wanshuiyin/paper-talk

Paper Talk: End-to-End Conference Talk Pipeline

Workflow: from a completed paper to a conference-ready talk artifact — slide outline, Beamer source, editable PPTX, speaker notes, full talk script, polished visuals, and assurance checks.

Pipeline target: $ARGUMENTS

Assurance Ladder

The skill takes an explicit — assurance: argument; default is polished.

| Level | Phases run | Use when | |---|---|---| | draft | 0 → 1 → 2 → 5 → 6 | Internal practice talk; quick first-pass deck | | polished (default) | 0 → 1 → 2 → 3 → 5 → 6 | Lab seminar, workshop, default conference talk | | conference-ready (or submission) | 0 → 1 → 2 → 3 → 4 → 5 → 6 | Top-venue oral / spotlight where slide claims are scrutinised; or anonymous submission |

Phase 4 (assurance checks) is opt-in via — assurance: conference-ready. The lower levels skip claim / citation / anonymity audits, which is correct when content has already been audited at the paper-writing stage.

Hard Invariants

These are non-negotiable across all phases:

1. Original paper is read-only. This workflow consumes the paper directory; it never modifies paper/main.tex or other paper artifacts.
2. Original deck is preserved. /paper-slides produces a baseline deck; /slides-polish writes a _polished versioned copy. The original Beamer + PPTX are never overwritten.
3. Speaker notes are byte-stable. Polish must not change slide.notes_slide content. Phase 4 verifies this.
4. No new content anywhere in the pipeline. All slide text, speaker notes, talk script, Q&A answers, claims, numbers, citations, URLs, author names, affiliations, anonymity placeholders, and experiment results must be either paper-grounded (extracted from PAPER_DIR/ artefacts) or explicitly user-provided. The pipeline never invents content during outline, build, polish, audit, or export. Phase-4 anonymity scan + claim audit verify this end-to-end.
5. No slide reordering. Add / drop / reorder requires explicit user flags.
6. Cross-model independence. Per-page Codex calls in /slides-polish use fresh spawn_agent calls (no send_input). See ../shared-references/reviewer-independence.md.
7. Anonymity fail-closed. If any audit (or any Codex fix proposal) would replace a placeholder with a real title / count / URL, the workflow halts and surfaces the proposal for human review. See ../shared-references/experiment-integrity.md.
8. Style references are guidance, not text source. A — reference: PDF or — style: preset informs visual weight and structural rhythm; never copy prose, examples, slide titles, or speaker-note text from the reference.
9. Final report cannot be conference-ready unless required audits pass. Phase 6 verifies and downgrades verdict if audits fail.
10. reasoning_effort: xhigh is invariant across all effort levels for any Codex call invoked by sub-skills.

Constants

• PAPER_DIR = paper/ — Source paper directory. Override via positional argument.
• OUTPUT_DIR = slides/ — Where the deck artefacts live.
• STATE_DIR = .aris/paper-talk/ — Workflow state, audit logs, final report.
• TALK_TYPE = spotlight — Default talk format. Inherited by /paper-slides.
• TALK_MINUTES = 15 — Default duration. Inherited by /paper-slides.
• VENUE = NeurIPS — Default venue (used by /paper-slides color schemes when — style is not passed).
• ASPECT_RATIO = 16:9 — Inherited by /paper-slides.
• STYLE_PRESET — generic if not passed; why-rf and venue presets supported by /slides-polish.
• REFERENCE_VISUAL — Required when assurance ≥ polished. The Beamer compile of this talk is an acceptable self-reference; an external academic talk PDF is preferred when the user wants visual alignment beyond defaults.
• AUTO_PROCEED = false — Each major phase pauses for user approval. Set true only when explicitly requested.

Inputs

Discovered from $ARGUMENTS and the project directory:

1. PAPER_DIR/ — compiled paper (main.tex + main.pdf). The paper must already be in good shape; this workflow does not write the paper.
2. PAPER_DIR/sections/*.tex — section sources for content extraction.
3. PAPER_DIR/figures/ — figures available for slide reuse.
4. (optional) slides/SLIDE_OUTLINE.md — pre-existing outline (resume mode); skips Phase 1 generation.
5. (optional) — reference: <pdf> — visual anchor for /slides-polish. If absent and the assurance level requires polish, the workflow uses the Beamer self-compile as reference.
6. (optional) — talk_type, — minutes, — style, — effort — see Constants.

