After I shipped a comprehensive SEO fix for rajiv.com and my two synthesis sites, the recovery work crossed Google Search Console, Bing Webmaster Tools, Cloudflare, and the live sites. The build-time check, schema additions, and redirect repair are in my May 5 technical writeup. What followed was an evening of submissions, inspections, cache purges, and checks across several interfaces.
I drove that work through one Claude Code session using MCP browser automation. The session used claude-in-chrome for browser navigation, with curl and git for the parts that did not need a GUI. I decided what to submit, in what order, and how to judge the response. The assistant carried the URL list and the latest observations from one tool to the next.
For engineers and technical leads, the gap between “I know what to do” and “I have to click through three consoles to do it” consumes attention. This session showed me how the Direction Dynamic (the synthesis pattern of human direction and AI execution) works when the deliverable is configured external state.
Human direction across several tools
In synthesis coding, the human leads: setting strategy, providing context, verifying output, and correcting course. The AI executes and surfaces options. Operations needs that same division of responsibility, with the added discipline of checking what happened in an external service.
My loop was:
- Describe the goal at a level a peer engineer could act on.
- Have the assistant navigate the tool, take the authorized action, and read back the response.
- Verify the readback, decide the next step, and direct it.
- Carry that context into the next tool or action.
I stayed responsible for the consequential decisions. The assistant handled navigation, forms, and readbacks. That let the task continue across interfaces without making me repeat the context at each handoff.
The recovery walkthrough
Google Search Console: sitemap resubmission and URL Inspection. I directed the assistant to select the rajiv.com property, open Sitemaps, remove the previous failing entry, submit the new /sitemap.xml, and read back the status. Once it returned “Success,” we moved to eight high-priority URLs I had identified: the homepage, the biography page, the article flagged with a 58 percent impressions drop, three other recent posts, and two articles that had fallen out of the index.
For each URL, the assistant read the indexing status (indexed, “Crawled - currently not indexed,” or “Discovered - currently not indexed”) and requested indexing where appropriate. Inspection reports what Google knows about a page. Requesting indexing asks Google to reconsider it; that request is not a finding that the page is now indexed.
I initially treated “Temporary processing error” in the sitemap field for two unindexed URLs as confirmation of the cause. Google’s URL Inspection documentation describes a temporary problem retrieving sitemap data in its reporting system. The message alone does not explain why a page is absent from the index.
Bing Webmaster Tools: submission and an incomplete import. For rajiv.com, we submitted /sitemap.xml and read back “Processing,” followed by “Success.” The synthesis-site imports from Search Console did not complete. I chose to try them again the following day, leaving them on the follow-up list. The properties had recently been created, but timing alone did not establish why Bing’s import failed.
Cloudflare: redirect verification and cache purge. I had the assistant check the www-to-apex redirect with curl, then open the Pages dashboard to confirm the successful deploy and perform the cache purge. The direct HTTP check and the dashboard served different purposes: one tested the redirect a visitor would receive; the other showed deployment status and provided the purge action.
Live verification. After deployment and the purge, the assistant fetched the sitemap, robots.txt, and a sample article on each of the three sites. We checked the URL contract, structured data, and absolute canonical links. These were checks of the deployed sites; indexing and search impressions needed later observations.
The whole sequence took about an hour. Doing it manually would have taken the same wall time, but I would have had to be in the chair for all of it, switching contexts between three browser tabs and the terminal, copy-pasting URLs, and tracking which submissions had succeeded. With direction, the ops loop ran in the background while I reviewed the output.
What I directed; what the assistant executed
| What I did | What the assistant did |
|---|---|
| Decided what to submit, in what order | Navigated the consoles |
| Chose the eight high-priority URLs | Inspected them, read the indexing state, and requested indexing where appropriate |
| Interpreted the sitemap error | Surfaced the status |
| Chose a later retry for the synthesis-site Bing imports | Reported the failed imports |
| Checked each readback against the expected state | Reported the response |
| Set the post-recovery follow-up plan | Carried out the directed actions |
The operator’s contribution includes interpretation, priorities, and deciding when to stop. Those responsibilities remain even when the assistant can perform every click.
Choose the interface and check before repeating
Google’s Sitemaps API documentation, updated in July 2024, already described programmatic submission before this session.
