Redundancy Pruner Agent Skill

Redundancy Pruner

Purpose: make the survey feel intentional by removing “looped template paragraphs” and consolidating global disclaimers, while keeping meaning and citations stable.

Role cards (use explicitly)

Compressor

Mission: remove repeated boilerplate without deleting subsection-specific work.

Do:

• Collapse repeated disclaimers into one front-matter paragraph (not per-H3 repeats).
• Delete repeated narration stems and empty glue sentences.
• Keep each H3’s unique contrasts/evaluation anchors/limitations intact.

Avoid:

• Cutting unique comparisons because they *sound* similar.
• Turning pruning into a rewrite (this skill is subtraction-first).

Narrative Keeper

Mission: keep the argument chain readable after pruning.

Do:

• Replace slide-like navigation with short argument bridges (NO new facts/citations).
• Ensure each H3 still has a thesis, contrasts, and at least one limitation.

Avoid:

• Generic transitions that could fit any subsection ("Moreover", "Next") without concrete nouns.

Role prompt: Boilerplate Pruner (editor)

You are pruning redundancy from a survey draft.

Your job is to remove repeated boilerplate and make transitions content-bearing, without changing meaning or citations.

Constraints:
- do not add/remove citation keys
- do not move citations across ### subsections
- do not delete subsection-specific comparisons, evaluation anchors, or limitations

Style:
- delete narration and generic glue
- keep one evidence-policy paragraph in front matter; avoid repeated disclaimers

Inputs

• output/DRAFT.md
• Optional (helps avoid accidental drift):
• outline/outline.yml (subsection boundaries)
• output/citation_anchors.prepolish.jsonl (if you are enforcing anchoring)

Outputs

• output/DRAFT.md (in-place edits)

Workflow

Use the role cards above.

Steps:

1) Identify repeated boilerplate (not content):

• repeated disclaimer paragraphs (evidence-policy, methodology caveats)
• repeated opener labels (e.g., Key takeaway: spam)
• repeated slide-like narration stems (e.g., “In the next section…”) and generic transitions

2) Pick a single home for global disclaimers:

• keep the evidence-policy paragraph once in front matter (Introduction or Related Work)
• delete duplicates inside H3 subsections

3) Rewrite transitions into argument bridges:

• keep bridges subsection-specific (use concrete nouns from that subsection)
• do not add facts or citations

4) Sanity check subsection integrity:

• each H3 still has its unique thesis + contrasts + limitation
• no citation-only lines and no trailing citation-dump paragraphs
• if outline/outline.yml exists, use it to confirm you did not prune across subsection boundaries
• if output/citation_anchors.prepolish.jsonl exists, treat it as a regression anchor (no cross-subsection citation drift)

Guardrails (do not violate)

• Do not add/remove citation keys.
• Do not move citations across ### subsections.
• Do not delete subsection-specific comparisons, evaluation anchors, or limitations.

Mini examples (rewrite intentions; do not add facts)

Repeated disclaimer -> keep once:

• Bad (repeated across many H3s): Claims remain provisional under abstract-only evidence.
• Better (once in front matter): state evidence policy as survey methodology, then delete duplicates in H3.

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.

Template synthesis stem -> content-first sentence:

• Bad: Taken together, these approaches... (repeated many times)
• Better: state the specific pattern directly (e.g., Across reported protocols, X trades off Y against Z...).

Troubleshooting

Issue: pruning removes subsection-specific content

Fix:

• Restrict edits to obviously repeated boilerplate; keep anything that encodes a unique comparison/limitation for that subsection.

Issue: pruning changes citation placement

Fix:

• Undo; citations must remain in the same subsection and keys must not change.