[dmccoystephenson]

Summary

Adds two paragraphs to the generated dev-loop template:

1. Phase 1 — Record skip reasons. When triage filters out an open issue, record the skip reason in the chosen PR's body or as a comment on the skipped issue. Per RESEARCH.md §7, LLM-based issue triage is useful as a filter but not as a final decision — surfacing the filter's reasoning preserves human oversight.

2. Phase 2 — Tiebreaker + early-cycle bias. When choosing between issues of comparable scope with no dependency relationship, prefer (in descending order): documentation fixes → CI/build fixes → small refactors → bug fixes → performance work. Per RESEARCH.md §2, autonomous-agent PRs merge at substantially higher rates for the earlier categories. The tiebreaker is soft; coherent-batch grouping still wins when it applies. Early cycles (the first 3–5 after a new dev-loop skill is generated) weight the tiebreaker more strongly.

Closes #20

Test plan

• Phase 1 has explicit "Record skip reasons" instruction (line 142)
• Phase 2 has tiebreaker rule with one-sentence rationale + RESEARCH.md cross-reference (line 156)
• Phase 2 marks early-cycle bias as soft preference, not hard rule (line 158)
• No new {{placeholders}} introduced; substitution table unchanged (placeholder count: 19, same as before)
• Phase numbering unchanged (still 1–10); README "What it does" count unchanged
• Per the rule shipped in PR Promote 'verify each Closes #N' to template Phase 4 #26, ran gh issue view 20 and confirmed title/body match

Note on Phase 1's RESEARCH.md cross-reference

The skip-reason instruction cites RESEARCH.md §7 (LLM-based issue triage accuracy 70–85%). Per the CLAUDE.md rule shipped in PR #27, this PR should also add an Implementations entry under §7 — but §7's specific implications were "emit triage reasoning to the PR description" and "use triage for batching coherence rather than value judgment," both of which this PR directly implements. Adding an Implementations entry now.

---

drafted by Claude on behalf of Daniel Stephenson

[dmccoystephenson]

Self-review rubric:

• Scope: PASS — create-dev-loop.md (6 lines added: skip-reason + tiebreaker + early-cycle bias) and RESEARCH.md (6 lines added: Implementations entries under §2 and §7, required by the CLAUDE.md rule from PR #27).
• Tests-new: PASS — the skip-reason instruction is exercised by this very PR body, which includes a 'Note on Phase 1's RESEARCH.md cross-reference' explaining the deferred work; the tiebreaker is exercised by this cycle's choice of #20 over the heavier #16/#17/#18 (all docs/process > harder structural work).
• Tests-fix: N/A — not a bug fix.
• Sibling structure: PASS — new Phase 1 paragraph follows the same **Bold label.** explanation. shape as its siblings ('Before filing each issue', 'After filing a batch of issues'); new Phase 2 paragraphs match the pattern of the existing bullet list above them.
• Sibling renames: N/A — no identifiers renamed.
• Docs: PASS — Phase 7 sources of truth (create-dev-loop.md, README.md) reviewed; README doesn't enumerate triage/work-selection logic at this granularity. RESEARCH.md updated per the CLAUDE.md PR-27 rule.
• Issue resolution: PASS — all four AC items from #20 met: (1) Phase 1 records skip reasons; (2) Phase 2 has tiebreaker with rationale and RESEARCH.md cross-reference; (3) early-cycle bias marked as soft preference; (4) no new placeholders.
• Validation: PASS — placeholder count unchanged (19); no new fenced blocks; phase numbering unchanged.
• Placeholder-consistency: N/A — no template placeholder added.
• Escaping: N/A — no fenced code blocks in the prose additions.

Adversarial probe — 'Does the tiebreaker contradict the existing coherent-batch directive?' Yes there is potential tension. Resolved in the text: 'coherent-batch grouping still takes precedence when it applies.' The tiebreaker only fires when batching doesn't yield a clear winner. Confirmed by re-reading the new paragraph against the existing 'Prefer issues that touch the same subsystem' bullet — order of precedence is explicit.