html-research
Skill0 starsChanged 3 months agoUnlicenseView on GitHub
Research a topic on the web and render one self-contained light-theme HTML report with searchable cards, clickable detail views, handwritten diagrams, file paths or GitHub links, and clickable sources. Use when the user runs /html-research or asks for a research report, a deep dive, a write-up, or how big tech solves a problem.
In this repository16 other files it gives its agents.
- claude-agents-md/AGENTS.mdAGENTS.md
- claude-agents-md/CLAUDE.mdCLAUDE.md
- bash-skillSkill
- bug-recordingSkill
- bundle-sizeSkill
- changelog-skillSkill
- container-skillSkill
- fact-checkerSkill
- git-gh-skillSkill
- index-html-skillSkill
- macos-electron-appsSkill
- pr-splitterSkill
- readme-skillSkill
- rerenderSkill
- scripts-skillSkill
- stacks-skillSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
13 rules. About 0 are common, 4 are distinctive, and 9 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
This author's own choices. Worth a look.
9
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 15
When the topic points at a local project, read the real files.
- Very specific: line 19
Every report is saved to /Users/diegopacheco/git/diegopacheco/html-research/.
- Distinctive: line 26
Run web research. Do not answer from memory.
- Distinctive: line 27
If the topic is an engineering matter (rate limiting, auth, caching, sharding, cloud security, observability, deploys, feature flags, …
- Very specific: line 32
If the topic touches a local repo, collect the real absolute paths.
- Very specific: line 36
Run python3 scripts/check_report.py <report.html> and fix everything it reports.
- Very specific: line 43
No CDN, no external stylesheet, no external script, no web font, no remote image.
- Very specific: line 44
Only <a href="https://..."> links to sources are allowed to point outside the file.
- Very specific: line 50
Every factual claim carries a clickable source.
- Distinctive: line 53
Never fabricate a URL. If you did not open it, do not cite it.
- Very specific: line 70
Add bullets, files, and links where they apply.
- Very specific: line 93
Every report MUST have a What The Industry Does group.
- Very specific: line 107
Use handwritten-style inline SVG.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
HTML Research
Turn a topic into one standalone HTML report backed by real web research.
Inputs
The topic is $ARGUMENTS. If it is empty, ask the user what to research and stop.
The topic can be a subject (rate limiting, auth, cloud security), a codebase question, or a mix. When the topic points at a local project, read the real files. Never invent paths.
Output Location
Every report is saved to /Users/diegopacheco/git/diegopacheco/html-research/.
Filename: {topic-slug}-{MM-yyyy}.html, lowercase, hyphenated.
Workflow
- Restate the topic in one line and name the angle you will research.
- Run web research. Do not answer from memory. Use several searches and fetch the strongest pages.
- If the topic is an engineering matter (rate limiting, auth, caching, sharding, cloud security, observability, deploys, feature flags, secrets, queues, and similar), you MUST research how big companies actually do it. Cover at least four of: Google, Amazon/AWS, Meta, Netflix, Stripe, Microsoft/Azure, GitHub, LinkedIn, Cloudflare, Shopify, Uber, or a relevant AI lab (Anthropic, OpenAI, DeepMind). Prefer engineering blogs, official docs, RFCs, papers, and talks.
- If the topic touches a local repo, collect the real absolute paths. If the file lives in a public
GitHub repo, also collect its
https://github.com/{org}/{repo}/blob/{ref}/{path}URL. - Copy
templates/report-template.htmlto the output path and fill theREPORTobject. - Draw diagrams only where they explain something a paragraph cannot.
- Run
python3 scripts/check_report.py <report.html>and fix everything it reports. - Run
bash scripts/open_report.sh <report.html>. - Print the full absolute path of the report as the last line of your answer.
Hard Rules
- One file. No companion CSS, JS, JSON, images, or fonts. Everything inline.
- No CDN, no external stylesheet, no external script, no web font, no remote image.
Only
<a href="https://...">links to sources are allowed to point outside the file. - Light theme only. Never a dark palette, never a
prefers-color-scheme: darkblock, never a theme toggle. - Embedded favicon via a
data:image/svg+xml<link rel="icon">. Keep it under 400 bytes. - Everything searchable through the single search box. Every finding is a card. Cards open a detail modal. Nothing important lives only in the modal-less page body.
- Every factual claim carries a clickable source.
sourcesat the bottom is mandatory, and each card that makes a claim gets its ownlinksentry. - Files are shown as full absolute paths. Link to GitHub when the file is public.
- Never fabricate a URL. If you did not open it, do not cite it.
Cards
Group cards so the chips are useful. A good report has 12 to 30 cards across groups such as:
Overview— what the thing is, why it matters, the tradeoff spaceWhat The Industry Does— mandatory, see belowPatterns/Algorithms/Mechanisms— the actual techniquesFailure Modes— what breaks in productionIn This Codebase— real files, real paths, when a repo is involvedRecommendations— what to do, ranked
Set groupOrder on the report to the order the sections should render in. Overview first,
What The Industry Does second.
Each card needs group, title, summary, tags, and a body of two to five {h, p} blocks.
Add bullets, files, and links where they apply. The card teaser is the hook;
the modal carries the substance.
Pros, cons and gaps are mandatory
Every card carries:
pros— exactly 5 concrete upsides of the thing that card is aboutcons— exactly 5 concrete downsidesgaps— 5 to 10 things it does not solve, does not cover, or leaves open
Rules for these three lists:
- One line each, no sub-clauses stacked with semicolons. A pro is a claim, not a paragraph.
- Concrete beats generic. "413k pending pages froze writes" beats "can be slow".
- A con is not a disguised pro. If you cannot find 5 real downsides, you do not understand the thing well enough yet — go back to the research.
- Gaps are about the world, not about your report: missing tooling, missing data, unanswered questions, things no vendor does yet, measurements nobody published.
- Omit
gapsonly when a card genuinely has none to state. Never omitprosorcons.
What The Industry Does
Every report MUST have a What The Industry Does group. It answers one question: how do the
companies with the hardest version of this problem actually solve it?
- One card per company or lab, with its concrete approach, real numbers, and its own sources.
- Cover at least 4 of: Google, Amazon/AWS, Meta, Netflix, Stripe, Microsoft/Azure, GitHub, LinkedIn, Cloudflare, Shopify, Uber, GitLab, Salesforce.
- Cover at least 1 AI lab or AI-platform angle: Anthropic, OpenAI, DeepMind, or the AI platform layer of a cloud vendor, when the topic touches models, agents, search or safety.
- Include one card naming where the public evidence runs out. Say plainly which companies publish nothing on this and do not invent coverage for them.
- Vendor docs show the win, incident reports show the cost. Cite both when both exist.
Diagrams
Use handwritten-style inline SVG. templates/handwritten-diagram.svg is the reference: a
feTurbulence + feDisplacementMap rough filter, a cursive font stack with real fallbacks,
white or #fbeee8 boxes, #1c1b19 strokes.
Diagram rules:
- Zero crossing arrows. Zero overlapping boxes. If the layout needs a crossing, split it in two diagrams.
- One flow direction, left to right or top to bottom, never both in one diagram.
- At most 7 boxes. Number the arrows when order matters.
- Set an explicit
viewBoxand leave 20px of margin around the drawing. - Skip the diagram entirely rather than shipping a tangled one.
Style
Plain language. No hype, no filler, no "in today's fast-paced world". Short paragraphs. Concrete numbers, limits, and names beat adjectives. State uncertainty when sources disagree.
Final Answer
End with the report path on its own line so the user can click it:
/Users/diegopacheco/git/diegopacheco/html-research/{file}.html
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

