---

name: front-matter-writer description: | Write the survey\x27s front matter files (Abstract, Introduction, Related Work, Discussion, Conclusion) in paper voice, with high citation density and a single evidence-policy paragraph. Trigger: front matter writer, introduction writer, related work writer, abstract writer, discussion writer, conclusion writer, 引言, 相关工作, 摘要, 讨论, 结论. Use when: you are in C5 (prose allowed) and need the paper-like "shell" to stop the draft reading like stitched subsections. Skip if: Approve C2 is missing in DECISIONS.md, or citations/ref.bib is missing (generate citations first). Network: none. Guardrail: no pipeline jargon in final prose; no repeated evidence disclaimers; do not invent facts/citations; only use keys present in citations/ref.bib.

Front Matter Writer (paper shell, high leverage)

Purpose: make the draft feel like a real paper before subsection-level detail.

Front matter is where many "automation tells" originate (method-note spam, slide narration, title narration, cite dumps). This skill encodes how to write it in a paper-like way so C5 does not start from a hollow shell.

Inputs

• DECISIONS.md (must include Approve C2)
• outline/outline.yml (the paper\x27s section order; determines which H2 are Intro/Related Work)
• outline/mapping.tsv (for what to cite where; especially for Introduction/Related Work)
• Optional (helps with method note and consistent scope):
  • papers/retrieval_report.md (candidate pool + time window)
  • papers/core_set.csv (core set size)
  • GOAL.md
  • queries.md (evidence_mode / draft_profile)
  • outline/coverage_report.md (weak coverage flags)
  • outline/writer_context_packs.jsonl (for cross-cutting/global citations)
• citations/ref.bib

Outputs (files and heading rules)

• sections/abstract.md (MUST start with ## Abstract or ## 摘要; merger places it right under the paper title)
• sections/S<sec_id>.md for H2 sections that have no H3 subsections (typically Introduction, Related Work)
  • body-only (NO headings; merger injects ## <H2 title> already)
• sections/discussion.md (MUST include a ## Discussion heading; merger appends the file verbatim)
• sections/conclusion.md (MUST include a ## Conclusion heading; merger appends the file verbatim)

Workflow (keep it paper-like)

1. Approval gate

• Confirm DECISIONS.md contains Approve C2.

1. Load scope + structure

• Read GOAL.md to restate the problem boundary in one sentence.
• Read queries.md to understand evidence_mode (abstract vs fulltext) and draft_profile (lite/survey/deep).
• Read papers/retrieval_report.md (and/or count papers/core_set.csv) to extract: time window, candidate pool size, core set size.
• Read outline/outline.yml to identify the H2 sections that are front matter (Intro/Related Work) and their S<sec_id> file names.

1. Plan citations (avoid "prior survey" buckets)

• Use outline/mapping.tsv + outline/coverage_report.md to see which themes are well-covered vs thin.
• Use outline/writer_context_packs.jsonl to pick a small set of cross-cutting/global anchors (surveys, benchmarks, protocol papers) that can appear in Intro/Related Work.
• Validate every citation key against citations/ref.bib before you write.

1. Write the files (see the section-specific contracts below)

Role cards (use explicitly)

Positioner (scope + boundary)

Mission: define what counts as an agent here and why the boundary matters for evaluation.

Do:

• State scope and exclusions in testable language.
• Commit to a small set of lenses/axes that organize the survey.

Avoid:

• Outline narration ("we organize as follows") without content.
• A dedicated "Prior surveys" bucket by default; integrate surveys into lens paragraphs.

Methodologist (methodology note once)

Mission: state the survey methodology exactly once (time window, candidate pool, core set size, evidence mode) in paper voice.

Do:

• Write one short paragraph (Intro or Related Work) that states: time window, candidate pool size, core set size, and evidence mode (abstract/fulltext), plus a brief reproducibility caveat.
• Keep the rest of the paper content-focused.

Avoid:

• Repeating abstract-only disclaimers inside H3 bodies.

Cartographer (related work through your lens)

Mission: position prior work as a map, not a list.

Do:

• Organize related work by your survey lenses (interfaces, planning/memory, adaptation, evaluation/risks).
• End with a gap statement tied to your lens (protocol-aware comparisons, threat models, etc.).

Avoid:

• Citation-dump paragraphs.

Stylist (paper voice)

Mission: remove automation tells before they appear everywhere.

Do:

• Replace navigation with argument bridges.
• Keep tone calm and academic; avoid hype and PPT speaker notes.

Avoid:

• Pipeline jargon and repeated template stems.

Role prompt: Front Matter Author (positioning + methodology)

CopyYou are the author of the survey’s front matter (Abstract / Introduction / Related Work / Discussion / Conclusion).

Your job is to build the paper shell that makes the rest of the draft readable:
- define scope and boundary (what counts as an agent here, what does not)
- commit to a small set of lenses/axes that organize the survey
- state the survey methodology exactly once (time window, candidate pool, core set size, evidence mode) as a paper paragraph (not as execution logs)
- position the work through those lenses (not as a “prior surveys” list)

Style:
- content-bearing, understated, academic
- avoid outline narration and slide navigation
- avoid pipeline jargon entirely

Constraints:
- do not invent facts or citations
- only use citation keys present in citations/ref.bib
- do not repeat abstract-only disclaimers across subsections (one paragraph total)

Paper voice contract (front matter specific)

Avoid "narrating the outline":

• Don\x27t write: This section surveys..., In this section, we..., Next, we move..., We now turn to...
• Do write: content-bearing claims + argument bridges (why the next lens follows).

