Skill: write-related-work

A file is considered a placeholder if it contains status: placeholder or if it has no ## Cluster: entries. A file is populated if it has at least one ## Cluster: entry with a non-empty **What this cluster does** field.

• If the file is populated: proceed directly to Step 1.
• If the file is a placeholder or missing: go to Step 0b to derive clusters from the literature pipeline before continuing.

Step 0b — Derive Clusters from Literature Pipeline

The clusters have not been defined yet. Derive them from the existing literature outputs rather than stopping or asking the user to run project-init.

Load available literature sources (in order of richness):

Glob: literature/synthesis/*.md          → thematic clusters from deep-paper-synthesis
Read: literature/gap_map.md              → coverage matrix from research-gap-mapper
Read: literature/papers.csv              → full paper list with relevance scores and gap_notes
Glob: literature/triage_report_*.md      → gap themes from paper-search-and-triage

Derive clusters using the best available source:

1. If synthesis files exist (literature/synthesis/*.md): Each synthesis file contains a "Thematic Clusters" section (from deep-paper-synthesis Step 4a). Read all synthesis files and collect every named cluster. Merge clusters with the same theme (e.g., "LLM-based APR" and "LLM for code repair" are the same cluster). Aim for 3–5 distinct clusters that cover the synthesized papers without overlap.

2. If literature/gap_map.md exists (but no synthesis files): The gap map's coverage matrix axes define the research dimensions. Use the axis values and their associated papers to construct clusters. Each axis value with 2+ papers becomes a cluster.

3. If only literature/papers.csv exists: Group papers by their gap_notes themes. Papers with similar gap notes belong to the same cluster. Use relevance scores to prioritize — score-4 and score-5 papers define the clusters; lower-scored papers fill in secondarily.

4. If no literature files exist at all: Tell the user:

   "I don't have enough literature context to derive clusters automatically. Please either:

   1. Run paper-search-and-triage to find relevant papers, then come back
   2. Tell me 3–5 thematic areas of related work and I'll build the clusters from your description"

   If the user provides descriptions inline, use those to construct clusters. Do not stop — proceed with whatever the user provides.

For each derived cluster, fill in:

## Cluster: {{cluster name}}
**What this cluster does**: {{1-sentence description derived from paper summaries}}
**How {{SYSTEM_NAME}} differs**: {{1-sentence differentiation — use gap_notes or synthesis
  "Relation to this work" fields; write "TBD" if not determinable yet}}
**Key papers to cite**: {{cite keys or titles from papers in this cluster}}

Write the derived clusters to project/related-work-clusters.md, replacing the placeholder:

# Related Work Clusters

project_name: {{project_name from research-focus.md}}
last_updated: {{TODAY}}
derived_from: {{list of source files used, e.g. "literature/synthesis/llm_repair_synthesis.md,
  literature/gap_map.md"}}

{{one ## Cluster: block per cluster}}

Then show the user what was derived:

"I derived [N] clusters from your literature pipeline:

1. [Cluster 1] — [N] papers: [paper titles or keys]
2. [Cluster 2] — ... ...

Does this look right? You can ask me to merge, split, rename, or reorder clusters before I draft the section. Or say 'looks good' to proceed."

Wait for confirmation. If the user requests adjustments, update project/related-work-clusters.md accordingly before continuing. If the user says "looks good" or equivalent, proceed to Step 1.

Step 1 — Check Review Mode

If review_mode: yes in project/venue-config.md:

• Do NOT mention author names, lab names, or institution names.
• Do NOT write "Unlike our previous system [X]..." — treat self-references as third-party.
• Use "{{SYSTEM_NAME}}" when referring to this paper's system.

Inform the user of current mode.

Step 2 — Load Literature

Read in this order:

Read: {{main_tex from project/paper-paths.md}}
Glob: {{sections_dir}}/*.tex
Read: {{bibliography from project/paper-paths.md}}
Read: literature/papers.csv        (if exists)
Glob: literature/synthesis/*.md    (if exists)

For each synthesis .md file found, read it — these contain the deep summaries from deep-paper-synthesis and are the primary source of content for related work paragraphs.

If papers.csv has entries, parse it for the full paper list with relevance scores.

If both literature/ files are empty:

"I don't see any synthesized papers in literature/synthesis/. I can still draft related work using the cluster descriptions in project/related-work-clusters.md, but the paragraphs will have placeholder citations. Would you like to proceed, or run deep-paper-synthesis first?"

Step 3 — Confirm Cluster Structure

Present the clusters from project/related-work-clusters.md to the user:

"I found [N] thematic clusters in project/related-work-clusters.md:

1. [Cluster 1 name]
2. [Cluster 2 name] ...

For a [N]-page paper with a [budget]-word related work section, [3-4] clusters is optimal. Should I use all [N] clusters, merge any, or adjust the order?"

Step 4 — Draft Each Cluster

For each cluster, write 1–2 paragraphs following this structure:

1. Cluster description (2–3 sentences): what does this line of work do? What is the general approach, and what problem does it address?
2. Key papers with differentiation: cite 2–5 key papers and distinguish each using the pattern: "Unlike [X] which [what X does], {{SYSTEM_NAME}} [what makes it different]."
3. Positioning sentence: 1 sentence making the cluster's relationship to {{SYSTEM_NAME}} explicit — whether competitive (same task), complementary (different task, shared infrastructure), or foundational (prior work this builds on).

Differentiation Pattern Library

Use these templates when writing differentiation within each cluster:

"Unlike [X] which [what X does], {{SYSTEM_NAME}} [how {{SYSTEM_NAME}} differs]."

