section-drafting

Announce after each subsubsection: "Self-review GATE for X.Y.Z: I will not begin X.Y.(Z+1) until I have critically reviewed X.Y.Z, identified weaknesses, and revised them."

A gate violation is: starting research for the next subsubsection before the current one is saved and self-reviewed, drafting prose for multiple subsubsections in one pass, or skipping the self-criticism step.

The full flow is:

for Z in [1, 2, …, last subsubsection]:
    Step 4  → Reference-guided research for X.Y.Z
    Step 5  → Figure discovery for X.Y.Z
    Step 6  → Draft writing (three-layer pipeline)
    Step 7  → Self-review & self-criticism (MANDATORY — revise before proceeding)
    Save revised draft to file
    ── SELF-QUALITY GATE: announce gate, confirm quality, then proceed ──

# After all subsubsections pass their self-quality gates:
Write section introduction (text before X.Y.1)
Verify inter-subsubsection transitions
Step 8  → Coherence check on the assembled section (fix issues)
Step 9  → Review pass: run the review skill, apply fixes
Step 10 → Present the revised section to the user

/draft X.Y.Z — Subsubsection Level

Draft a single subsubsection. Before writing, read existing context:

• Subsection intro (X.Y text, if present) — understand the framing
• Previous subsubsections (X.Y.1 … X.Y.(Z-1)) — ensure continuity
• Next subsubsection outline (X.Y.(Z+1) from chapter_outline.md) — set up proper transitions

Workflow Steps

Step 0: Parse Input & Determine Mode

1. Parse the user's /draft command:
   • X introduction → chapter introduction mode
   • X.Y → subsection mode
   • X.Y.Z → subsubsection mode
2. Map chapter X → directory chapter_0X/
3. Locate the target in chapter_outline.md
4. If chapter introduction (X introduction): proceed to draft the X.0 block — read the full chapter outline and thesis connections, skip per-subsection research
5. If subsection (X.Y): ask user for intro-only or full-section mode
6. If subsubsection (X.Y.Z): proceed directly

Step 1: Context Loading

1. Read outline.md — thesis arc, how this chapter connects to predecessors and successors
2. Read chapter_outline.md — section's Goal, Narrative, key points, transitions, style notes
3. Read references.md — key references with provenance (✅ directly queryable vs ❌ cited-within-review)
4. Read existing drafts in chapter_XX/sections/:
   • Subsection intro text (if drafting a subsubsection)
   • Previous subsubsections (for continuity)
   • Next subsubsection outline (for transitions)

Step 2: Source Registry

→ Call source_registry skill
→ Obtain: author_paper_ids (001), review_ids (002), general_ids

Step 3: Knowledge Retrieval

→ Call knowledge skill (retrieve mode)
→ Input: section topic keywords
→ Output: existing insights that can skip or accelerate research

Step 4: Reference-Guided NotebookLM Research

Research is guided by references.md — the literature groundwork done during /chapter N. Start with the sources already identified as relevant to this section, then broaden.

1. Parse references.md for this section: Read Section 3 ("References Breakdown by Section") to find the key sources listed for this specific section (X.Y or X.Y.Z). From the Reference Data Table (Section 4), extract ✅ Direct Sources (individually queryable in NotebookLM) and ❌ Referenced Sources (cited within reviews — note which review(s) to query from the "Cited In" column). Build a prioritized query plan: these sources are queried first.

2. Priority queries — section-specific sources:

   # For ✅ Direct Sources: query them directly
   mcp_notebooklm_notebook_query(
       notebook_id="1b7df790-7858-4fc8-879c-39f41238c4ae",
       query="<specific question about a key point in section X.Y>",
       source_ids=<direct_source_ids_from_references_md>
   )
   
   # For ❌ Referenced Sources: query the review(s) that cite them
   mcp_notebooklm_notebook_query(
       notebook_id="1b7df790-7858-4fc8-879c-39f41238c4ae",
       query="What does this review say about [Referenced Paper] regarding [topic]?",
       source_ids=<citing_review_source_ids>
   )
   

3. Broadening queries — fill gaps: After priority queries, identify any key points from the outline that still lack sufficient detail. For these gaps:

   • Reviews (002) source IDs): consensus definitions, equations, structural context
   • All sources (omit source_ids): specific citations, numbers, arXiv IDs

4. Author papers (if applicable): Only if this section references the author's own work (001) source IDs).

5. Track provenance: For each claim, record which source backs it → \cite{bib_key}. Prefer bib keys already in references.md over new ones.

Step 5: Figure Discovery

Before drafting, identify and download figures from the literature that would strengthen this section.

1. Query NotebookLM for figure recommendations:

   mcp_notebooklm_notebook_query(
       notebook_id="1b7df790-7858-4fc8-879c-39f41238c4ae",
       query="For the topics covered in section [X.Y] about [topic],
              which published figures are commonly used to illustrate
              these concepts? For each, provide the paper's arXiv ID,
              figure number, and a brief description.",
       source_ids=<review_002_ids>
   )
   

2. Cross-reference with references.md Figure Candidates Table (if literature_research Phase 3 was run for this chapter) — merge any pre-identified candidates.

3. Attempt download for each candidate figure:

   # Get figure URLs from InspireHEP
   mcp_inspirehep_get_paper_figures(arxiv_id="https://arxiv.org/abs/<ID>")
   
   # Match the target figure number from the returned list
   
   # Download to chapter directory
   # run_command: curl -L -o "chapter_XX/figures/<bib_key>_figN.png" "<figure_url>"
   

   See paper_lookup Recipe 3 for full details.

4. Handle failures: If the figure is not indexed on InspireHEP, record it for a % TODO placeholder in the draft (see Step 6 fallback below).

Step 6: Draft Writing (Three-Layer Pipeline)

1. Layer 1 (scientific-writing): Outline key points from Step 4 research, then convert to flowing paragraphs
2. Layer 2 (humanizer): Anti-AI pass — strip patterns, vary rhythm
3. Layer 3 (personal style): Apply voice, physics-first motivation, clarity check — break overlong sentences

Output format (LaTeX-ready markdown):

• Equations: $$ for display, $ for inline
• Citations: \cite{bib_key}
• Cross-references: \label{sec:X.Y.Z}, \ref{sec:...}
• No bullet points in final prose — everything in flowing paragraphs

Literature figures: Insert figures identified in Step 5. For each successfully downloaded figure, use:

\begin{figure}[t]
    \centering
    \includegraphics[width=\columnwidth]{figures/<bib_key>_figN.png}
    \caption{<Description of what the figure shows>, from \cite{<bib_key>}, Figure~N.}
    \label{fig:<descriptive_label>}
\end{figure}

If a figure was not available on InspireHEP (Step 5 failure), insert a placeholder with a % TODO comment:

\begin{figure}[t]
    \centering
    % TODO: insert Figure 3 from arXiv:2302.01947 (Amerio et al., 2023)
    \includegraphics[width=\columnwidth]{figures/placeholder.pdf}
    \caption{Reconstructed source-count distribution from \cite{Amerio:2023uet}, Figure~3.}
    \label{fig:dnds_result}
\end{figure}

The % TODO comment must include: (1) the figure number, (2) the arXiv ID or DOI, and (3) the author/year for easy lookup. The user will replace the placeholder with the actual figure file.

Step 7: Self-Review & Quality Gate (per subsubsection)

This step runs after every subsubsection draft and is the quality gate that must pass before proceeding.