Most AI-generated "summaries" of long podcasts are walls of text. On October 6, 2026, Boris Cherny, who works on Claude Code at Anthropic, shared something different: an interactive website companion to the latest Acquired episode on The Home Depot, built by Claude Opus 5.5 from one natural-language prompt. It retells a 3-hour-35-minute episode as 17 chapters, with a tape-measure timeline you can scrub, original watercolor illustrations and a set of small interactive visualizations.
This post describes what the artifact contains, shows the prompt Cherny used, explains why a prompt this short produces something this rich, and gives a method for doing the same with your own material, including the accuracy and copyright cautions that matter. We reviewed the shared artifact text and the thread, and we have not independently fact-checked the episode's claims.
TL;DR: the questions people are asking
| Question | Short answer |
|---|---|
| What is it? | A Claude artifact: an illustrated, interactive companion to the Acquired Home Depot episode. |
| Who made it? | Boris Cherny, using Claude Opus 5.5. |
| How long is the source? | 3 hours 35 minutes, retold in about 17 chapters ("aisles"). |
| Was there a special skill? | No. "All Opus, just ask it to draw watercolor." |
| How are images made? | OpenCV and NumPy code that simulates pigment, paper and wet edges. |
| Did it read the transcript? | Cherny says Claude chose to read it plus other primary sources. |
| Can I copy it? | Yes, the method. Mind accuracy and rights. |
What the page contains
The artifact is called an "illustrated companion to Acquired," retelling the Home Depot story in order and stopping at most chapters to name a business lesson. Features visible in the text:
- A scrubbable timeline. A "tape measure" pinned to the top of the screen tracks position in the 3-hour-35-minute recording, with chapter markers, and you can tap to jump.
- Chapters as aisles. Each chapter is an "aisle" with a timestamp, a narrative, a key statistic and a boxed "lesson."
- Original watercolors for most chapters, such as a warehouse store at dawn or a still life of a pharmacist's counter.
- Interactive data visualizations. A log-scale chart comparing a hypothetical $1,000 investment in the company and in the S&P 500; a stepper walking through a negotiation's price changes; a state-by-state grid that animates store growth over time; a slider model for how supplier payment terms and inventory turns affect who finances the shelf; a margin-versus-volume calculator; a flywheel diagram you can tap through; and a line chart of sales, stores and sales per store over decades.
- A "playbook" summary that lists every lesson in order and links back to its chapter.
- A sources section and a disclosure that the page is an independent companion written in the author's own words and is not affiliated with Acquired or The Home Depot.
The result reads like an interactive magazine feature, which matches what the prompt asked for.
The prompt
Cherny posted the prompt as a screenshot. In summary, he wrote that the episode was "really amazing" and asked Claude to make an artifact that narrates the entire episode, beautiful and interactive, focused on communicating the nuggets of business wisdom throughout. He added: "Make it really good, use lots of tokens and iterate till you're proud of it." Then: maybe add watercolor illustrations and historical photos, aim for New Yorker or New York Times data viz vibes but better, and "I think you can use OpenCV for water color images."
The model began by researching: the screenshot shows it compiling revenue and store-count history and then moving to build. The app also displayed a notice that "Opus draws down usage faster than Sonnet," and the model selector showed Opus 5.5 at an extra-high setting.
Why a short prompt can produce a rich result
Several ingredients in the prompt are doing real work, and you can reuse each.
- A clear unit of value. "The nuggets of business wisdom" tells the model what to extract, which is why every chapter ends in a lesson.
- A style reference. "New Yorker or New York Times data viz vibes" gives a concrete aesthetic target for charts and typography, better than "make it pretty."
- Permission to spend effort. "Use lots of tokens and iterate till you're proud" encourages longer planning, self-review and revision rather than a first draft.
- A suggested tool. Naming OpenCV for watercolors gave the model a feasible path to illustrations without an image-generation API. It could write code that simulates the look.
- Autonomy over sources. Cherny says Claude chose to read the transcript and other primary sources, which explains the filings and book citations at the end.
This fits a pattern we described in our post on thin prompts and thick artifacts: a short instruction that sets a goal and style can produce a large artifact if the model has tools and room to work. It also reflects the growing role of Claude artifacts as shareable, interactive outputs.
How the watercolors work, and what that implies
Asked whether the illustrations were SVG, Cherny said the model used OpenCV. The artifact states the watercolors were produced by a small OpenCV and NumPy program that models pigment, paper texture and wet edges, and that none reproduces a photograph.
