DL Paper Code Blog

Overview

This skill turns the standardized "paper + repo" bundle into a publish-ready article: you will read the PDF derivatives, inspect the code snapshot, fill the scaffold outline, write the final Markdown post, and keep figure assets synced with Lsky. Everything lives inside a pre-generated workspace, so the work is editorial/reporting instead of scraping.

Key ingredients live in materials_manifest.json, paper/, code/, .lsky_upload_cache.json, and article_scaffold.md. The assets/article_scaffold_template.md file can be copied whenever you need a fresh outline.

1. Intake the Bundle

1. Map the workspace. Use the checklist in references/materials.md to understand where each artifact sits (paper text, figure crops, repo context, bibliography, cache files).
2. Verify availability. If the manifest references missing directories (e.g., repo not captured), pause and ask the user; otherwise continue.
3. Prepare the scaffold. Copy assets/article_scaffold_template.md into the project root (or overwrite the existing article_scaffold.md). This is your scratchpad for planning.

2. Research & Outline

Follow the phased process described in references/workflow.md:

• Paper deep dive: mine paper_text.txt, figures_manifest.json, and rendered pages to capture the narrative, metrics, and figure references. Take notes directly in the scaffold.
• Code reconnaissance: read code/repo_context.md (plus repo_manifest.json) to understand repo layout, commands, configs, and practical pitfalls worth mentioning in the blog.
• Extra context: skim anything the user left in sources/ and cite it explicitly if used.
• Outline: fill every TODO inside article_scaffold.md so each section already contains bullet-point evidence (source + snippet). This keeps the drafting phase deterministic and makes reviews easy.

3. Draft & Polish the Article

Use the structure in references/blog_outline.md:

1. Metadata/title block in Chinese followed by the hero figure.
2. ## 5-minute TL;DR with bullet proof points referencing both paper and repo.
3. Sections 1-6 covering task motivation, method storyline, training/experiments, code reproduction guide, learner takeaways, and paper-vs-code differences.
4. "References" citing the paper and any supporting DOIs/links.

Writing tips:

• Everything stays in Simplified Chinese with a friendly-but-technical tone (see references/writing_spec.md for precise requirements, including title rules, section depth, citations, and limited use of code/inline formulas).
• Target length: 5000-7000 non-whitespace characters. Use scripts/check_article_length.py during QA to keep output stable.
• Use scripts/check_article_requirements.py to enforce mechanical constraints (title/metadata, references section, ban paper/pages, and optionally require https image URLs after Lsky sync).
• Do a final human self-review pass with references/self_review.md to ensure the article "explains clearly" (no guessing, evidence-backed claims, readable structure, figure-text consistency).
• Alternate paragraphs and bullet lists so readers can skim.
• Whenever you mention a claim, cite both the PDF section (page or figure) and the repo file/command that corroborates it.
• Keep article_scaffold.md alongside the final <slug>_blog.md so reviewers can trace back to sources.

4. Figures & Asset Logistics

Manage visuals according to references/figures.md:

• Select 4-6 figures: one front-matter image, at least one method diagram, several experiment/ablation plots or tables.
• Draft using local cropped images from paper/figures/ (and paper/front_matter/). Then run scripts/sync_lsky_images.py to upload via the installed lsky-uploader skill, update .lsky_upload_cache.json, and rewrite the blog Markdown to use returned URLs.
• If downstream platforms need bundled PNGs, mirror the used figures under salad_blog_assets/images/figN.png.
• Double-check all links resolve before handing off.

5. Journal Metrics (JCR + CAS)

If you need accurate journal impact factor / quartiles and CAS partitions, use the local SQLite DB generated from your Excel tables:

• Build/update DB from your Excel source: scripts/build_journal_metrics_db.py --overwrite
• Query by journal name or ISSN: scripts/query_journal_metrics.py --journal "Advanced Science"

The DB file lives at references/journal_metrics_2025.sqlite3. If it is missing, run scripts/build_journal_metrics_db.py --overwrite to generate it locally.

Deliverables Checklist

When you finish a project, ensure the workspace contains:

• Updated article_scaffold.md with your filled outline.
• Final <slug>_blog.md (or blog.md if the user specified one filename) following the outline.
• Updated .lsky_upload_cache.json if any new figures were uploaded.
• Optional: salad_blog_assets/images/ containing all referenced media.

Document blockers (missing repo, corrupt PDF, etc.) at the top of the blog before the TL;DR if you cannot complete a section.

Resources

• references/materials.md - directory map + file purposes.
• references/workflow.md - phased process from intake to QA.
• references/blog_outline.md - mandatory section order, tone, and formatting.
• references/figures.md - figure selection, Lsky uploads, cache maintenance, and optional asset packaging.
• references/writing_spec.md - condensed SML prompt constraints (audience, tone, title format, length, references, image callouts).
• lsky-uploader skill - required for scripts/sync_lsky_images.py (set LSKY_TOKEN before running).
• scripts/build_journal_metrics_db.py - converts your Excel tables into a queryable SQLite DB under references/.
• scripts/query_journal_metrics.py - looks up JCR/CAS metrics for the metadata block.
• scripts/sync_lsky_images.py - uploads local images and rewrites Markdown links to Lsky URLs.
• scripts/check_article_length.py - counts characters and reports per-section breakdown.
• scripts/check_article_requirements.py - validates title/metadata/references and blocks forbidden figure sources.
• references/self_review.md - human checklist for clarity and final polish.
• assets/article_scaffold_template.md - copy this template into each project root to start outlining.