Transition Weaver (LLM-first; NO NEW FACTS)

Purpose: produce a small, low-risk “transition map” so adjacent subsections do not read like islands.

This skill is intentionally low-risk and deterministic:

• the script writes outline/transitions.md from adjacent subsection titles + brief bridge handles
• the generated transitions stay paper-safe and avoid adding new facts or citations

Transitions should answer:

• what the previous unit established
• what gap/tension remains
• why the next unit follows

Injection contract (treat transitions as draft text)

outline/transitions.md is not planning notes: section-merger injects it into output/DRAFT.md. So each transition line must be safe to read as paper prose.

Format contract (for merge insertion):

• Only lines matching - 3.1 -> 3.2: <text> are inserted by default (within-chapter H3 -> next H3).
• Keep <text> as one sentence without list formatting.

Notes:

• Use the ASCII arrow -> (not a unicode arrow) to avoid invisible/control-character encoding issues.
• section-merger accepts both -> and → for backward compatibility, but -> is the preferred contract.

Hard rules:

• Write the transition sentence as final prose: content-bearing, not process-bearing.
• No planner-talk openers: avoid stems like "To keep ...", "The remaining uncertainty is ...", "setting up a cleaner ...".
• No slash-list axis labels (A / B / C; planning/memory). Rewrite using natural prose.
• Keep it short: one sentence is preferred; rarely two.
• Avoid semicolon-heavy multi-clause construction notes.

Rewrite triggers (if you see these, rewrite):

• "To keep ..." / "We next focus on ..." / "The remaining uncertainty is ..."
• "as the comparison lens" / "reference point" / "to make the next trade-offs easier to interpret"

Role prompt: Linker (coherence without narration)

text
You are the coherence linker for a survey.

Your job is to write short, content-bearing transitions between adjacent subsections:
- restate what was established (one clause)
- name the remaining tension/gap (one clause)
- justify why the next subsection is the right lens (one clause)

Style:
- argument bridge, not navigation
- no “Now we discuss / Next we move / In this section…”
- no semicolon planning notes

Constraints:
- NO NEW FACTS
- NO citations
- only reuse handles that already exist (titles, RQs, bridge_terms)

Style targets (paper-like, still NO NEW FACTS):

• Prefer argument bridges: content-bearing sentences, not outline narration.
• Keep it short (often 1 sentence).
• Avoid title narration once merged: do not write “From Section A to Section B”.
• Avoid “Now we discuss / Next we introduce / In this section we ...”.

CRITICAL: Transitions must be real content sentences, NOT construction notes.

• Bad: “After X, Y makes the bridge explicit via …; …; setting up a cleaner A-vs-B comparison.”
• Good: “While loop design determines what actions are possible, tool interfaces define how those actions are grounded in executable APIs and orchestration policies.”

Also avoid (reads like axis/planning notes once merged):

• Slash-list axis labels (e.g., A/B/C, planning/memory); rewrite using natural prose (and/or).

Inputs

• outline/outline.yml (ordering + titles)
• outline/subsection_briefs.jsonl (expects rq and optional bridge_terms/contrast_hook)

Output

• outline/transitions.md (used by section-merger; keep paper voice)

Workflow (NO NEW FACTS)

1) Read outline/outline.yml to determine adjacency (which H3 follows which). 2) Read outline/subsection_briefs.jsonl to extract each subsection’s rq and any bridge handles (bridge_terms, contrast_hook). 3) For each boundary, write 1–2 transition sentences:

• no new facts
• no citations
• no explicit “we organize this section as …” meta narration
• no placeholders (TODO, …, <!-- SCAFFOLD -->)

4) Write outline/transitions.md.

Role cards (use explicitly)

Linker (argument bridge)

Mission: write short, content-bearing transitions without narration.

Do:

• Restate what was established (one clause).
• Name the remaining tension/gap (one clause).
• Justify why the next unit follows (one clause).

Avoid:

• Title narration ("From X to Y") and slide navigation ("Now we turn").
• Semicolon planning notes or meta commentary.

Skeptic (template killer)

Mission: delete anything that reads like construction notes.

Do:

• Remove generic transitions that could fit any subsection.
• Force subsection-specific nouns from titles/RQs/bridge terms.

Avoid:

• Smuggling new facts into transitions.

Script (generator + validation)

You usually do not run this manually; it exists so a pipeline runner can deterministically generate and validate the artifact.

Quick Start

• uv run python .codex/skills/transition-weaver/scripts/run.py --workspace <workspace>

All Options

• --workspace <dir>: workspace root
• --unit-id <U###>: unit id (optional; for logs)
• --inputs <semicolon-separated>: override inputs (rare; prefer defaults)
• --outputs <semicolon-separated>: override outputs (rare; default validates outline/transitions.md)
• --checkpoint <C#>: checkpoint id (optional; for logs)

Examples

• Validate after you write outline/transitions.md:
• uv run python .codex/skills/transition-weaver/scripts/run.py --workspace <workspace>

Troubleshooting

Issue: transitions read like templates

Fix:

• Ensure subsection briefs include subsection-specific bridge signals (bridge_terms / contrast_hook).
• Rewrite the transitions to mention those handles (as content, not as axis-label lists).

Note: insertion is explicit

By default, section-merger does not insert generated transitions. Create outline/transitions.insert_h3.ok for within-chapter H3 transitions or outline/transitions.insert_h2.ok for between-H2 transitions, then run the post-merge voice gate.