Nsf Thrust Refiner

If the request is primarily a citation pass, asks for explicit citation density, requires careful external verification, or needs new entries added to <proposal-dir>/bibs/working.bib, hand off that part to nsf-bibref-filler instead of recreating its workflow inside thrust-refiner.

Read these references only as needed:

• refinement modes: references/refinement-modes.md
• local literature grounding: references/local-literature-grounding.md
• figure planning patterns: references/figure-patterns.md
• technical depth patterns: references/technical-depth-patterns.md
• unit quality checklist: references/unit-quality-checklist.md
• house style and example anchors from composer when needed: ../nsf-proposal-composer/references/house-style.md, ../nsf-proposal-composer/references/example-anchors.md, ../nsf-proposal-composer/references/running-examples.md

When literature grounding touches overview, motivation, or background paragraphs, do not let the team's own papers become the primary support for broad field or infrastructure claims. Use external anchors first and use the team's own work mainly for direct technical inheritance or prior capability.

When thrust-refiner adds or rewrites local sectioning commands such as \subsection{} or \subsubsection{}, inherit the heading style owned by nsf-proposal-composer: use Title Case and do not leave content words in sentence case.

In a multi-proposal repo, use <proposal-dir> to mean one concrete proposal workspace such as proposals/nsf-25-533-fairos.

If solicitation constraints materially affect the unit, consult <proposal-dir>/guardrail/solicitation-qc.md before patching prose.

When a numbered unit file contains % composer:sync-* and % refiner:body-* markers, treat the refiner:body region as the primary editing target. Only patch the sync region when the refined body makes the top-level framing stale.

Workflow

1. Gather Only The Inputs Needed For The Target Unit

Read the smallest set of files that support the refinement request:

• <proposal-dir>/11-aim1.tex through <proposal-dir>/14-aim4.tex
• <proposal-dir>/00-project-description.tex for intro framing and the pre-unit bridge
• <proposal-dir>/20-evaluation.tex if the refinement changes measurable outcomes
• <proposal-dir>/00-project-summary.tex only if the user wants summary sync
• <proposal-dir>/guardrail/solicitation-qc.md when the request touches solicitation-facing criteria such as measurable outcomes, cyberinfrastructure, adoption, or open-science impact
• relevant .bib files and a small number of prior example units

Do not reload the whole proposal unless the user explicitly asks for a global rewrite.

2. Choose One Or More Refinement Modes

Use one or more of these modes, but keep the patch focused:

1. gap-tightening
2. literature-grounding
3. technical-deepening
4. local-evaluation
5. figure-planning
6. sync-followthrough

See references/refinement-modes.md for what each mode should produce.

3. Ground Literature Locally First

Before inventing new citations, prefer local evidence sources:

• <proposal-dir>/bibs/*.bib
• template/bibs/*.bib
• example proposal .bib files under examples/proposals/
• citations already used in neighboring units

Use scripts/search_local_bib.py when the user asks for related work or when a unit clearly needs stronger grounding.

This step is for lightweight unit-level grounding only. thrust-refiner may add or tighten a few citations that are directly adjacent to the prose it is rewriting, but it should not become a full citation-management pass.

Escalate to nsf-bibref-filler when any of these are true:

• the user specifies sparse, moderate, or heavy citation density
• the section needs a broad citation sweep rather than 1-3 local fixes
• external literature needs to be verified and added to working.bib
• the main task is checking citation accuracy rather than deepening the method

Rules:

• prefer cite keys already available in the active proposal bibliography
• use prior example proposals to learn what literature categories typically support this kind of unit
• if a broad framing claim is currently supported mostly by the team's own papers, treat that as incomplete grounding and look for better external anchors instead
• if the relevant paper title is known but the exact cite key is uncertain, leave a visible \todo{Add citation for "..."}
• do not silently insert fabricated cite keys
• do not duplicate nsf-bibref-filler's density-control or external-entry workflow inside this skill

4. Deepen The Unit Without Changing The Whole Proposal

When refining a thrust or aim, keep the persuasive burden local:

• sharpen the problem and gap
• make the technical object explicit: data structure, representation, signal, algorithmic move, workflow, or system component
• clarify what is learned, built, measured, or released
• add local dependencies on earlier or later thrusts when helpful
• preserve the unit title and role unless the user asks to rethink it

If reviewer comprehension risk is high because the unit relies on abstract infrastructure objects or layered handoffs, consider tightening one shared running example instead of adding more definitions. Reuse the same example across nearby units when possible, and use a compact callout or pseudo-artifact sketch only when it improves scanability. Treat this as optional; do not force every refined unit to contain a box.

If the body change implies a narrower or sharper high-level claim, make the smallest corresponding sync-region change needed for coherence instead of rewriting the whole unit.

Do not stop at prose expansion when the method is still under-specified. Use references/technical-depth-patterns.md to add the kind of lightweight formalism that matches the proposal family and the user's prior house style.

Do not turn one unit into a new proposal direction without checking whether it breaks the intro, summary, or neighboring units.

5. Add Figure Thinking Only When It Compresses Logic

Use figures when they clarify mechanism, workflow, or evaluation logic better than prose alone.

If the target unit has no figure and the mechanism is pipeline-heavy, dependency-heavy, or infrastructure-heavy, the default behavior should be:

1. propose one concrete figure concept
2. add a placeholder in the unit
3. leave a visible \todo{...} for the final asset

Typical outputs:

• refine an existing figure caption
• add a figure placeholder with a filename and caption skeleton
• add a visible \todo{Draft figure showing ...}
• describe one concrete figure concept in prose before drawing anything

See references/figure-patterns.md.

6. Leave Visible TODOs For Human Follow-Up

This skill must make unresolved issues easy for human coauthors to see.

Use visible LaTeX TODOs for:

• uncertain cite keys
• collaborator-provided facts not yet confirmed
• external validator names not yet final
• figure assets not yet drawn
• open choices that need PI review

Prefer \todo{...} over hidden % comments for anything the coauthor team still needs to notice.

7. Sync Only The Nearby Files That Actually Changed

By default, patch the target unit file only.

Patch nearby files only when the refinement clearly changes them:

• 00-project-description.tex if the unit title, local promise, or bridge is now stale
• 20-evaluation.tex if the unit's measurable outcomes materially changed
• 00-project-summary.tex only when the user asks for summary sync

If the change is local, keep it local.

Output Style

When using this skill, prefer one of these outcomes:

• a refined unit file patch
• a short refinement plan followed by the patch
• a literature-grounding patch plus visible TODOs for uncertain cites
• a figure concept with caption placeholder and integration point

Keep prose aligned with the repository's existing house style. Reuse the function of prior examples, not their wording.