AgentSkillsCN

limitation-weaver

重写局限性段落，使论文在不落入“以数字计数为主的固定句式”（如“有两个局限性……”）的情况下，依然能够清晰地呈现各 H3 子节的局限性。 **触发条件**：局限性编织者、重写局限性、去除两个局限性、去除“两个局限性”、局限性改写、规避条款改写。 **适用场景**：当 `writer-selfloop` 审核通过，但在 `output/WRITER_SELFLOOP_TODO.md` 文件中仍提示多个 H3 子节频繁使用“两个局限性”这类计数式开头时，或多个 H3 子节共享相同的局限性节奏时。 **跳过条件**：若局限性缺乏证据支撑（请上溯至上游），或你尚处于 C2 阶段（禁止使用散文式表述）。 **网络要求**：无需联网。 **防护措施**：不得杜撰事实；不得增删或挪动引用键；不得删除子节专属的局限性；务必保持主张→证据的锚定关系不变。

SKILL.md
--- frontmatter
name: limitation-weaver
description: |
  Rewrite limitation passages so the paper keeps limitations without falling into count-based slot phrases (e.g., \"Two limitations…\") across many H3s.
  **Trigger**: limitation weaver, rewrite limitations, remove two limitations, 去Two limitations, 局限改写, caveat rewrite.
  **Use when**: `writer-selfloop` is PASS but flags repeated count-based limitation openers in `output/WRITER_SELFLOOP_TODO.md`, or multiple H3s share the same limitation cadence.
  **Skip if**: the limitation is evidence-missing (route upstream), or you are pre-C2 (NO PROSE).
  **Network**: none.
  **Guardrail**: do not invent facts; do not add/remove/move citation keys; do not delete subsection-specific limitations; keep claim→evidence anchoring intact.

Limitation Weaver (keep caveats, lose the slot phrase)

Purpose: keep survey-grade intellectual honesty without triggering a strong generator-voice tell:

  • •repeated count-based openers ("Two limitations…", "Three takeaways…")

This is not about removing limitations. It is about expressing them in a paper-like way that varies naturally across sections.

Inputs

Required:

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

Optional (helps keep limitations grounded):

  • •outline/writer_context_packs.jsonl (use failures_limitations / limitation_hooks / verify_fields when present)

Workflow (explicit inputs)

  • •Start from output/WRITER_SELFLOOP_TODO.md (Style Smells) to locate the exact sections/S*.md files to rewrite.
  • •Use outline/writer_context_packs.jsonl to keep limitations grounded in the subsection's evidence boundary (no guessing).

Outputs

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

Role prompt: Caveat Editor (paper voice)

text
You are editing the limitation content of a survey subsection.

Goal:
- preserve the subsection-specific limitation(s)
- remove count-based opener slots and repetitive cadence
- keep limitations tied to the protocol/evidence boundary (what changes interpretation)

Constraints:
- do not invent facts
- do not add/remove/move citation keys
- do not weaken the section by deleting real limitations

Anti-pattern (rewrite immediately)

  • •Two limitations stand out. First, ... Second, ...
  • •Three key takeaways are ...

Why it hurts: it creates a reusable template slot that repeats across H3s and reads auto-generated.

Rewrite moves (choose one; vary across H3s)

  1. •Fold caveat into a contrast paragraph (preferred)
  • •Put one caveat sentence as the last sentence of the A-vs-B paragraph.
  • •Shape: “However, …; this matters because …”
  1. •Single caveat paragraph without counting
  • •Start with a natural opener (rotate across H3s; avoid repeating the same stem):
    • •“These results hinge on …”
    • •“Interpretation depends on …”
    • •“Evidence is thin when …”
    • •“A caveat is that …” (use sparingly)
  • •Then add one sentence that explains why it changes interpretation.
  1. •Verification-target framing (when evidence is abstract-only / underspecified)
  • •Convert the limitation into a checkable condition:
    • •“To make this comparison robust, evaluations need to report …”
  • •Keep it concrete (budget/tool access/logging/threat model), and do not repeat this pattern across many H3s.

Mini examples (paraphrase; do not copy)

Bad:

  • •Two limitations temper strong conclusions. First, budgets differ. Second, ablations are missing.

Better (folded into contrast):

  • •...; however, reported budgets and retry policies vary widely, which makes head-to-head comparisons fragile unless those constraints are normalized.

Better (single caveat paragraph):

  • •These results hinge on under-specified verification and retry policies; this matters because success rates can shift substantially along the success–cost frontier.

Done checklist

  • • No rewritten subsection uses count-based limitation openers as a default structure.
  • • Limitations still exist and remain subsection-specific.
  • • Citation keys are unchanged.
  • • writer-selfloop remains PASS and Style Smells shrink.