{"slug":"claesbackman-paper-version","name":"paper-version","description":"Convert a LaTeX research paper into a policy brief, 1-page summary, or 5-page summary for a general audience, with factual review and a standalone HTML page for GitHub Pages.","long_description":"---\nname: paper-version\ndescription: Convert a LaTeX research paper into a policy brief, 1-page summary, or 5-page summary for a general audience, with factual review and a standalone HTML page for GitHub Pages.\nuser-invocable: true\nargument-hint: [brief|1page|5page]\nallowed-tools: Read, Write, Edit, Glob, Grep, Bash, Agent\ndisable-model-invocation: true\n---\n\n# Paper Versions\n\nConvert a LaTeX research paper project into an accessible audience-facing version, review it for accuracy, and produce a standalone HTML page.\n\n## Input\n\n- `$ARGUMENTS[0]` — Output format: `brief`, `1page`, or `5page` (required)\n\n## Format definitions\n\n- **`brief`** — A structured 2-page policy brief with fixed sections: *The Question*, *What We Do*, *Key Findings*, *Policy Implications*, *Caveats*. Formal but accessible. No jargon.\n- **`1page`** — One flowing page of prose for a general audience. No section headers. Written like a short article. Leads with the main finding.\n- **`5page`** — A five-page narrative summary with light structure (Background, What We Did, What We Found, Why It Matters, Limitations). Accessible to an educated non-specialist.\n\n## Instructions\n\n### Step 0 — Validate input\n\nIf `$ARGUMENTS[0]` is not one of `brief`, `1page`, `5page`, stop and tell the user: \"Please specify a format: `brief`, `1page`, or `5page`.\"\n\nSet `FORMAT` = the argument value.\n\n---\n\n### Step 1 — Reader Agent\n\nLaunch an Agent (subagent_type: general-purpose) with this task:\n\n> You are reading a LaTeX research paper project. Your job is to extract the full content of the paper into a clean, structured plain-text representation that will be used by a writer agent.\n>\n> **Instructions:**\n> 1. Find the main `.tex` file in the current directory. It is usually the file that contains `\\documentclass` or `\\begin{document}`. Use Glob to search for `**/*.tex` files, then identify the root file.\n> 2. Read the root `.tex` file. Wherever you find `\\input{...}` or `\\include{...}` commands, read those files too, recursively, until you have assembled the full paper text.\n> 3. Strip LaTeX markup: remove commands like `\\textbf{}`, `\\emph{}`, `\\cite{}`, `\\label{}`, `\\ref{}`, `\\footnote{}`, equation environments, figure environments, table environments. Keep the text content. For tables, summarize the key numbers in prose form. For figures, note what the figure shows based on the caption.\n> 4. Extract and clearly label these components:\n> - **Title**\n> - **Authors**\n> - **Abstract** (verbatim)\n> - **Introduction** (full text)\n> - **Data/Empirical Setting** (if present)\n> - **Methods/Approach** (condensed)\n> - **Main Results** — list every key quantitative finding, with exact numbers, units, and confidence intervals as stated in the paper. Be exhaustive here.\n> - **Robustness/Heterogeneity** (condensed)\n> - **Conclusion**\n> - **Key claims made by the authors** — list every causal or interpretive claim the authors make explicitly, noting the exact language they use (e.g., \"we find that X causes Y\" vs. \"X is associated with Y\")\n> 5. Extract figure information: scan the full `.tex` source for `\\includegraphics` commands. For each one, record:\n> - The filename as specified (e.g., `figures/fig1` or `fig_results`)\n> - The caption text from the nearest `\\caption{}` command\n> - Which section of the paper the figure appears in\n> List these under a **Figures** section in your output, formatted as:\n> `FIGURE: [filename] | CAPTION: [caption text] | SECTION: [section name]`\n> 6. Identify the 1-2 figures that best illustrate the paper's main finding and mark them with `[KEY FIGURE]`.\n> 7. Return the full extraction as structured markdown. Do not summarize or interpret — just extract.\n\nSave the Reader Agent's output to `output/paper_extraction.md`. Create the `output/` directory if it does not exist.\n\nAfter the Reader Agent finishes, scan the project directory for image files that match the extracted figure filenames. Use Glob to search for `**/*.png`, `**/*.jpg`, and `**/*.jpeg`. For each figure listed as `[KEY FIGURE]` in `paper_extraction.md`, check whether a matching PNG or JPG file exists (try the filename with and without extension, and with common path prefixes like `figures/`, `Figures/`, `fig/`). Build a list of embeddable figures — those where a `.png` or `.jpg` file was found — and save it to `output/figures_list.md` in this format:\n\n```\nEMBEDDABLE: [relative path from project root] | CAPTION: [caption text]\nNOT FOUND (PDF or missing): [filename] | CAPTION: [caption text]\n```\n\nIf no PNG/JPG figures are found at all, note that in `output/figures_list.md` and continue — the HTML will be text-only.\n\n---\n\n### Step 2 — Writer Agent\n\nLaunch an Agent (subagent_type: general-purpose) with this task, passing it the content of `output/paper_extraction.md` and the value of FORMAT:\n\n> You are a science writer creating an accessible version of an economics research paper for a general audience.