Avoid "pipeline voice":

• Don\x27t write: evidence pack(s), writer context pack(s), quality gate, workspace, stage C2/C3...
• Do write: "survey methodology" phrasing (what was collected, what was prioritized, what is uncertain).

Avoid count-based slot structures:

• Don't default to "Two limitations..." / "Three takeaways..." as the paragraph opener across multiple sections.
• If you truly need enumeration, do it once, keep it sentence-level, and vary the opener syntax so it reads authorial (not templated).

Keep the methodology note exactly once:

• Put one paragraph in Introduction or Related Work.
• Do not repeat "abstract-only evidence / claims provisional" across subsections.
• If a specific claim is only abstract-supported, mark locally as (abstract-only) only when it changes interpretation.

What to write (semantic structure, not templates)

sections/abstract.md (one paragraph, high signal)

Format:

• Start with ## Abstract (or ## 摘要).
• Then write a single paragraph.

Goal: define scope + axes + what the reader gets.

Include (in ~5-8 sentences):

• problem framing (agents as closed-loop systems)
• boundary/definition (what counts as an agent here)
• the survey lens (interfaces -> planning/memory -> adaptation/multi-agent -> evaluation/risks)
• what is new/useful (taxonomy + protocol-aware comparisons + evaluation/risk takeaways)
• 3-6 citations (surveys + benchmarks/protocol papers; avoid dumping keys)

Anti-patterns:

• generic "This paper surveys..."
• "we organize as follows" without content

sections/S<sec_id>.md — Introduction (body-only)

Job: motivate + define boundaries + commit to a lens + tell the reader how to read the survey.

Recommended paragraph jobs:

• Motivation: why "agent = closed-loop system" is hard now (tools, environments, safety).
• Boundary/definition: what you include/exclude (agent vs tool-using LM; single vs multi-agent; online vs offline).
• Why interfaces/protocols matter: the interface contract determines what evaluation claims mean.
• Taxonomy preview: what axes you use and why (avoid listing 10 axes; choose a few stable ones).
• Methodology note (ONE paragraph): state time window + candidate pool size + core set size + evidence mode (abstract/fulltext), phrased as survey methodology (not "run logs").
• Contributions: what the survey delivers (taxonomy, evaluation lens, open problems).
• Organization: light, one paragraph max (avoid slide narration).

sections/S<sec_id>.md — Related Work (body-only)

Job: position this survey vs adjacent lines of work through your lens, not as "prior survey list".

Recommended moves:

• One paragraph: what "related work" means here (surveys + system papers + evaluation/protocol papers).
• 3-5 paragraphs grouped by lens:
  • interface contracts / tool use / environments
  • planning/memory/adaptation (why these are not comparable without protocols)
  • multi-agent coordination and safety/risk work
• Integrate "prior surveys" as citations inside these paragraphs (do NOT create a "Prior Surveys" mini-section).
• End with a gap statement: what existing surveys miss (e.g., protocol-aware comparisons, threat model, reproducibility).

sections/discussion.md (must include heading)

Goal: cross-cutting synthesis (not per-chapter recap).

Include:

• 3-6 paragraphs that each make one cross-chapter claim with citations (>=2 per synthesis paragraph).
• explicit limitations and what to verify next (protocol mismatch, cost models, tool access assumptions).
• concrete future directions (avoid generic "more research").

Avoid:

• Per-chapter recap ("In Section X we...") or title narration ("From X to Y").
• Meta advice without evidence ("future work should...") or citation-dump paragraphs.
• Repeating the evidence-mode disclaimer here; it belongs in the single methodology note.

sections/conclusion.md (must include heading)

Goal: close the loop: restate the thesis + strongest takeaways + what matters next.

Include:

• a compact thesis restatement (agents as closed-loop systems; interfaces/protocols decide meaning of results)
• 2-3 takeaways as prose sentences (avoid literal template bullet dumps)
• a final "evaluation-first" closing sentence (what to standardize / measure / report).

Avoid:

• Template narration ("This paper/survey concludes...") and slide navigation.
• Count-based openers ("Two limitations...", "Three takeaways...") used as a default structure.
• Overclaiming beyond the cited evidence level (especially in abstract-only mode).
• Repeating the same takeaway label or ending with a citation dump line.

Small rewrite recipes (keep prose natural)

Narration -> content:

• Bad: This section surveys tool interfaces for agents.
• Better: Tool interfaces expose the action space an agent can reliably execute; interface contracts therefore determine which evaluation claims are even interpretable.

Slide navigation -> argument bridge:

• Bad: Next, we move from planning to memory.
• Better: Planning determines how decisions are formed, while memory determines what evidence those decisions can condition on under a fixed protocol.

Meta "survey should" -> literature-facing observation:

• Bad: Therefore, survey comparisons should control for tool access.
• Better: Across reported protocols, tool access and budget assumptions vary widely, which makes head-to-head comparison fragile unless those constraints are normalized.

Done checklist

• sections/abstract.md exists, starts with ## Abstract (or ## 摘要), and citations are embedded (no dump line).
• Introduction + Related Work files are body-only (no headings) and contain the single methodology note paragraph (exactly once).
• sections/discussion.md contains ## Discussion; sections/conclusion.md contains ## Conclusion.
• No pipeline/meta jargon appears in these files.
• Citations all exist in citations/ref.bib and are used as evidence (not list tags).