Output Layout

File names match /paper-slides's emit contract — this workflow is a non-renaming consumer.

slides/
├── SLIDE_OUTLINE.md                # Phase-1 outline (claim-first per slide)
├── main.tex                        # Beamer source (from /paper-slides)
├── main.pdf                        # compiled Beamer
├── presentation.pptx               # editable PPTX (from /paper-slides; emit name fixed by /paper-slides)
├── presentation_pre_polish.pptx    # snapshot before /slides-polish
├── presentation_polished.pptx      # /slides-polish output (when assurance ≥ polished)
├── presentation_polished.pdf       # rendered (LibreOffice or user-export)
├── speaker_notes.md                # per-slide notes (from /paper-slides)
├── TALK_SCRIPT.md                  # full word-for-word talk script + Q&A (from /paper-slides)
└── assets/                         # per-slide PNG previews

.aris/paper-talk/
├── PIPELINE_STATE.json             # phase pointer, status, timestamps
├── FINAL_REPORT.md                 # human-readable summary at end
├── audit-input/                    # Phase-4 staging copies of slide text + notes + script (so /paper-claim-audit and /citation-audit can run on slide content as a synthetic "paper")
│   ├── main.tex                    # \input{sections/slide_text.tex} \input{sections/notes.tex} \input{sections/script.tex}
│   ├── sections/
│   │   ├── slide_text.tex          # all visible slide bullets/titles/callouts as \section{Slide K}
│   │   ├── notes.tex               # speaker_notes.md → LaTeX, one \section per slide
│   │   └── script.tex              # TALK_SCRIPT.md → LaTeX, one \section per slide
│   ├── references.bib → ../../../paper/references.bib   # symlink to source paper bib for /citation-audit
│   ├── results/ → ../../../paper/results/                # symlink — claim audit needs real evidence
│   └── figures/ → ../../../paper/figures/                # symlink — citation audit may reference figure captions
└── audits/
    ├── slide_claim_audit.json      # 6-state verdict per claim (PASS/WARN/FAIL/NOT_APPLICABLE/BLOCKED/ERROR)
    ├── citation_audit.json         # citation existence / metadata
    ├── anonymity_scan.json         # placeholder discipline check
    └── export_integrity.json       # page-count / aspect / blank-slide / notes-preserved

The audit JSON files follow the shared 6-state schema; see ../shared-references/assurance-contract.md.

Workflow

Phase 0: Setup

1. Validate paper directory: ls $PAPER_DIR/main.tex $PAPER_DIR/main.pdf. If absent → halt and ask user to point to the compiled paper.
2. Check prerequisites:
   • LaTeX: which xelatex pdflatex latexmk
   • PPTX rendering: which soffice (LibreOffice headless) — required for Phase 4 export integrity check; otherwise prompt user to export PDF manually.
   • PDF tools: which pdfinfo pdftoppm (poppler) — required for /slides-polish PNG rendering.
   • Codex MCP availability.
   • python-pptx (python3 -c 'import pptx').
3. Resolve overrides from $ARGUMENTS: talk_type, minutes, assurance, reference, style, effort.
4. State init: write .aris/paper-talk/PIPELINE_STATE.json with phase: 0, timestamp, all resolved overrides.
5. Resume mode: if slides/SLIDE_OUTLINE.md exists and PIPELINE_STATE.json shows recent in-progress work, prompt the user — resume from last phase or start fresh.

Phase 1: Slide Outline (Checkpoint)

Goal: produce or accept a slides/SLIDE_OUTLINE.md that the user signs off on before any deck-building happens.

If slides/SLIDE_OUTLINE.md already exists, present its summary (slide count, time budget, claim-per-slide map) to the user and ask:

"Use existing outline (Y/n)? Modify? Regenerate?"

Otherwise, delegate to /paper-slides Phase-1 only (content extraction + slide outline generation). /paper-slides writes the outline into its own state; we adopt it as slides/SLIDE_OUTLINE.md.

The outline must contain, per slide:

• A claim-first title (states the slide's point, not the topic).
• One main idea per slide.
• Bullet content (sentence fragments, not paragraphs).
• Figure / equation / callout note.
• Time budget (seconds).
• Transition cue.
• Speaker note seed (1-3 sentences; expanded in Phase 2).

Checkpoint: present the outline. Default behavior: pause for user approval. Set AUTO_PROCEED = true only when the user is explicitly running unattended.

Phase 2: Build Baseline Deck

Invoke /paper-slides to generate Beamer source + PPTX from the approved outline.

/paper-slides "<paper-dir>" — talk_type: <T> — minutes: <N> — venue: <V> — aspect: 16:9 — notes: true

Forward — style-ref: if the user originally passed it (writer-side; the reviewer-side polish in Phase 3 must NOT see it — see invariant 8).