\n>\n> **Paper content:** [paste full content of output/paper_extraction.md]\n>\n> **Output format requested:** [FORMAT]\n>\n> **Format instructions:**\n>\n> For `brief` — Write a structured policy brief. Use exactly these section headers:\n> - **The Question** — What problem does this paper address? (2-3 sentences)\n> - **What We Do** — How do the authors study it? Data, setting, method in plain language. (3-4 sentences)\n> - **Key Findings** — The main quantitative results. Report actual numbers. (4-6 bullet points, each 1-2 sentences)\n> - **Policy Implications** — What do these findings suggest for policy? Be concrete. (3-5 sentences)\n> - **Caveats** — Limitations the authors themselves acknowledge. (2-3 sentences)\n> Total length: approximately 500-600 words.\n>\n> For `1page` — Write one page of flowing prose. No section headers. Open with the main finding stated plainly. Use accessible analogies where helpful. Do not use academic hedging or jargon. End with why it matters. Total length: approximately 350-400 words.\n>\n> For `5page` — Write a 5-page narrative summary with these light headers:\n> - **Background** — Context and motivation\n> - **What We Did** — Data, setting, approach\n> - **What We Found** — Results in detail, with numbers\n> - **Why It Matters** — Implications\n> - **Limitations** — What the paper cannot establish\n> Total length: approximately 1400-1600 words.\n>\n> **General writing rules (all formats):**\n> - Write for an intelligent adult with no economics background\n> - Report key numbers — do not vague them out\n> - Do not overclaim causality beyond what the paper itself claims — match the paper's own language precisely\n> - No passive voice where avoidable\n> - No bullet points except where explicitly called for above\n> - No jargon without explanation\n>\n> **Key findings callout:** Identify the single most important quantitative finding — the one number or result a reader should walk away remembering. Write it as a short punchy sentence (max 25 words). Mark it clearly in your output with the tag `[CALLOUT]: ` at the start of the line, e.g.:\n> `[CALLOUT]: Homeowners in the top wealth decile hold 40% of all housing wealth, despite representing only 10% of households.`\n> This will be displayed as a highlighted box on the webpage.\n>\n> Return only the finished draft (including the [CALLOUT] line), with no preamble or meta-commentary.\n\nSave the Writer Agent's output to `output/[FORMAT]_draft.md` (e.g., `output/brief_draft.md`).\n\n---\n\n### Step 3 — Reviewer Agent\n\nLaunch an Agent (subagent_type: general-purpose) with this task:\n\n> You are a fact-checker and editorial reviewer. You will compare a summary/brief of an economics paper against the original paper's extracted content, and produce a structured review report.\n>\n> **Original paper extraction:** [paste full content of output/paper_extraction.md]\n>\n> **Draft to review:** [paste full content of output/[FORMAT]_draft.md]\n>\n> **Your job — check for three categories of issues:**\n>\n> **1. Factual accuracy**\n> Go through every quantitative claim, statistic, or finding stated in the draft. For each one, verify it against the original extraction. Flag anything that:\n> - Uses a wrong number\n> - Changes units or direction of an effect\n> - Attributes a finding to the wrong group or condition\n> - Omits a critical qualifier (e.g., \"only for renters\" or \"only in the short run\")\n>\n> **2. Overclaiming**\n> Compare every causal or interpretive claim in the draft against the exact language in the original paper. Flag anywhere the draft uses stronger language than the paper (e.g., draft says \"causes\" when paper says \"is associated with\"; draft says \"proves\" when paper says \"suggests\").\n>\n> **3. Framing consistency**\n> Flag anywhere the draft:\n> - Shifts the emphasis of the findings away from what the paper presents as primary\n> - Buries or omits a key finding the paper treats as central\n> - Introduces an implication the paper does not make\n>\n> **Output format:**\n> Produce a numbered list of issues found. For each issue, state:\n> - **Category** (Factual / Overclaiming / Framing)\n> - **Location in draft** (quote the relevant phrase)\n> - **The problem** (what is wrong or overstated)\n> - **Suggested fix** (what it should say instead, with reference to the source)\n>\n> If no issues are found in a category, say so explicitly. End with an overall verdict: PASS (ready to use with minor edits), REVISE (needs targeted fixes before use), or MAJOR REVISION (significant accuracy or framing problems).\n\nSave the Reviewer Agent's output to `output/[FORMAT]_review.md`.\n\n---\n\n### Step 4 — Pause and present review\n\nAfter the Reviewer Agent finishes:\n\n1. Read `output/[FORMAT]_review.md` and display its full contents to the user.\n2. Tell the user: \"The draft is saved at `output/[FORMAT]_draft.md`. Edit it as needed before proceeding. When you are ready to generate the HTML, reply **proceed**. To cancel, reply **cancel**.\"\n3. Wait for the user's response.\n - If the user says **cancel** or anything indicating they want to stop, end the skill.\n - If the user says **proceed** or any clear affirmation, continue to Step 5.\n\n---\n\n### Step 5 — Website Agent\n\nRead the current contents of `output/[FORMAT]_draft.md` (which may now be edited by the user).\n\nLaunch an Agent (subagent_type: general-purpose) with this task:\n\n> You are building a clean standalone HTML page for an economics research paper summary. This page should have no JavaScript dependencies, use a single `