For a new task, I would check the supported operation, available permissions, and verification method before choosing an interface. An API can make inputs and responses easier to preserve. A browser can expose a workflow the operator needs to inspect directly. For an infrequent recovery, I am comfortable directing work through a dashboard; recurring work is a reason to invest in a script or integration.
Whether an action can be repeated safely depends on the operation, not on whether it uses an API or a GUI. A repeated read and a repeated state change have different consequences in either interface. Check the current state and the previous response before sending another request. If a sitemap is already accepted, decide whether anything has changed that warrants resubmitting it. If an indexing request was already sent, do not repeat it simply because another tool has become available.
My original estimate of roughly ten to twelve daily inspections should not be used as a general Search Console quota. Google’s Search Console API limits list 2,000 URL inspection queries per day and 600 per minute per site. Those are API quotas, not the limits for browser URL inspection or browser indexing requests.
I had the assistant report the number of inspections completed so I could prioritize the remaining URLs. If an interface reports a limit, retain its exact message and the operation that produced it. A hanging request or a retry banner can have several explanations. For an action with side effects, check whether the first attempt took effect before retrying.
Authentication is a firm boundary. When a captcha, two-factor prompt, or another owner-only step appears, the assistant stops that action and asks me to handle it. It can retain the pending work so we resume at the right place.
Keep actions and outcomes in the same record
For recovery work across several services, I would use these fields:
| Field | What belongs there |
|---|---|
| Observation | The exact page, response, or status, with its URL or property and observation time |
| Working diagnosis | The explanation being tested and the evidence that would distinguish it from alternatives |
| Authorized action | The target and permitted change, including any boundary on retries |
| Immediate response | What the service returned after the action |
| Later outcome | The subsequent check needed, its owner, and the result when observed |
A sitemap accepted by a service belongs in the immediate-response field. A later inspection showing a page indexed belongs in the outcome field. A change in impressions needs its own observation period and comparison.
Google describes sitemap submission as a hint, with no guarantee that it will download the sitemap or use it to crawl the listed URLs. That makes the later check part of the work. Google’s sitemap guidance explains the limitation alongside the submission methods.
For a future migration, I would begin the record before the change: save the advertised sitemap URL and its contents, inspect a representative set of URLs, and export impressions with the date range and report filters recorded. I would test the new build before cutover, then repeat the sitemap and URL checks after deployment. Search observations would need a later comparison period, also recorded. These dated observations give an investigation a comparison point; they do not, by themselves, establish what caused a change in impressions.
When direction fits the work
I use this pattern for one-time or infrequent recoveries, investigations, and audits that span tools and need the operator’s judgment along the way. A shared transcript helps me see what was attempted, what succeeded, and what still needs attention.
Frequent, repeatable work deserves a script, scheduled job, or integration with explicit assertions and recovery behavior. The choice turns on the cost of building and maintaining that automation compared with directing the task. A supported API does not make every one-off task worth automating; an available browser assistant does not make repeated manual navigation a sound long-term design.
I keep financial transactions and account-permission changes under direct human control. A person merely ratifying an assistant’s consequential clicks is not the division of responsibility I want for those actions.
From code to operations
Framework update (May 2026): The summary below uses the expanded fourth principle, the self-improving system.
Synthesis engineering is the discipline of how expert humans direct AI agents to produce reliable work. Its early articulation has centered on code: the Foundation-First pattern, the Direction Dynamic, and the five pillars of human architectural authority, systematic quality standards, active system understanding, the self-improving system, and transferable knowledge.
This recovery is a practical case for extending that discipline into operations. The artifact changes from code to configured state, but the operator still decides what matters, what counts as success, and when to retry, escalate, or stop. That shared method makes synthesis engineering a discipline spanning code and operations.
I want to extract the browser-automation patterns for Search Console, Bing Webmaster, and Cloudflare into reusable templates, carrying the navigation steps, expected responses, and stopping conditions into future sessions. The migration checklist belongs before the next cutover, where a baseline can improve the later investigation. Naming the Direction Dynamic in ops gives me a pattern to use again in recoveries and investigations that cross several consoles.
The technical companion explains the site repairs and the build-time check intended to catch a future regression. This account explains how I directed the operational work around them. Together they show how an engineer can combine code changes, browser actions, and judgment in one continuing task and keep responsibility for deciding whether the work has achieved its purpose.
This article is part of the synthesis engineering series. Synthesis engineering and synthesis coding are released to the public domain (CC0). Build on them, adapt them, share them.