/paper-slides produces (emit names fixed by /paper-slides):

• slides/main.tex + slides/main.pdf (Beamer)
• slides/presentation.pptx (editable PPTX)
• slides/speaker_notes.md
• slides/TALK_SCRIPT.md

After return: verify the four artefacts exist. If any missing, halt with a diagnostic. Do not rename outputs — downstream phases bind to these exact paths.

Phase 3: Visual Polish (when assurance ≥ polished)

Skip when assurance == draft.

Invoke /slides-polish against the freshly generated PPTX, with a reference. If the user passed — reference:, use that. Otherwise, use the Beamer compile (slides/main.pdf) as the visual reference — the Beamer is the design intent for the same talk, so it is the correct anchor when no external reference is available.

/slides-polish "slides/presentation.pptx" — reference: slides/main.pdf — style: <preset> — effort: <effort>

After return: verify slides/presentation_polished.pptx exists; if rendering tools are available, also verify slides/presentation_polished.pdf.

The polish phase is read-only on content (see Hard Invariants). If /slides-polish emits a BLOCKED verdict because a Codex fix proposal would alter content, surface that block to the user and halt rather than overriding.

Phase 4: Assurance Checks (when assurance == conference-ready)

Skip when assurance ∈ {draft, polished}.

/paper-claim-audit and /citation-audit expect a paper-shaped input (paper/main.tex + paper/sections/*.tex + paper/results/) and emit JSON under paper/. Slides have a different shape, so this phase first stages slide artefacts into a synthetic paper directory, then invokes the audits against that synthetic input. Verdicts are written under .aris/paper-talk/audits/ using the shared 6-state schema (PASS, WARN, FAIL, NOT_APPLICABLE, BLOCKED, ERROR) — see ../shared-references/assurance-contract.md.

4.0 Staging adapter

The staged "paper" mirrors /paper-claim-audit and /citation-audit's expected layout: main.tex at the root that \input{}s sections/*.tex, plus a .bib file and results/+figures/ symlinks.

.aris/paper-talk/audit-input/
├── main.tex
│   # contains: \input{sections/slide_text.tex}
│   #           \input{sections/notes.tex}
│   #           \input{sections/script.tex}
│   #           \bibliography{references}
├── sections/
│   ├── slide_text.tex      # one \section{Slide K — <title>} per slide; visible bullets, titles, callout text, in-slide \cite{...}
│   ├── notes.tex           # speaker_notes.md converted to LaTeX, one \section per slide
│   └── script.tex          # TALK_SCRIPT.md converted to LaTeX, one \section per slide
├── references.bib → ../../../paper/references.bib  # symlink to source paper bib (or copy if symlink not allowed)
├── results/ → ../../../paper/results/              # symlink — claim audit needs real evidence
└── figures/ → ../../../paper/figures/              # symlink — citation audit may reference figure captions

If the source paper uses a different bib name (e.g., paper.bib), mirror the actual file. If multiple bibs are used, link them all.

The staged "paper" exists only for the duration of Phase 4 and is removed in Phase 6 unless the user passes — keep-audit-input.

4.1 Slide claim audit

Invoke /paper-claim-audit against the staged input. Scope is slide text + speaker notes + full talk script — talks often smuggle unsupported claims into spoken parts that the visible bullets don't show.

/paper-claim-audit ".aris/paper-talk/audit-input"

The audit emits audit-input/PAPER_CLAIM_AUDIT.json with the shared 6-state schema. Move it to .aris/paper-talk/audits/slide_claim_audit.json and de-stage the path prefix so the verdicts cite slide K rather than synthetic-paper sections.

A FAIL or BLOCKED verdict on any claim downgrades the Phase-6 final verdict from conference-ready to polished.

4.2 Citation audit

Invoke /citation-audit over the staged input. Verify any \cite{...} in slides + notes + script via DBLP / CrossRef; flag fabricated entries.

/citation-audit ".aris/paper-talk/audit-input"

Output → .aris/paper-talk/audits/citation_audit.json (6-state).

4.3 Anonymity scan

When the talk is for an anonymous-submission venue or the user passed — anonymous, scan slides + notes + script for:

• Real paper titles where placeholders should be ([anonymized], Withheld for anonymous review, etc.).
• Real author names beyond the talk's own author block.
• Real submission counts.
• Real URLs that would deanonymize.

Output → .aris/paper-talk/audits/anonymity_scan.json (6-state). Any FAIL (real-content leak) blocks conference-ready.

