[dmccoystephenson]

Summary

• Replaces the free-form Phase 4 self-review with an objective yes/no rubric anchored on CI status.
• Adds an adversarial framing instruction ("assume FAIL unless you have direct evidence of PASS") to counter LLM self-preference bias.
• Caps intrinsic critique at one pass per PR. Phase 6 now distinguishes internal critique (capped) from external signals — reviewer comments and CI failures — which remain uncapped.
• Adds a new SELF_REVIEW_RUBRIC placeholder so generated skills can append repo-specific objective rubric items.
• Adds an edge-case entry for CI failures on the PR.

Motivation (per RESEARCH.md)

Per the new convention in CLAUDE.md, design changes cite the empirical basis:

• §1 — Self-critique without an external signal is unreliable. Huang et al. ICLR 2024 (arXiv:2310.01798), Kamoi et al. TACL 2024 (arXiv:2406.01297), and arXiv:2604.06996 on self-preference bias all converge on: intrinsic self-correction is neutral-to-harmful without an external grounding signal. This change uses CI status as that signal and replaces the previous "review for logic errors, edge cases, anti-patterns…" framing with an objective rubric.
• §5 — Reflection iterations plateau early. Self-Refine (arXiv:2303.17651) and Reflexion (arXiv:2303.11366) show gains concentrate in iterations 1–2; later iterations regress without external feedback. The one-pass cap is a direct application.

The previous flow already implicitly ran self-review once (linear phases, no loop back to Phase 4), so the cap is more of a documentation/intent change than a behavioral one — but making it explicit prevents future drift and gives the agent a principled reason to refuse re-critiquing.

Retrofit follow-up

Per CLAUDE.md's "Promoting a rule into the template" rule, this template change needs retrofit PRs in existing child skills so they don't silently lag:

• dmccoystephenson/example-l-dev-loop
• dmccoystephenson/example-q-dev-loop
• dmccoystephenson/example-a-dev-loop

Not included in this PR — those are separate repos. Will follow up after this merges (or can be assigned).

Test plan

• All {{placeholder}} tokens still have rows in the Step 4 substitution table (verified: SELF_REVIEW_RUBRIC added)
• Conditional block syntax {{#if SELF_REVIEW_RUBRIC}}...{{/if}} matches the pattern used by other optional placeholders (CODE_PATTERNS, LINT_CMD, REVIEWER, CLAUDE_MD)
• Generated skill compiles cleanly when /create-dev-loop is run against a real repo with and without repo-specific rubric items
• The new Phase 4 reads as actionable instructions, not advice
• Phase 6's "external vs. internal" distinction is clear without re-reading Phase 4

🤖 Generated with Claude Code

---

drafted by Claude on behalf of Daniel Stephenson