Skip to main content
Building Silan Viking9 min read

Researchers Should Not Rebuild a Website for Every New Result

Silan Hu#academic-website#geo#research-workflow
Researchers Should Not Rebuild a Website for Every New Result cover

A research update should have identity, evidence, relationships, review state, publication state, and observable delivery signals instead of forcing the researcher to rebuild a website each time.

Body

When a paper is accepted, an experiment reaches a new result, or an archive changes the evidence, the researcher usually wants to do something small: explain the current claim, point readers to the evidence, and make clear which public page is now authoritative. In practice, that small task is often postponed. The project page is stale, the resume is stale, a slide deck contains the new conclusion, and the public website still reflects work from months ago.

The problem is not that researchers cannot build websites. It is now fairly easy to vibe-code another page. The real problem is that research output lacks a low-friction incremental publication mechanism. Every new result forces the researcher to handle titles, summaries, images, language variants, project relationships, publication state, deployment, and feedback again. A small research update becomes a front-end and operations task.

The abstraction I want is a Research Update Lifecycle: a research update is not a loose paragraph of text. It is a content event with stable identity, evidence, relationships, review state, publication state, and observable delivery signals. The website should not merely display these events as static pages. It should become external memory for the research work.

This post uses one concrete scenario: a useful result appears on Friday, while the context is still fresh. How does it become a clear public explanation, and how does the researcher know whether the current version reached readers? Silan Viking models that path as five state transitions: capture -> structure -> propose -> publish -> observe.

Figure 1

The five explicit state transitions from a fresh result to a reviewed public page

1. Capture the smallest durable object

The first step should be smaller than writing an article. The researcher opens the quick editor, types or dictates the one sentence that should not be lost, and saves it while the evidence is still close at hand. What is being saved is not an article yet. It is the smallest claim worth preserving.

Figure 2

Drafting a research update by voice in the current quick editor

For example, in a systems-history or deployment-history project, a new archive source might only support one sentence:

The new archive changes our date for the first deployment from 2019 to 2017.

That sentence can live as a dated moment. It does not need a polished title, search keywords, or a complete argument yet, but it should have stable identity, time, source location, optional evidence links, and current confidence. Otherwise it quickly decays into a chat message, a slide note, or a temporary Markdown file in some folder.

Figure 3

Dated research notes kept on the current Moments timeline

The value is simple: the threshold is low, but the note is not disposable. It has stable identity. Later, even if the title, language, or page path changes, the claim can still be linked to the article, project page, or resume entry that explains it. Stable identity matters more than a beautiful editor because the easiest thing to lose in research maintenance is not text. It is the provenance between pieces of text.

2. Structure the note into a citable page

When the result is ready to share, the researcher creates a separate article or project update and links it to the original moment. The important part is not pretending that everything is "the same document." Moment -> Article -> Project are three kinds of objects, not three views. They should be connected by explicit relationships instead of copy-paste consistency.

One research update page only needs to answer four questions:

  1. What is the new claim?
  2. Which previous judgment does it change?
  3. Where is the evidence, and where are the boundaries?
  4. What should the reader look at next?
Figure 4

Editing the article while its source structure remains visible

The editor presents a readable page while the source keeps the fields needed for titles, summaries, language variants, media, and related-work links. The researcher maintains epistemic state, not a React template. A good tool should let the researcher change the explanation itself instead of making them relearn the front-end directory, build script, and deployment path for every update.

If the result also belongs on a project page or resume, those remain separate items. The same result has different pragmatic roles in an article, a project page, and a resume: the article explains evidence, the project page organizes the line of work, and the resume compresses contribution. The system should preserve relationships, not force one paragraph to serve every audience. Later, a person or an Agent can recover where each claim came from.

3. Give the Agent proposal authority, not publication authority

An Agent is most useful when the task is specific and the output is verifiable. It can find the project and related moments, draft a summary, prepare another language version, add a missing relationship, or report pages that no longer match recent work. These are maintenance actions, not the research judgment itself.

The default Agent workflow should stop at a reviewable proposal. The researcher checks the diff and decides whether to accept it. Production publication and deployment remain owner-run operations. In other words, the Agent can have proposal authority, but it should not have publication authority.

