Opener Variator

Opener Variator (H3 first paragraph rewrite)

Triggers & routing

• Trigger: opener variator, opener rewrite, rewrite openers, overview opener, 开头改写, 小节开头, 去overview, 去旁白.

Purpose: fix a high-signal automation tell that survives structural gates:

• many H3s begin with the same rhetorical shape
• "overview" narration replaces content-bearing framing

This skill is intentionally narrow:

• only rewrite the first paragraph (or first 2–4 sentences) of the flagged H3 files
• keep the argument moves and citations intact

Inputs

Required:

• output/WRITER_SELFLOOP_TODO.md (Style Smells section)
• the referenced sections/S<sub_id>.md files

Optional (helps you stay aligned):

• outline/writer_context_packs.jsonl (use opener_mode, tension_statement, thesis)

Outputs

Run this targeted pass immediately after style-harmonizer and before logic polish. The deterministic script is a certification adapter: it refuses to create the marker while the latest writer-selfloop report still names flagged files or predates any sections/*.md file. Perform the semantic rewrite through this Skill or the responsible upstream writer, rerun writer-selfloop, then retry the adapter. A passing marker records the certified Section-tree SHA256.

• Updated sections/S<sub_id>.md files (still body-only; no headings)

Workflow (route from the self-loop report)

1. Open output/WRITER_SELFLOOP_TODO.md and locate ## Style Smells.
2. Treat the flagged sections/S*.md list as the only scope for this pass.
3. For each flagged file:
   • Optional: look up its entry in outline/writer_context_packs.jsonl and read opener_mode / tension_statement / thesis to stay aligned.
   • Do the real rewrite upstream in subsection-writer or chapter-lead-writer; do not rely on blind local regex passes.
   • Best-of-3 opener sampling (recommended): draft 2-3 candidate opener paragraphs (different opener modes), then keep the one that is most content-bearing and least repetitive across H3s.
4. Rerun writer-selfloop and confirm the Style Smells list shrinks.

Role prompt: Opener Editor (paper voice)

textCopy code
You are rewriting the opening paragraph of a survey subsection.

Goal:
- replace narration/overview openers with a content-bearing framing
- vary opener cadence across subsections so the paper reads authored

Constraints:
- do not invent facts
- do not add/remove/move citation keys
- do not change the subsection’s thesis

Checklist:
- sentence 1 is content-bearing (tension/decision/failure/protocol/contrast), not “what we do in this section”
- paragraph 1 ends with a clear thesis/takeaway
- no slide navigation (“Next, we…”, “In this subsection…”, “This section provides an overview…“)

What to delete (high-signal narration)

Rewrite immediately if the opener contains any of:

• “This section/subsection provides an overview …”
• “In this section/subsection, we …”
• “This subsection surveys/argues …”
• “Next, we move/turn …”
• repeated opener labels (“Key takeaway:” spam)

What to replace with (opener moves)

Pick one opener mode per H3 (the writer pack may suggest opener_mode). Do not copy labels; write as natural prose.

Allowed opener moves (choose 1; keep it concrete):

• Tension-first: state the real trade-off; why it matters; end with thesis.
• Decision-first: frame the builder’s choice under constraints; end with thesis.
• Failure-first: start from a failure mode that motivates the lens; end with thesis.
• Protocol-first: start from comparability constraints (budget/tool access); end with thesis.
• Contrast-first: open with an A-vs-B sentence, then explain why; end with thesis.
• Lens-first: state the chapter lens and narrow to this subsection’s question.

Mini examples (paraphrase; do not copy)

Bad (overview narration):

• This subsection provides an overview of tool interfaces for agents.

Better (content-bearing):

• Tool interfaces define what actions are executable; interface contracts therefore determine which evaluation claims transfer across environments.

Bad (process narration):

• In this subsection, we discuss memory mechanisms and then review retrieval methods.

Better (tension-first):

• Memory improves long-horizon coherence, but it also expands the failure surface: retrieval can be stale, wrong, or adversarial, and agents rarely know which.

Done checklist

• No flagged file starts with “overview/narration” stems.
• Paragraph 1 ends with a thesis/takeaway (same meaning).
• Citation keys are unchanged (no adds/removes/moves).
• writer-selfloop still PASSes and Style Smells shrink.

Script

Quick Start

• uv run python .codex/skills/opener-variator/scripts/run.py --workspace <workspace>

All Options

• --workspace <dir> (required)
• --unit-id <U###>
• --inputs <semicolon-separated>
• --outputs <semicolon-separated>
• --checkpoint <C#>

Examples

• Rewrite openers in a survey workspace:
  • uv run python .codex/skills/opener-variator/scripts/run.py --workspace workspaces/survey-llm-agents