AI Agents on Bilibili: Video Briefs, Auto-Comments, and Account Boundaries

Imagine two people opening the same video community. One wants an agent to find useful engineering talks and deliver a daily brief. The other wants a way to hide repetitive, apparently automated comments. Both are responding to a change in how AI is used: a model is no longer confined to a private chat; it can read a public community and, if given an account, speak inside it. This guide is for developers testing video agents, teams running public accounts, and readers trying to manage comment noise. It turns the available project documentation into a read-only trial, an interaction approval policy, and a practical stop plan.

The two open-source projects are dsh-bili-scroll and dafeiyu-cleaner. The former searches videos, extracts available material, and drafts a Markdown report, with optional interactions. The latter is a browser extension that locally identifies and hides suspected AI comments. They solve different problems. Neither should be presented as an official Bilibili feature, and a local hide action is not a platform deletion or enforcement action.

Confirm the feature set before giving an account access

The dsh-bili-scroll repository describes four stages: discovery, understanding, reporting, and optional interaction. It can search videos by keyword, read available subtitles and related information, then select useful items for a daily Markdown brief. The repository also describes comment and direct-message settings. Automatic commenting is off by default. When enabled, the stated default is one comment per round, with a configurable maximum of 20, and the same video should be commented on only once. Those limits are properties of the plugin implementation; they do not constitute consent from creators, approval by the platform, or a guarantee that one comment will not be disruptive.

The safest evaluation therefore begins with discovery and reporting only. Run several days without write permission. Measure whether the agent finds the right videos, whether its claims trace to subtitles, and how much correction the report needs. Only then decide whether there is a legitimate reason to draft public comments. An agent's ability to read a video is not evidence that it should publish under your identity.

Figure 1. The project page is a starting point for checking installation and supported functions; verify the current repository and your local settings before use.

Think of the pipeline as four separate gates. The input gate controls which public videos the agent may inspect. The interpretation gate determines what it may infer when subtitles are missing. The output gate controls where reports are saved. The interaction gate controls whether it may affect other people using your account. Most useful trials require only the first three. Keeping the fourth independent prevents an innocent-sounding “full automation” switch from bundling analysis with public speech.

Run a read-only pilot that produces an auditable brief

Read the current installation section in the repository before running commands: dependencies, minimum DeepSeek Harness (DSH) version, account login flow, and plugin source can change. The DSH repository describes the platform as a rapidly changing developer preview. At the time checked, the plugin repository provided dsh plugin --profile web add github:c1overzz/dsh-bili-scroll for a GitHub install. Inspect the package and requested permissions before installation, and do not assume a similarly named registry package is the same build. Record the installed version in the pilot log so a later failure is attributable to a specific version.

Start with one narrow assignment: “Each morning, find at most five videos about AI-agent security and provide the title, link, duration, subtitle availability, two evidence-backed notes, and one question to verify.” Include positive and negative search terms. For example, include agent security and evaluation methods while excluding compilation clips, reuploads, and unrelated product announcements. If a video has no accessible subtitles, the agent may still record its title and public description, but it must not write a content-level summary as if it had watched the complete video.

A useful brief row has six fields: source URL, discovery time, video title, evidence source, claims that can be checked, and reason for inclusion. Have the agent distinguish a quote or timestamped subtitle from an inference. Where evidence is missing, write “not verified.” Store the brief locally or in an access-controlled team workspace. Avoid automatically publishing it or pushing it into a large team chat before someone has checked it.

Figure 2. The settings illustrate why read access and interaction should be evaluated as different permissions.

Use a small daily sample rather than judging the first attractive report. Open several selected videos and ask: Is this actually about the requested topic? Does the summary cite content that exists? Are the links valid? Did the system deduplicate video IDs? Did it treat an older upload as new because the discussion recently became popular? Record each mistake and change one filter at a time. Three days of this review reveal more than a single polished demo because search order, subtitle coverage, and new uploads all vary.

Why a public comment is not just another generated summary

A private brief mainly changes your own decisions. A public comment changes the experience of creators and viewers, speaks under an account identity, and may be read as a human endorsement or criticism. Even a factually correct generated response can sound intrusive when it ignores conversation context. Page content can also contain instructions designed to redirect the agent, links that lure it away from the task, or requests for private information. Once account credentials are present, the risk includes token exposure and unwanted messages as well as content quality.

Use four permission levels. Level 0 is read-only search and private reporting. Level 1 produces comment drafts that a human must review and send. Level 2 permits sending only to an explicit allowlist of accounts, videos, and time windows, with every proposed message approved before dispatch. Level 3 is unattended interaction and should be treated as a specialized exception that requires a clear policy and platform-rule review. Many teams get most of the time savings at Level 1. They avoid handing the model the final “publish” action.

If assisted publishing is necessary, show an approval card with the target video, exact text, evidence for claims, sending account, planned time, and person responsible for withdrawal. Define hard stops: the agent claims details from a video without content evidence; asks viewers for sensitive data; adds an unauthorized marketing link; repeats near-identical text; receives a platform warning; or triggers a credible complaint. On any hard stop, disable sending and preserve the event log. Merely telling the prompt to “be careful next time” is not a control.

Figure 3. The comments and labels illustrate a moderation dispute; the screenshot alone cannot establish who controls an account or prove safe operation at scale.

Prompt instructions against leaking private data are helpful but weak on their own. Keep secrets and private chats out of the comment-generation context. Put account tokens only in the sending component. Run a separate outbound check over the complete comment, and let an approver see precisely what will be posted. A brand account should also ask whether a proposed message is a transparent contribution or an attempt to imitate an ordinary user. Minutes saved on posting are not worth a long-term credibility loss.

