comparison.md (the original-vs-generated comparison).
The comparison lists the lines a rule captured — with the rule id, a verdict, and what was kept or lost — and then the lines no rule captured, so the gap in intent is visible rather than hidden.
make prompts regenerates all of it; tests/31 checks the two new forms.
Fixed a manifest bug: the full build appended to a stale manifest, duplicating entries across repeated runs.

krank is a production-rules engine whose notation an LLM can emit reliably, plus a renderer that writes the rules back into prose for a round trip.
The object model (dataflow.in/object-model.json, 79 objects) names the facts; the extractors (scripts/site-facts.mjs) produce them from files; the markers (documents/13-markers.md) tie prompt lines to rules.
prompts.out/ is the generated formal form (rules.json, marked.md, reduced.md, rendered.md, comparison.md); make prompts builds it.
The studio (local, make studio) drives the interaction cycle: write prose, review candidates, approve, render, compare, check; it streams live progress and the model’s reasoning.
The shared rules split prompts/capabilities/ into artifacts/ and execution/ (v2026-11-06.1, breaking).
The general pattern, recorded in the skeleton, is: capture intent in prose, render it into an exact form, iterate with the human to verify and elaborate.

13 prompt families are compiled; the rest of the corpus is not. The comparison makes the gap explicit — e.g. prompts/01-contract.md has 4 lines captured and 150 not captured.
The comparison is deterministic (the renderer’s output next to the original), and it now names the uncovered lines explicitly.
The five capability-artifact prompts (01–05) are now compiled; their long bullets are line-wrapped in the shared files, so their cases quote a single-line fragment rather than the whole sentence.
sources/browser/index.html carries an uncommitted reformatting (functionally identical, roughly 189 → 549 lines) done outside this session; decide whether to keep or revert it.

Compile the remaining rule-dense families: capabilities/execution/06–09, features/01–08, episodes/01–02, how-to-write-research.md, flavors/02-cpp-conventions.md, and 10-project.md.
If the deterministic comparison is still unclear, add a model-written plain-language explanation of each mismatch.