Newsletter AI

Newsletter: AI Intelligence Brief

This skill's job is to produce one edition of the weekly AI intelligence newsletter. All of the editorial substance — role, audience, research window, sourcing rules, the 15-part structure, editorial rules, and writing style — lives in assets/PROMPT.md in this same directory. This file only says how to use it; assets/PROMPT.md is the single source of truth for what the newsletter contains and how it reads, so don't reconstruct the brief from memory or summarize it — read the file itself.

Steps

1. Read assets/PROMPT.md in full before writing anything. Treat it as the complete editorial brief for this edition, not a template to paraphrase.
2. Parse the invocation for flags:
   • Output format — -o/--output followed by md, yaml, or yml; or the shorthand --md / --yaml / --yml. No flag means md (the original behavior). An unrecognized value is an error — ask the user to pick md or yaml rather than guessing.
   • Report path (yaml/yml output only) — -r/--report followed by a path: either a folder (gets report.html + report.yaml inside it) or an explicit .html/.htm file path (gets a .yaml sidecar with the same basename next to it). Given with md output, it's a no-op — say so rather than silently ignoring it.
3. Do the research it calls for: AI developments from the last 7 days, pulling in 30-90 days of prior context only where a development needs it to make sense. Use current web research and prefer the primary sources assets/PROMPT.md lists (company announcements, papers, model cards, technical reports, repos, earnings, regulatory filings) over secondary reporting; cross-check important claims across multiple sources.
4. Write the newsletter following assets/PROMPT.md's structure and rules exactly, in order: the Top 20 Developments, then Parts 2-15 (Model Intelligence through The Big Picture), applying the Editorial Rules and Writing Style sections throughout. Skip a section's content honestly ("No material development this week.") rather than manufacturing a story to fill it.
5. Keep confirmed fact, company claim, reported information, and analyst inference clearly distinguished throughout, as assets/PROMPT.md requires — don't blur them for narrative flow.
6. Attach a real source to every sourced claim, and never invent one. Whenever a development traces to an actual URL you found during research, carry that URL forward as a citation (see the YAML sources field below, or an inline link in markdown mode). When no verifiable URL exists for a claim, say so in the text instead of fabricating a link — a missing citation is honest; a fake one is not.
7. Produce the output in the requested format:
   • md (default): return the finished edition as markdown chat output, structured per assets/PROMPT.md, unless the user asks for it to be saved to a file. Cite sources as ordinary markdown links.
   • yaml/yml: instead of prose, produce a single YAML document conforming exactly to the schema builder/report_data.py validates (see its module docstring and validate() for the authoritative field list) — a meta block plus an ordered sections list (one kind: home landing section, then one kind: standard section per Part, each with items that carry title, an icon/badge from the template's set, and whichever of body / facts / table / stats / kv fits that item's content, plus sources: [{label, url}] for every claim with a real link). Use assets/templates/sample-report.yaml as a fully worked example — copy its shape rather than inventing a new one. Do not hand-number citations or write a references list yourself: list each item's sources, and the render step below dedupes them by URL and builds the numbering and per-tab References accordions automatically, which is what guarantees every citation link actually resolves.

     • No -r/--report given: write the YAML to a file (or return it as a fenced code block if the user hasn't said where), and mention it can be rendered into the interactive report with:

       python skills/newsletters/newsletter-ai/builder/build_report.py <file>.yaml <output>.html
       

     • -r/--report PATH given: run the render yourself:

       python skills/newsletters/newsletter-ai/builder/build_report.py <file>.yaml -r PATH
       

       This writes the rendered HTML and a copy of the YAML side by side (see Report path above for how PATH resolves), and by default fuses the source YAML into the HTML too (a hidden, base64-encoded <script> blob) so the single .html file remains fully reproducible even without its .yaml sidecar — the page's "Download data" button in the top bar lets a viewer pull that YAML back out. Pass --no-fuse only if the user explicitly doesn't want the data embedded. Report the paths written back to the user.

Development

See meta/MAINTAINERS.md for this skill's layout, the YAML schema, how to update assets/PROMPT.md or assets/templates/report.html, how to run the test suite, and versioning conventions. It is not read as part of carrying out a /newsletter-ai request — don't act on it while producing an edition.