Phase 5: Final Export + Integrity

Resolve the final PPTX path:

FINAL_PPTX = slides/presentation_polished.pptx  if Phase 3 ran
           = slides/presentation.pptx           if assurance == draft (Phase 3 skipped)

Then:

1. Recompile Beamer cleanly: cd slides && latexmk -pdf -xelatex main.tex (or pdflatex for non-CJK). Confirm main.pdf page count matches outline slide count.
2. Render FINAL_PPTX → PDF via soffice --headless --convert-to pdf <FINAL_PPTX> -outdir slides/. If unavailable, prompt user.
3. Export integrity check → .aris/paper-talk/audits/export_integrity.json (6-state):
   • PPTX-PDF page count == Beamer-PDF page count.
   • Aspect ratio == declared (16:9 default).
   • No fully-blank slide.
   • Speaker notes preserved on every slide that had them in the baseline. When Phase 3 ran, sha256 match against presentation_pre_polish.pptx's notes; when Phase 3 skipped, against presentation.pptx's own notes (no-op).
   • Embedded fonts are appropriate for projector use (e.g., Calibri / Helvetica Neue / PingFang SC for Chinese).

If any integrity check fails, downgrade the Phase-6 verdict and surface specifics to the user.

Phase 6: Final Report

Write .aris/paper-talk/FINAL_REPORT.md with:

• Verdict: draft | polished | conference-ready (downgraded if audits failed).
• Artefact paths: Beamer source / PDF, baseline PPTX, polished PPTX, polished PDF, notes, script.
• Slide count, time budget vs target.
• Audit summary (one line per audit run + PASS / WARN / FAIL).
• Open warnings: any unresolved Codex polish notes the user should review by hand.
• Next steps: e.g., "drop in real QR images on slide N", "verify HF Daily Papers screenshot", "rehearse to confirm 25-min budget".

A final conference-ready verdict requires:

• Phases 0-5 all completed without halt.
• All Phase-4 audit JSONs returned PASS (or NOT_APPLICABLE when the audit's content detector did not match — e.g., no citations on slides). Any WARN, FAIL, BLOCKED, or ERROR downgrades the verdict.
• Phase-5 export integrity returned PASS.
• No anonymity-scan FAIL when anonymous submission is declared.

Otherwise the report emits the highest passing level (polished or draft) and itemises why conference-ready was not granted, citing the specific audit JSON and the failing verdict line.

Effort Levels

See ../shared-references/effort-contract.md.

| effort | Behavior | |---|---| | lite | Forwarded to sub-skills. /slides-polish runs in lite (BLOCKERS only). Phase 4 still runs at the requested assurance level — effort controls depth, never audit gating (see ../shared-references/assurance-contract.md). | | balanced (default) | Standard pipeline. /slides-polish runs balanced. | | max | /slides-polish runs max (per-page review on every slide). Phase-4 audits read all artefacts in full. | | beast | /slides-polish runs beast (second polish round). Phase-4 audits include extended assurance checks; final report adds an executive summary. |

reasoning_effort: xhigh is invariant.

assurance and effort are independent axes. The user may legally combine effort: lite, assurance: conference-ready to mean "fast pipeline, but every audit must emit a verdict before finalisation."

Parameter Pass-Through

/paper-talk "paper/" — talk_type: oral — minutes: 25 — assurance: conference-ready \
            — reference: ./refs/why_rf_2025.pdf — style: why-rf — effort: max

Forwarded to:

• /paper-slides (Phase 2): paper-dir, talk_type, minutes, venue, aspect, notes, style-ref if writer-side.
• /slides-polish (Phase 3): polished pptx target, reference, style, effort.
• /paper-claim-audit (Phase 4.1): scoped to slide deck artefacts (not paper).
• /citation-audit (Phase 4.2): scoped to slide-deck-only citations.
• Anonymity scan (Phase 4.3): inline tool, no external skill.

When NOT to Use

• The paper is not yet compiled or claims are not stable. Run /paper-writing first.
• The user wants a poster, not a talk. Use /paper-poster-html.
• The user already has a finished deck and only needs visual polish. Use /slides-polish directly — /paper-talk would needlessly rebuild.
• The talk content is unrelated to a paper (general lecture, demo). The orchestration assumes a paper-grounded talk; for ad-hoc decks, use /paper-slides directly with manual outline.

Prior Skill Relationship

• Composes /paper-slides, /slides-polish, /paper-claim-audit, /citation-audit.
• Does not call /kill-argument by default — that is upstream intellectual stress-test, not deck QA. Users who want talk-story stress-test before slide build can run /kill-argument paper/ first.
• Sister workflow to /paper-writing (paper) and /paper-poster-html (poster).

Empirical Origin

This workflow generalises a 30+ iteration polish run on a Chinese-spoken academic conference talk (May 2026). The convergent learning was that talk preparation has the same shape as paper preparation — plan, build, polish, audit, export, report — and benefits from the same assurance ladder. The per-page Codex polish pass (Phase 3) is the single most expensive but highest-value step, and is hidden behind the polished default so users get it without thinking about it.