That is notable for two reasons. First, it shows a language model producing stylized raster art by writing a rendering algorithm, not by calling an image model. The images are generated by deterministic code, so they can be regenerated and tweaked. Second, the output is original, which sidesteps the usual worry about reproducing copyrighted images, though you should still look at what the code produced. The page notes that for photographs of the early stores, readers should see Acquired's own artifacts, which is the right way to handle third-party images.
The tradeoff is control: algorithmic watercolors have a recognizable look, and complex scenes with people are harder to render convincingly this way. For illustration styles that need fine detail, a dedicated image model may still do better.
A method you can reuse
If you want a companion to a long episode, talk or book chapter, here is a structured version of the same approach.
- Start with material you may use. Use your own content, content you have licensed, or content where the owner allows derivative explainers. Do not feed in a transcript you are not allowed to process or redistribute.
- State the audience and the takeaway. For example: "operators who want practical lessons" or "students new to the topic."
- Specify structure. Chapters tied to timestamps, one idea per chapter, a lesson box and a closing summary.
- Specify interaction types. Ask for a timeline scrubber, a few charts with sliders, a stepper for a story with phases and a flip-card comparison. Do not ask for everything everywhere, since too many widgets distract.
- Reference a visual style. Name publications or designers whose charts you like.
- Allow research and iteration. Say you want it to read primary sources, check numbers and revise before showing you.
- Add rules for facts. Ask it to mark every number with a source and to flag anything it could not verify.
- Ask for accessibility. Keyboard navigation, readable contrast, alt text for illustrations and a reduced-motion option.
- Review in passes. Read the text for accuracy, test each interaction, and check it on a phone.
- Disclose. Say plainly that it is an independent companion, name the source and add a sources list.
Accuracy and rights: what to check
A beautiful page makes errors look authoritative, so verification matters more, not less.
- Numbers. Check every statistic against the primary source: filings, official reports or the episode itself. The page cites sources such as annual reports and a companion PDF. Spot-check the charts, since a log-scale chart with a wrong data point is persuasive and wrong.
- Quotes and stories. Anecdotes retold from a podcast are second-hand. Treat them as the hosts' account, not established fact, and attribute them.
- Computed illustrations. The page includes hypothetical calculations, such as what a stake would be worth today. These are labeled as thought experiments, and any such figure should be labeled the same way.
- Copyright. Summarizing and commenting on an episode in your own words is generally different from republishing its transcript or artwork, but rules vary by country and by how closely you track the source. Avoid long verbatim passages, credit the original and link to it.
- Affiliation. Say you are not affiliated with the creators or the company.
- Sensitive subjects. For topics involving health, legal or financial advice, add a clear disclaimer, as the page does.
Cost and practicalities
The thread's screenshot includes a warning that Opus uses your usage allowance faster than Sonnet, and the prompt explicitly asks for "lots of tokens." A page this ambitious likely consumed a large share of a session. If you are on a limited plan, consider drafting the structure with a cheaper model and reserving the strongest model for the build. Keep the artifact self-contained, since shared artifacts are hosted pages, and be aware of privacy settings, a topic we covered in our note on shared Claude chats and artifacts being indexed.
Why this matters beyond one podcast
The notable shift is the form of the output. Instead of a summary, the model produced an explorable explanation with charts you can manipulate. That is the direction of research on interactive learning, such as the benefits of "doing" described in our post on AI interactive textbooks. It also suggests a new genre for creators: a companion site for every long episode, course or report, generated in an afternoon.
Other recent showcases of Opus 5.5 point the same way, including community builds with generated scenes and simulations and historical discovery experiments. The common thread is that a strong model plus a clear creative brief and room to iterate now yields polished artifacts, not rough drafts.
What this means for what you build or pay
If you teach, publish or run a podcast, the lesson is practical: write a better brief, not more prompts. Specify the value you want extracted, the style, the interaction types and the rules for facts, then let the model iterate. Budget for the usage cost and plan time for human review. If you build products, notice the demand: readers want explorable explanations, and tools that make them reliable, with sourced data, accessible design and clear attribution, have room to grow.
Related reading
- Thin prompts, thick artifacts
- Claude Code artifacts and shareable sessions
- Claude artifacts and MCP connectors
- Shared Claude chats and artifacts: indexing and privacy
- Claude Opus 5.5 community showcase
- Opus 5.5 and a historical discovery experiment
- AI interactive textbooks and the doer effect
- Boris Cherny on Claude Code
Primary: Boris Cherny's thread on X (October 6, 2026) and the shared Claude artifact · the Acquired episode on The Home Depot (the source the companion is based on)
Details are accurate as of October 7, 2026 and come from the shared artifact and thread. We did not independently verify the episode's historical or financial claims, and this post is not affiliated with Acquired, Anthropic or The Home Depot. Do not republish third-party transcripts or artwork without permission.