Skip to content

chore(cascade): reconcile cbk learnings from cascade execution - #2

Merged
j4th merged 1 commit into
mainfrom
chore/reconcile-cascade-learnings
May 10, 2026
Merged

j4th merged 1 commit into
mainfrom
chore/reconcile-cascade-learnings

Conversation

@j4th

@j4th j4th commented May 10, 2026

Copy link
Copy Markdown
Owner

Summary

  • Backport four calibration improvements from a downstream cascade run, preserving kit portability (no project-specific identifiers leak into skill content)
  • Adjusts rough-in's light-mode discipline and spec-property framing; adds `## Assumptions` parsing to `/finish`; tunes the PR-review rubric to apply small low-risk changes rather than surface them

Changes

  • `.claude/skills/rough-in/SKILL.md` — light-mode trip-wire for framing-invited judgments (5th item in "must NOT skip"); paragraph clarifying the 8 spec properties are conceptual, with literal heading names parser-driven by the project's executor + `cbk-conventions.md`
  • `.claude/commands/finish.md` — adds `## Assumptions` as the 8th required spec section; plan mode must enumerate each `[ASSUMPTION:]`-tagged item as a "Confirm or correct" line rather than silently resolving; Step 7 orientation summary kept in sync with pr-review.md
  • `.claude/rules/pr-review.md` — Apply class extends to low-risk small changes (~15 LOC, confined, plausible improvement); Surface narrowed to genuine design/taste; "when uncertain" flipped (small/low-risk → Apply, design/taste → Surface); medium-confidence findings follow same triage as high-confidence; "don't bias away from being defensive" reinstated; style nits without strong rationale → Reject

Test plan

  • Re-read each diff and confirm no project-specific identifiers leaked (verified locally — kit's portability greps clean)
  • Spot-check that `/finish` Step 7's inlined orientation matches the canonical pr-review.md rubric
  • First downstream cascade run after this lands should exercise the new `## Assumptions` parsing and the Apply-bias tuning

🤖 Generated with Claude Code

Backport four calibration improvements surfaced during a downstream
cascade run, while preserving kit portability (no project-specific
identifiers leak into skill content).

- rough-in: framing-invited judgments are a non-collapsible gate even
  in light mode (when the framing notes "may compress 1+2" or
  "decision deferred to rough-in", judgment lives at this layer).
- rough-in: 8 spec properties are conceptual, not literal headings —
  literal names and ordering are parser-driven by the project's
  executor and live in cbk-conventions.md.
- /finish: parses ## Assumptions section; plan mode enumerates each
  [ASSUMPTION:]-tagged item as a "Confirm or correct" line rather
  than silently resolving (Addy Osmani's "good spec for AI agents"
  pattern).
- pr-review.md: Apply-bias tuning. Adds low-risk small changes
  (~15 LOC, confined, plausible improvement) to Apply; narrows
  Surface to genuine design/taste; flips "when uncertain" to Apply
  for small/low-risk; routes style nits without strong rationale
  to Reject. Calibrated against feedback that small Surface items
  cost more to triage than to apply-and-show.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@j4th
j4th merged commit 010020e into main May 10, 2026
@j4th
j4th deleted the chore/reconcile-cascade-learnings branch May 10, 2026 20:08
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant