How I Use Claude Subagents to Audit 30 Webflow Pages in 10 Minutes

What Pushed Me to Try Subagents on a Real Project?

Two weeks ago a Bengaluru SaaS client asked me to audit a 30-page Webflow site for AEO readiness before a Google AI Mode push. The old method, prompting Claude one page at a time, would have eaten my entire Thursday. So I tried Anthropic's new dynamic workflows feature instead, the one that ships hundreds of small subagents in parallel and stitches the results. I expected something flashy. What I got was a full 30-page audit completed in nine minutes and forty seconds with usable findings for every URL.

According to Anthropic's launch post for Claude Opus 4.8 on May 28, 2026, dynamic workflows can spin up hundreds of small subagents to deliver results faster than a single long conversation. The pricing stays at $5 per million input tokens and $25 per million output tokens, the same as Opus 4.7. The cost for my 30-page audit ended up at roughly $1.85, which I billed back to the client as a tooling pass-through.

I want to walk through the exact workflow I now use, what to watch out for, and why this is the first AI feature in 2026 that has changed how I quote audit work for clients.

What Is a Claude Subagent and How Does It Differ From Regular Prompting?

A Claude subagent is a small focused instance of the model that handles one slice of a larger task in parallel with other subagents. Instead of one long conversation where Claude reasons about everything, you spawn many shorter conversations that each tackle a single page, file, or question, then merge the results. The orchestrator agent decides how to split the work and how to combine the output.

Anthropic's documentation describes the pattern as a way to handle embarrassingly parallel tasks: things like file-by-file code review, page-by-page content audits, and document-level summarization where each unit is independent. For a 30-page Webflow audit each page is independent of the others, which is the textbook case for subagents.

This is different from a single long prompt because Claude inside a long context window will start losing detail past about 200,000 tokens of working memory, even on Opus 4.8. With subagents each one sees only the page it is responsible for, so it stays fresh and specific. Princeton's GEO-bench research from January 2026 found that focused per-section analysis yields roughly 38% more actionable findings than whole-document passes. Subagents are the practical way to get that focus at scale.

How Do I Set Up the Audit Workflow Step by Step?

The setup is three pieces: a sitemap extractor, a per-page audit prompt, and an orchestrator that fans the work out and reduces the results. I run all three through Claude Code, which has built-in subagent support for the Opus 4.8 model. Total setup time the first time was about 45 minutes. Repeat runs take about 30 seconds to start.

First I run a small Node script that hits the site's sitemap.xml, fetches the raw HTML of every URL, and writes one file per page into a working directory. For a 30-page site this takes about 12 seconds. Second I write a per-page audit prompt that asks for five things: the primary query the page should rank for, missing schema, AEO weaknesses, citation potential on a 1 to 10 scale, and a single recommended rewrite. Third I tell the orchestrator to spawn one subagent per page file and combine the outputs into a single markdown report.

The orchestrator prompt is fifteen lines long. The audit prompt is forty lines long. The whole thing is reusable across clients. I keep both as templates in a folder I version with Git, and I update them whenever I learn something new from a client engagement.

What Do the Findings Actually Look Like?

The output is a markdown file with one section per page. Each section has a citation score, three to five concrete issues, and a recommended rewrite for the weakest paragraph. For the Bengaluru SaaS client, the report flagged 14 pages as citation-weak, 8 as schema-incomplete, and 3 as duplicative. The recommended rewrites were specific enough that a junior writer on their side picked them up and shipped half of them in two days.

What I would not get from a single long conversation is the per-page scoring rigor. With one big prompt Claude tends to summarize and average. With subagents each page is judged on its own and the comparison across pages is much sharper. The client could see exactly which pages were the bottom ten in citation potential, which is the list they actually need.

Why Not Just Use Webflow's Built-In AEO Audit Tool?

Webflow's AEO audit is good for surface-level signals like missing meta tags, broken schema, and obvious thin pages. It costs Workspace AI credits and runs inside the Designer. I run it monthly on retainer clients as the baseline. The Claude subagent pass is what I use on top for the deeper rewrites and the citation scoring that Webflow's tool does not yet give you.

The two are complementary, not competing. Webflow's audit answers what is technically wrong. The subagent audit answers what would actually make AI cite this page. For most client work I now do both, and I include the Claude pass in my flat-fee audit deliverable. The Webflow credit cost is small once you know the patterns. I covered the credit math in my note on the May 2026 pricing reset for context.

How Much Does a 30-Page Audit Actually Cost in API Tokens?

For 30 pages averaging 800 words each, total input tokens land around 280,000 and output tokens around 80,000. At Opus 4.8 pricing that is $1.40 in input and $2.00 in output, total $3.40, minus the prompt caching discount of about 45% that kicks in for repeated runs of the same orchestrator. My last run cost $1.85 net. Even at full price the audit would cost less than a single Semrush keyword pull.

The economics matter because they unlock pricing I could not previously offer. I used to quote 30-page audits at 40,000 rupees because they took me a full day. Now I quote them at 25,000 rupees and finish them in an afternoon, including the time to read the subagent output and edit recommendations into client voice. The margin per hour went up, the price for the client went down, and the deliverable got better.

What Goes Wrong With Subagents and How Do I Catch It?

Subagents fail silently more often than long-context Claude does. If a page returns a 404 or a rendering issue, the subagent for that page will happily score it as having very thin content rather than telling you it could not load the page. The orchestrator does not catch this by default. I now add a sanity check: each subagent must report the word count and H1 of the page it analyzed. If word count is under 50 I treat that page as failed and re-run it manually.

The other gotcha is hallucinated citation scores. Claude will sometimes confidently give a 7 out of 10 to a page that has zero schema and no answer-style content. I cross-check the bottom and top scoring pages against the Webflow AEO report and the actual citations from Perplexity and ChatGPT to make sure the rankings are calibrated. About one in twelve scores gets bumped after the cross-check.

How Do You Get the Same Output Without Writing Code?

You do not need to write code. Claude Code's built-in Task tool exposes subagent spawning behind a simple instruction: use multiple agents to analyze each page in parallel. For a non-developer the workflow is feeding the URL list and the audit prompt into Claude Code as a project, then asking the main agent to delegate one page per subagent and write a combined report. It works but it is slower than a Node script for the fetch step.

If you do not use Claude Code, the Anthropic Console has a similar capability through the Managed Agents API now in general availability. Same model, same parallelism, more setup. For a one-time audit I would just use Claude Code. For a recurring monthly audit I would invest the 45 minutes to wire up the Node fetcher and reuse it. My broader writeup on integrating Claude with Webflow lives in my piece on the Webflow MCP server inside Claude Code.

How to Run Your First Subagent Audit This Week

Pick a single Webflow client site you know well, ideally one with 20 to 40 pages. Pull its sitemap. Open Claude Code with the Opus 4.8 model selected. Tell it to fetch every URL, spawn one subagent per page using a focused AEO audit prompt, and write a combined markdown report scoring each page from 1 to 10 on citation potential. Block out 30 minutes and you will have a report by lunch.

The first audit will teach you what to tighten in the prompt. The second audit will run cleanly. By the third you will know what to charge clients for it. For more on how I structure the Claude side of my workflow, my note on Claude Opus 4.8 dynamic workflows for web teams covers the broader pattern beyond just audits.

If you want help wiring up a subagent audit on your own client work, or just want a worked example, I am happy to walk through it. Let's chat.