"In contrast to [X], which [limitation of X], {{SYSTEM_NAME}} [advantage]."

"[X] shares our goal of [shared goal] but differs in [key dimension]:
 [X does Y]; {{SYSTEM_NAME}} [does Z instead]."

"The closest prior work is [X] \citep{key}, which [brief description].
 However, [X] [specific limitation], whereas {{SYSTEM_NAME}} [specific advantage]."

"{{SYSTEM_NAME}} builds on [X]'s [shared idea] but extends it with [new contribution],
 enabling [new capability not in X]."

Choose the template based on how competitive vs. complementary the cited work is:

• Competitive (same task, different approach): use "Unlike X..." or "In contrast to X..."
• Complementary (different task, shared infrastructure): use "{{SYSTEM_NAME}} builds on..." or "This work is orthogonal to X..."
• Seminal (foundational paper everyone cites): one citation in the cluster description, no special differentiation needed.

Using Synthesis Cards

If a synthesis card exists in literature/synthesis/ for a paper being cited:

• Use the "Limitations" and "Relation to this work" fields from the card to derive the differentiation sentence.
• Use the "Results" field to include a specific number when citing the paper's performance.

If no synthesis card exists, use the abstract from literature/papers.csv, or the cluster differentiation sentence from project/related-work-clusters.md as the basis.

Step 5 — Write Closing Positioning Paragraph

After all clusters, add a 2–3 sentence paragraph that summarizes {{SYSTEM_NAME}}'s unique position across all clusters. Use the "What Makes This Novel" field from project/research-focus.md as the basis. Also reference the headline result from project/contributions.md if appropriate.

Pattern:

"In summary, {{SYSTEM_NAME}} occupies a distinct position at the intersection of [cluster 1] and [cluster 2]. Existing systems either [limitation 1] or [limitation 2]. {{SYSTEM_NAME}} is the first to [unique contribution from research-focus.md]."

Step 6 — Check Against Background Section

Read the background section (if it exists) and ensure related work does NOT repeat background content:

Grep(pattern=r"\\subsection", path="{{sections_dir}}/background.tex", output_mode="count")

Read background.tex to find which concepts it already explains. Flag any content in the related work draft that duplicates those explanations. Related work should only cite concepts with "as described in Section~\ref{sec:background}" rather than re-explaining them.

Step 7 — Check Citations Against Bibliography

For every \cite{} key in the draft, check if it exists in the bibliography:

Grep(pattern=r"{{cite_key}}", path="{{bibliography from project/paper-paths.md}}")

For each missing key, note as TODO and provide guidance on finding the entry. Never invent DOIs or page numbers.

If a paper appears in literature/papers.csv but has no BibTeX entry yet, offer to construct one from the CSV metadata (arxiv_id, title, authors, year, venue, doi fields).

Step 8 — Apply Anonymization Check

If review mode is active:

Grep(pattern=r"our lab|our prior work|we previously|our previous|our earlier|our group",
     path="{{sections_dir}}/related_work.tex", output_mode="content")

Also check for any system names that uniquely identify the research group. Flag all hits and suggest neutral rewrites.

Step 9 — Page Budget Check

wc -w {{sections_dir}}/related_work.tex

Compare against budget in project/venue-config.md. Defaults if not specified:

• ACL 8-page: 400–550 words (0.75–1.0 page)
• NeurIPS 9-page: 500–650 words
• USENIX 13-page: 700–900 words

If over budget:

• Reduce complementary/foundational clusters to 2–3 sentences each.
• Keep competitive clusters (same task) at full length — reviewers scrutinize these.
• Cut citations to ≤ 3 per cluster; keep only the most directly related papers.

If under budget:

• Expand the most competitive cluster (same task, different approach) — reviewers look for this.
• Add a paragraph on datasets/benchmarks used by prior work if multiple benchmarks exist in the field.

Step 10 — Write Output File

Write to {{sections_dir}}/related_work.tex.

The file should begin:

% Related Work section — {{SYSTEM_NAME}} paper
% Generated by write-related-work skill

\section{Related Work}
\label{sec:related}

Each cluster uses \paragraph{Cluster Name.} (not \subsection) to keep the section compact. This is standard for related work in most venues.

After writing, check that \input{sections/related_work} is in the main .tex:

grep "input{sections/related_work" {{main_tex}} 2>/dev/null || \
grep "input{sections/related" {{main_tex}} 2>/dev/null

If not present, tell the user where to add it (typically after Experiments, before Conclusion).

Step 11 — Final Checklist

• \section{Related Work} with \label{sec:related} present.
• [N] thematic clusters using \paragraph{} headers.
• Each cluster has explicit differentiation using "Unlike X..." or "In contrast...".
• Closing positioning paragraph summarizing {{SYSTEM_NAME}}'s unique contribution.
• No repeated content from the background section.
• All \cite{} keys verified in bibliography; missing entries noted as TODO.
• No author/lab names in review mode.
• Word count is within venue budget.
• \input{sections/related_work} placement checked in main .tex.

Step 12 — Remind User to Sync with Overleaf

SYNC REMINDER:
If your paper/ directory is a git submodule linked to Overleaf:

  cd {{paper_root_dir}}
  git add latex/sections/related_work.tex latex/custom.bib
  git commit -m "Add related work section"
  git push

After pushing, verify on Overleaf:
- All citations resolve (no "?" in the PDF)
- Related work does not duplicate content from the background section
- The section fits within the page budget