# Technical Paper Writer > Technical Paper Writer - Skill: `matteotitta/technical-paper-writer` (Agent Skill) - Install (CLI): `npx skillmds@latest add matteotitta/technical-paper-writer` - Raw SKILL.md: https://api.skillmd.com/api/skills/matteotitta/technical-paper-writer/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: matteotitta (https://skillmd.com/u/matteotitta) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/matteotitta/technical-paper-writer --- # Technical Paper Writer Render first-party research as an arXiv-style technical paper. A stats-heavy paper in preprint format gets cited — by journalists, by other content, and by LLMs, which prefer a single authoritative source with defensible numbers over loose vendor claims. The paper becomes the citable centre of gravity for a data-led PR programme. **Requires the tectonic LaTeX engine** (`brew install tectonic` on macOS; `cargo install tectonic` or the distro package elsewhere) and network access on the first run — it downloads its package bundle then. A slow or offline first run is that download, not a hang. ## Doctrine inherited Output complies with: - [`output-tenets.md`](../../../../../rules/output-tenets.md) — the seven tenets - [`evidence-bound-outputs.md`](../../../../../rules/evidence-bound-outputs.md) — every claim carries its source, or its confidence drops - [`quantitative-evidence-floors.md`](../../../../../rules/quantitative-evidence-floors.md) — state the volume floor, name the confounds and lag. A paper is the one format where this is load-bearing: the Methods and Limitations sections exist to carry it - [`ontology.md`](../../../../../rules/ontology.md) — confidence levels + attribution - [`ai-speak-anti-patterns.md`](../../../../../rules/ai-speak-anti-patterns.md) — a paper in robot prose reads as generated, which is the opposite of the goal **Boundary with `/thought-leadership`:** that skill *writes* the whitepaper — argument-led prose, contrarian angle, three human gates. This skill *renders* an existing study as a paper. They're sequential, not rival. "Write a whitepaper on X" → `/thought-leadership`. "Turn this study into a paper" → here. ## When to run - A whitepaper, benchmark study, survey, or internal dataset exists and needs a citable, research-grade rendering - Building the anchor asset for data-led PR / link earning - NOT for marketing collateral, one-pagers, or content without real data — the format only works if the substance is research ## Steps ### 1. Intake Collect: source material (findings, tables, method notes), author name(s) + affiliation + contact, whether this re-renders an earlier publication (if so: original date, sponsor, research partners), intended hosting URL (own domain under `/research/` works well). ### 2. Fact-gate (mandatory, before any LaTeX) Every headline claim must trace to a named primary source. For any market-sizing claim ("X million...", "Y% of..."), run the **denominator test**: divide the claim by its total population. If the share implied is implausible for the stated definition, the claim is definitionally unsound even if a source exists. > **Worked example.** "Two-thirds of households hold £100k+ in investable assets" — only true if "investable" quietly includes property and pensions. A source exists; the claim is still unsound. It gets replaced with a sourced one, or cut. Unsourced or unsound claims get replaced or cut. Never rendered. ### 3. Structure Follow the premium reference. Required elements, in order: | Element | Rule | |---|---| | Title + subtitle | Findings-oriented, method in the subtitle | | Author block | Name / affiliation / URL / email, one column per author | | Date + `\thanks` footnote | Provenance: sponsor, research partners, and what changed if re-rendering an earlier publication | | Abstract | Findings-led, numbers up front, one paragraph | | Keywords | 4-6 | | Introduction | Ends with a numbered **Contributions** list | | Data and Methods | One subsection per method or data source | | Results | One claim per subsection heading; booktabs tables | | Discussion | What the findings mean; second-order observations | | **Limitations and threats to validity** | Mandatory — this section IS the credibility mechanism. Include a "commissioned research" paragraph when true. | | Conclusion | Restate findings with numbers | | Acknowledgements + References | Every stat in the paper maps to a `\citep` entry | **Worked example — finding to claim-heading:** - Input (source material): "Survey: 62% of failed deployments traced back to configuration drift" - Output (Results subsection): `\subsection{Configuration drift dominates deployment failures}` — the heading asserts the finding; the prose and table under it carry the number and the source. ### 4. Build Copy the premium reference into a dedicated working folder as `main.tex` — **never edit the file in the premium reference** (it is the reusable template). Then `tectonic main.tex`. Notes: the first-ever run downloads tectonic's package bundle over the network (slow, fails offline — not a hang); `\bibitem` optional args must keep the `Source(Year)` shape or natbib errors. Fix every error **and every overfull hbox** (wide tables: `\footnotesize` + tighter `\tabcolsep`; long URLs: the template loads `xurl`). Then **read the compiled PDF** — page 1 and the references page — and confirm each stat cites the right source. An overfull table is invisible in the compile-log summary. ### 5. Integrity guards - "Pending on arXiv" may only be claimed after an actual submission exists. - Commissioned or sponsored research must say so, in the footnote and in Limitations. - A re-rendering of earlier work must declare what changed and what didn't. ## What good looks like **Evaluations** — Every statistic maps to a `\bibitem`. Every market-sizing claim passed the denominator test. Limitations section present and specific (not "further research is needed"). Provenance footnote names sponsor + partners where they exist. Sample sizes and confounds stated in Methods per `quantitative-evidence-floors.md`. The compiled PDF was read, not just compiled. No claim in the paper is softer or harder than its evidence. ## Common mistakes - Rendering marketing claims in academic dress — the format amplifies scrutiny, not just credibility - Skipping Limitations to "look stronger" — it works the other way: the Limitations section is what makes commissioned research citable - Wide tables overflowing a two-column layout — check the compiled PDF visually, not just the compile exit code - Citing one source for a sentence containing three claims — split the citations - Claiming arXiv status before submitting --- ## Final ship gate Run `/premortem --output` before ship. See [`/premortem` skill](../../../../meta/orchestration/premortem/SKILL.md) for the 5 execution domains (will-it-resonate / will-it-convert / will-it-stay-on-brand / will-stakeholder-push-back / will-it-degrade-over-time) and output template. Then run `/voice-reviewer` — the content ship gate: voice + brand quality (pm-loop.md). Trivial-case escape: `## Premortem\nNo failure modes — trivial change` satisfies the contract for genuinely trivial outputs. --- ## Persuasion & stickiness pass Output complies with [persuasion-and-stickiness.md](../../../../../rules/persuasion-and-stickiness.md) — Cialdini's 7 persuasion levers + Heath's SUCCESs. A paper is the narrow case: **Authority** and **Credible** carry nearly all the load, and both are earned by the Limitations section and the bibliography, not by adjectives. **Concrete** applies to the Results headings — the heading asserts the finding, in numbers. Do not reach for Scarcity, Emotional, or Stories here; a preprint that reaches for them stops reading as research, which is the whole asset. ## Attribution Contributed by Ben Moore (Discovered Labs / ClientCo), July 2026. The fact-gate (incl. the denominator test), paper structure table, template, and integrity guards are his. Adapted to Genesys frontmatter, doctrine, and gates.