That boundary makes a natural-language request practical:

Update the project page with this result, preserve the uncertainty in the article, and show me the exact changes.

A useful Agent output should not merely say "done." It should list which files changed, which relationships were added, which uncertainties were preserved, which pages need owner confirmation, and which claims are still not safe to publish. The Agent removes repeated maintenance. It does not decide what the researcher is willing to claim in public.

4. Model publication as an explicit state change

Articles, project pages, language variants, media, and search metadata are managed together, but publication must remain an explicit state change. More precisely, a research update should move from private draft to reviewed, then to published, and finally to deployed. Editing, syncing, translation, and Agent proposals are not publish events.

Figure 5

Managing articles and publication state in the current desktop workspace

A private draft does not become public because it was edited, synchronized, or included in an Agent proposal. The owner reviews the page, publishes the item, then deploys it through a configured site. This boundary may look conservative, but it matters for research writing: the researcher must retain final responsibility for public claims.

Visitors receive an ordinary public page that they can read without knowing anything about Silan Viking. A good system should reduce the reader's understanding cost, not require the reader to understand the author's toolchain.

Figure 6

The public blog listing produced from reviewed content

SEO and GEO have a narrow meaning here. They are not promises that a search engine or AI system will cite the work. They mean the published page can carry a stable address, clear title and summary, language information, structured data, and explicit links to related work. Those signals make the page easier for search engines and answer engines to process; they do not guarantee indexing, understanding, citation, or ranking.

5. Treat post-publication feedback as diagnostic signals

Publication is not the end of the lifecycle. For researchers, traditional page views are too blunt: a visit is not a full reading, a crawler request is not understanding, and a citation does not prove that readers saw the evidence. The more useful move is to split post-publication feedback into diagnostic signals.

Silan Viking's dashboard compares the local and deployed content versions and separates human activity from requests attributed to search bots and AI crawlers.

Figure 7

Human traffic, crawler requests, and deployed-version status in the current dashboard

Crawler attribution is a discovery clue, not proof that an AI system understood or cited the work. The useful questions are smaller and more actionable:

  • Delivery: did the current version reach the site?
  • Attention: are people opening the result or only the project index?
  • Machine reachability: did a search or AI crawler request the page?
  • Opening clarity: is the summary clear enough to lead readers to the evidence?

The next action might be a clearer opening, a missing relationship, another language version, or no change at all. The key is that the researcher can now separate failure modes: the content was not deployed, readers did not enter, machines did not request the page, or the page itself failed to lead readers to the evidence.

What is available now

Today, Silan Viking can create and validate plain-text content, connect related records, publish language variants, preview a site with Docker, deploy to a configured target, and report delivery and traffic. Agent proposals are available through MCP, and Git is the current way to move the source across machines.

The screenshots above come from the working desktop interface in the Silan Viking source checkout. They show the current editing experience with existing site content. They are not a staged walkthrough of the archive example and do not represent the installation path for a new workspace. A newly initialized external workspace can use the CLI and ordinary Markdown files immediately; packaged desktop onboarding and more direct cross-device work are still in progress.

That is the boundary I want to keep explicit. The current system supports a reviewable, publishable, observable research-update loop. It does not yet turn all personal website maintenance into full automation. The important part is not the automation story. It is turning a research update from a one-off page project into a repeatable state machine.

Start with the update you are already postponing

Do not organize every old project. Start with one private article for the next result that would otherwise stay in a slide deck. These commands create a local workspace and one valid private research update. They do not publish it for you:

bash
curl -fsSL https://raw.githubusercontent.com/Qingbolan/Silan-Personal-Website/main/engine/install.sh | sh
mkdir my-research-site && cd my-research-site
silan init
silan blog new new-research-result
silan content lint
silan index sync
silan blog show new-research-result

Edit the generated Markdown file in any editor and answer the four questions above. That is the first useful result: one valid, private page whose publication remains under your control.

Once that update has identity, evidence, relationships, state, and observable signals, the website is no longer just a display surface. It starts becoming external memory for the research work.

0 likes
Silan Hu3 readsShare:
G

No comments yet