What a comment-cleaning extension actually does

The dafeiyu-cleaner repository describes a Chrome, Edge, and Firefox extension that recognizes suspected bot comments. It offers offline rules and optional DeepSeek API assisted judgment. Its central action is to hide matching content in the viewer's browser. It does not erase the server record, issue an official platform penalty, or make the same comment disappear for everyone else.

Local hiding is useful because a viewer can reduce noise without waiting for platform moderation. It is also reversible, which matters when a genuine user writes in a style that resembles an automated response. However, detection rules and models can misclassify people. Optional API judgment introduces a separate decision about transmitting comment text, storing API keys, and paying for calls. Before installation, inspect the extension's site permissions, update history, and network behavior. Begin with reversible hiding and keep a whitelist and restore action available.

Figure 4. Browser-side identification and hiding are personal viewing controls, not a platform-wide enforcement result.

Tone alone is a poor authorship test. A person can write a tidy numbered answer or a short generic reply; an edited automated message can look highly personal. For your own reading experience, a reversible filter can be acceptable even when it is imperfect. For reporting an account, publicly naming a supposed bot, or restricting participation, require stronger evidence: behavior over time, source traces where available, frequency patterns, and human review.

Figure 5. A label can guide a personal filter, but one label does not establish who controls an account.

A seven-day trial for a real team

On day one, define a purpose and an owner. “Give the product team a useful video brief by 9 a.m.” is testable; “let the AI explore the internet” is not. Name the person who can stop the job, decide where logs live, and choose a retention period. On day two, install from a checked source and record version and permissions. Leave comments, DMs, and automatic replies disabled. On day three, create a ten-video human baseline and compare the agent's retrieval and summaries against it.

On day four, edit inclusion and exclusion rules using observed errors. Avoid vague commands such as “be smarter.” On day five, inject failure cases: missing subtitles, a removed video, misleading title, and a comment containing instructions directed at the agent. A safe workflow skips or marks uncertainty instead of inventing content or obeying page instructions. On day six, give the brief to the actual readers. Ask which entries saved time and which required correction. On day seven, choose among continued read-only use, human-approved draft comments, or disabling the workflow. Base the decision on useful-item rate, review minutes, and account risk, not novelty.

For teams using Code0 to manage development work, treat the brief as material for review before it becomes a requirement, test task, or public position. If the team routes models through Apito, request logging and cost controls can help with operational oversight, but a model gateway cannot replace account authorization or a human publish decision. “The model generated a valid response” and “the organization should publish it” are different judgments.

Budget, privacy, and maintenance questions to answer

The first budget is for model and interface usage. Search, subtitle processing, summarization, and comment generation may touch different services. Optional API-assisted comment classification adds cost that grows with the volume of comments inspected. Track videos processed per day, average subtitle length, brief-generation frequency, and classified comments. Calculate from actual billing data rather than a decorative sample price.

The second budget is for identity and data exposure. Cookies, login states, messages, and browsing preferences may be sensitive. Keep credentials in a managed secret store, not in a prompt, report, or screenshot. Retain logs for the shortest useful period, separate team roles, and document whose browsing history a shared-device extension can access. If a trial uses a personal account, be especially cautious about giving software permissions that exceed the task.

The third budget is maintenance. Site structure, subtitle availability, plugin dependencies, browser permissions, and platform controls can change. Perform a weekly health check: did the plugin update, are failures concentrated on one endpoint, are brief links dead, is the extension hiding genuine discussion? Write the rollback sequence before a problem: stop the scheduled task, revoke login, disable the extension, restore hidden material, and inspect recently sent comments.

Choose the workflow that matches the problem

If the goal is faster discovery, use read-only briefs. Their benefit is easy to measure and their side effects are limited. If the goal is maintaining a public account, begin with comment drafts and human approval; any later automation should be narrow and traceable. If the goal is reducing noise while reading, use reversible browser filtering and do not convert its label into a public accusation. These workflows can coexist, but they need separate permissions, logs, and success measures.

Before a pilot goes live, check seven things: repositories and versions are current; read and write permissions are separate; comments and DMs are off by default; reports preserve evidence; false positives can be restored; credentials and API keys can be revoked; and a named person can stop and investigate a complaint. Then the agent is a bounded tool in a community, rather than an unowned automatic speaker.

Put the boundary into a Code0 workflow

Represent video candidates, evidence review, and publishing as separate tasks. A candidate task stores only public links and subtitle excerpts. A review task must record uncertainty. Only a named human can close the publishing task. An agent-generated draft therefore cannot become a public comment just because an earlier task finished. Review false selections and rejected drafts weekly, then update narrow rules rather than copying raw comment threads into the system prompt.

Sources

• dsh-bili-scroll repository and current instructions
• dafeiyu-cleaner repository and extension documentation

FAQ

Will installing the video briefing plugin make it post comments automatically?

The repository describes automatic commenting as optional and off by default. Still verify the installed version, local settings, and account grants. Keep comments, messages, and replies disabled during a read-only trial.

Can an agent summarize a video when subtitles are unavailable?

It can record public metadata such as title and description, but should mark content understanding as unavailable. It must not invent a detailed account of a video it cannot inspect.

Does an “AI comment” label prove the account is a bot?

No. Rules and models produce false positives. The label is appropriate as a reversible personal filter, not as sole evidence for a report, public accusation, or ban.

Can a prompt be the only safeguard for a company account?

No. Separate target allowlists, sending permissions, outbound checks, human approval, logs, and a stop switch from the language model. A prompt cannot substitute for those controls.

When is optional API-based comment classification worth using?

Consider it only after offline rules have a measured error problem and the organization accepts the data transfer and cost. Test a small sample, preserve restore controls, and review how the key is stored.