paper-talk

Files (1)

• SKILL.md

SKILL.md (first lines)

# Paper Talk: End-to-End Conference Talk Pipeline

> **Before each reviewer slot**, follow [Selectable Reviewers](../shared-references/reviewer-backends.md):
> resolve invocation → `.aris/reviewer.json` → Claude Code default (Codex MCP)
> with `reviewer_route.py` through the canonical helper resolver, preserving this
> slot's tier, fresh contexts, source/image inputs and deterministic gates.
> Codex examples/pins below apply only to `backend=codex`; other routes use their
> own first-call/reply contract. Save the route and actual returned identity;
> unavailable selections are `REVIEW_UNAVAILABLE`. Classify acceptance from actual
> families/completion/scope. Forward invocation reviewer/model/effort overrides
> to reviewer-bearing children. Generation workers are unchanged.


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 threads (no `codex-reply`). 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`

From wanshuiyin/Auto-claude-code-research-in-sleep, MIT.

More from wanshuiyin/Auto-claude-code-research-in-sleep