Virality monitoring skill spec: finding conversations worth joining

This is the specification for the listening skill: watching public conversations for the moments where a reply would genuinely help somebody, and where the product is genuinely part of the help. It ends at a queue item and a drafted reply. It does not end at a posted comment, and the reason is in the escalation section.

The word virality is doing specific work here. The skill is not counting mentions of your brand, which is a reporting job. It is looking for conversations that are moving, where a question is live and unanswered, and where an answer arrives while people are still reading.

Only public posts, only through access the platform actually grants. That is a boundary rather than a preference: terms of service, rate limits and the difference between an official interface and scraping around one are part of the specification, not an operational detail to sort out later.

Two constraints apply to every source. Anything that requires evading a rate limit, an authentication wall or a robots directive is out of scope. Anything a platform prohibits does not become acceptable because an automated system is doing it.

Five families of signal, evaluated separately. Collapsing them into one relevance number is how a system ends up replying enthusiastically to a two-year-old thread that was resolved on the first comment.

Intent

Is there a job to be done in the post? The strong forms are an explicit request for a recommendation, a described problem with no solution yet, frustration with a current approach, and a stated plan to change something. Everything else is discussion.

Fit

Would this person plausibly be helped by what you have? Role, scale, stack and vocabulary all carry it, and a thread full of practitioners describing a problem you do not solve is a bad fit however relevant the keywords look.

Timing

Is the conversation live? Recency, whether replies are still arriving, whether the question is answered, and whether the thread is locked or archived. A perfect answer to a settled question is noise.

Reach

How many people will read this later? Threads that rank in search keep receiving readers for years, which can matter more than the audience present today.

Risk

What could a reply cost? Community rules against promotion, a thread about a competitor, a complaint about your own product, a legal or medical context, personal distress, or a pile-on in progress. Risk does not lower a score, it vetoes.

Keyword matching is the retrieval layer, not the decision. It casts a wide net over product names, category terms, problem descriptions and the phrasings people use when they do not know the category term. Then everything it catches has to be read.

These are the classes a keyword-only system gets wrong, and each needs its own test:

• Negation, where the post says it is not looking for a tool or has already chosen one
• Past tense, where the problem was solved last month and the thread is a retrospective
• Sarcasm and complaint, where the category term appears inside criticism rather than a request
• Homonyms, where a product name is also an ordinary word, a place or somebody else's brand
• Quotation, where the post is describing what someone else asked rather than asking
• Answered questions, where a reply already solved it and another voice adds nothing
• Language and register, where a machine translation reads as a stranger arriving in a local community

The decision the skill makes is a classification with a stated reason, not a similarity score. Every queue item records the sentence that established intent, quoted from the post. If no such sentence can be quoted, there is no intent and the item is dropped.

Scoring exists to order a queue that is longer than the time available, not to decide whether to reply. That decision belongs to the veto list, and the two mechanisms stay apart.

Any weights you start with are guesses. They are worth writing down so they can be argued with and adjusted against what actually happens, and they are not evidence. Treat a scoring configuration as an opinion with a version number.

These are the vetoes:

• The community prohibits promotional replies, or prohibits links, and the reply would breach that
• The thread is a complaint about your product, which belongs to support and not to outreach
• The thread names a competitor as its subject, where a reply from you is an advert wearing a helpful voice
• The context is legal, medical, financial or safety-critical, or the author is in evident distress
• The thread is a pile-on, a brigade or an active controversy
• The account has already replied in this thread, or in another thread by the same author recently
• The post is older than the freshness window agreed for that source

The same question travels. One person asks in a forum, links it on social, and someone else quotes it a day later. Three queue items for one conversation produce three replies that read as a campaign.

• Group by conversation rather than by post, matching on the author, the link and the substance of the question
• Hold one reply per conversation, and let a human decide if a second is warranted
• Cap replies per community per period, so an account that shows up in every thread does not become the thing people notice
• Keep a per-author cooldown, because following one person across communities reads as surveillance regardless of intent
• Expire items that were not acted on, since a stale reply is worse than silence

The test is simple and hard to pass.

If not, it is an advert, and the community will read it as one.

Answer first

The reply opens by addressing the actual question, in enough detail to be usable on its own. A reply whose value is behind a link has given nothing.

Disclose the connection

Say who you are and that you work on the thing you are mentioning, in the reply itself rather than in a profile. Beyond being the rule in many communities, undisclosed endorsement by someone with a material connection to the product is a regulatory problem in several jurisdictions.

One link at most

And only when the link is the answer. A reply carrying two links is being written for the poster of the reply.

Match the register

Communities have a voice. Marketing phrasing in a technical thread is detected instantly and read as an intrusion, whatever the content underneath.

Concede the limits

Say plainly where your product does not fit and what you would use instead. It reads as honesty because it is honest, and it costs a sale you were not going to make.

Leave without a hook

No closing pitch, no offer, no request to book anything. The reply ends when the answer ends.

Sometimes, do not mention it

A helpful answer with no product in it is a legitimate output. A skill that cannot produce one will eventually force the product into a thread where it does not belong.

The skill drafts. A person posts. That division is the point at which the whole design stays defensible, and it should hold until the queue has a long record of a human agreeing with the draft.

A listening system reads people who did not ask to be read by a company. The constraints follow from that, and they are not negotiable against volume.

• Public posts only, never private messages, never a private group joined under a pretext
• One identity, disclosed, rather than several accounts creating the appearance of independent voices
• No profiles of individuals beyond what the current conversation requires, and no linking of accounts across platforms to build one
• Deleted and edited posts are honoured, so a queue item whose source has gone is dropped rather than acted on from a cached copy
• Retention has a limit and a reason, because a permanent archive of what strangers said about your category is a liability nobody asked for

Each queue item carries the same fields, and a missing field blocks the item.

A listening system that cannot see what happened to its replies is running blind, and it will keep making the same mistake in a new community every week. The feedback that matters is not engagement. It is whether the reply survived, whether a moderator removed it, whether the author responded and how, and whether anyone else in the thread objected to its presence.

These are the failures to watch for, in roughly the order they appear:

• Volume creep, where the queue grows and the bar quietly drops to fill the time available
• Keyword drift, where a broad term starts pulling in a neighbouring topic nobody noticed
• Voice collapse, where every reply reads as the same paragraph with the nouns swapped
• Rule rot, where a community changed its policy and the stored copy did not
• Removal blindness, where nobody checks which replies were taken down and the pattern in them

Review on a fixed schedule, reading the actual replies rather than the counts. The question at the review is not how many conversations were joined. It is whether a regular in each of those communities would still be glad the account is there.