Everyone in your organisation now works next to a capable LLM. The highest-leverage thing you can ask it for is not a wall of text. It is one self-contained HTML file: interactive, printable, attachable to a ticket, and openable on any locked-down laptop.
The thirteen artifacts below are real, working files, generated with an LLM. Each carries the brief that produced it. Run that brief in the model of your choice, whether that is GPT, Claude, Gemini, Grok or DeepSeek, and iterate a few rounds. You can often get broadly similar output, not identical. That is the point: the capability belongs to LLMs, not to any one product.
These artifacts were built with Claude Code, an LLM coding agent, over several rounds of iteration. They are shown as the bar to aim for, not as a guaranteed first answer. Most capable models can get close if you iterate; none will reproduce these byte for byte. Treat each artifact as a first draft to check, not a finished record: an LLM can be confidently wrong, so a person reviews and owns the contents before anyone relies on them. All names and numbers are fictional. ABC Corp does not exist. This is a demonstration pattern, not a production system or security certification.
Thirteen artifacts is a lot to browse. Pick the lane that matches your role and open three or four that will land. You can wander the full set afterwards.
Documents you produce every week, upgraded. No IDE, no repo: paste a brief, save the HTML, open it.
Evidence an agent can write next to the code: reviews, decisions, releases, and a map of the legacy.
The argument itself, as a talk you can run from one file: arrow-key slides with speaker notes.
The documents reviewers work from: a post-mortem, release evidence, and a dependency exposure view — drafted by the agent, checked by you.
Same weekly status, two output formats. Ask the LLM for the one on the right.
# Project Lighthouse — week 24 **Status:** Amber. Wave-2 started 2 days late; decommission sign-off moved to 17 Jul. - Budget: 1.94M of 4.2M (under plan) - Milestone: wave-2 migration, due 26 Jun - Risk: duplicate records (high), API limits (med) - Decision needed: approve front-loading reviews
Same content. One you skim, one you use — print it, attach it to a ticket, open it on a locked-down laptop.
No IDE, no repo, no deploy. You paste a brief into the assistant you already have, save its answer as a
.html file, and open it in a browser. That is the whole workflow. The six artifacts below show
how everyday business documents can be upgraded, and in some cases replaced, by self-contained HTML.
Engineering teams live on evidence: reviews, decisions, incident records, release packs. An LLM coding agent can write that evidence next to the code it describes, as a file in the repo. HTML earns its place here because the result diffs like code, attaches to a change ticket, and still opens in two years with no tooling. The agent drafts the evidence; a person still reviews and signs it — the format is what's dependable here, not the first draft's contents.
The five questions that come up the moment you suggest this inside an enterprise. Short, honest answers.
| Is it secure? | Lower security overhead than a typical web app: no package dependency tree, no third-party runtime, and no network calls in these examples. The whole artifact is inspectable as plain text, but the HTML and JavaScript still need human review before enterprise use. |
| Is it maintainable? | It is plain text. It diffs in git, reviews like code, and has no framework or build step to churn or break next quarter. |
| Does it work as a record? | It attaches to a change ticket or evidence pack as one file, and still opens in two years with nothing but a browser. A person signs off on the contents. |
| Better than a PDF or PowerPoint? | For interactive evidence, often yes: HTML can be filterable, printable and diffable. PDF and PowerPoint are still valid records, but they are weaker when interaction, inspection or version diffs matter. |
| Will every LLM nail it first try? | No — and that honesty is the point. You iterate a few rounds, in any capable model. These are the bar to aim for, not a guaranteed first answer. None will reproduce them byte for byte. |
Six properties of the format itself. Judge them on their merits; no vendor promises them.
Nothing above depends on any vendor. But if your organisation runs Microsoft 365 Copilot or GitHub Copilot, these documented features render or author HTML there, under admin control.
| Microsoft 365 Copilot Pages supports "lightweight apps": create, edit, and preview runnable code directly inside a page. | Microsoft Support |
| Code previews in Copilot Chat and Pages are a Cloud Policy: enabled by default, and your admin can disable or scope it. | Microsoft Learn |
| GitHub Copilot agent mode determines which files to change and proposes edits for the user's approval, governed by enterprise policies. | GitHub Docs · policies |
| With GitHub's data residency enforced, "your code, prompts, and Copilot responses never leave your region during inference processing" (EU and US). Business and Enterprise data is not used for model training. | GitHub Docs · Trust Center |
Thank you, Thariq. The idea that an AI agent should hand you an HTML artifact instead of a wall of markdown comes from Thariq Shihipar's "The unreasonable effectiveness of HTML". Credit for the original insight is entirely his. The enterprise lens, the briefs, and all thirteen examples here are new work inspired by it, not copies of it.
This site, including every artifact on it, was built with Claude Code, an LLM coding agent. Fittingly.