Social listening

State: to-be-approved. Deploy-verified against a live workspace: yes, on 2026-10-04: a limit: 20 search landed 21 posts from the past week, the listener surfaced 3 of them in one Slack digest, read no comments because none of the picked posts had any, called no engagement action, and wrote 4 ledger rows (3 posts and the week's digest row); a re-run the same day found the digest row and posted nothing. Treat Done when below as the acceptance test and review cargo-ai cdk plan before deploying. Make no outcome claim for this skill until it is approved.

The outcome

The conversations your buyers have in public stop going past unseen. Every Monday:

1. A model buys one capped search. linkedin_posts syncs LinkedIn's post search for the past week, on keywords taken from your workspace context: the pains in the buyer's own words, the category term, competitor names. The query is the spend, so it lives in the model's config.
2. An agent judges. It reads the ICP and pains from the context, drops every post already surfaced, and picks at most five worth joining, each with the context line that decided it.
3. It reads comments only on those five, and names at most ten commenters whose headline matches an ICP persona, quoting the first sentence of what they said.
4. One digest lands in a locked Slack channel, and the picked posts go into the surfaced_posts ledger, so a post never headlines twice.

It never likes, comments, connects or messages. Those actions sit on the same LinkedIn connector and none is wired; the contract fails if one is. A human reads the digest and decides whether to join. The ledger keeps posts, not people: an engager appears in that week's digest, quoted from a public comment, and is not stored as a list.

Example

Every Monday, tell #gtm-signals which LinkedIn conversations about lead routing are worth joining.

Illustrative output, fictional records:

:ear: *What the market said this week*
_Routing breakage after CRM migrations came up three times._

*Worth joining*
• Dana Ruiz, VP RevOps at Fabrikam: "Our Salesforce migration broke every
  routing rule we had" — ICP persona, pain "routing breaks on field changes"
  (linkedin.com/posts/dana-ruiz_fab…) · 41 comments
• Wingtip Toys Engineering: a teardown of hand-built lead routing — category
  term "speed to lead" (linkedin.com/posts/wingtip…) · 12 comments

*Warm engagers*
• Sam Okafor, Head of Revenue Operations at Contoso, on Dana's post:
  "We rebuilt ours three times this year." (linkedin.com/in/sam-okafor)

_Searched: "lead routing" OR "speed to lead" · 40 posts read, 3 already surfaced_

Forty posts were read; two were picked and recorded in surfaced_posts, one commenter matched an ICP persona, and nothing was sent to anyone.

Put it in your project

This folder is a worked example: real CDK resources written for some other company. The job is to end up with the code your company would have written, in your project, and an agent does the adapting. If the cargo-cdk skill is in your session it carries the long form of this; if not, this is enough.

1. Install it — the CLI does the copy. From inside the CDK project, cargo-ai cdk add cookbook/social-listening writes this example to infra/social-listening/ and this procedure to .claude/skills/social-listening/. No project yet? cargo-ai cdk init <dir> --cookbook social-listening && cd <dir> && npm install does both. If you are reading this from the project's .claude/skills/, the install already happened — start at step 2.
2. Reconcile it with what is already declared. If the project already has a LinkedIn, Slack or Anthropic connector, rewire the imports to the existing one and drop the copy; two resources with one slug is a collision at deploy.
3. Write the keywords from the context. Replace searchKeywords in infra/models/linkedin-posts.ts with phrases a buyer would post, read from the ICP and pains in the workspace context, plus competitor names. Run the search once by hand (cargo-ai orchestration action execute on linkedin.searchPosts with the same keywords and datePosted: "Past week") and read ten results out loud before you keep them.
4. Point Slack at your channel. Set channelId in infra/agents/listener.ts to a channel id (C…) from the Slack connector's autocomplete, and invite the bot.
5. Adapt. Work the sections below in order: What should not change is what you argue back about; What you can change is what you offer unprompted; What you will be asked is the floor, and you derive before you ask. Record what you changed and why under a ## Decisions section in your copy of this file.
6. Check, then plan. node --import tsx evals/contract.mjs from this skill's folder, then npm run check && cargo-ai cdk plan from the project root. Show the diff, and deploy only on an explicit yes: cargo-ai cdk deploy.
7. Verify. Trigger the model sync once, then send the agent its trigger text by hand (cargo-ai ai message create), and walk Done when with evidence. Send it again: the second run must post nothing.

What you will be asked

Derive before you ask. An input with a lookup is looked up, not asked.

What you can change

What should not change

• No engagement action on uses. (infra/agents/listener.ts) Liking, commenting, connecting or messaging from an agent is automated outreach to people who never asked. The contract fails on any of them.
• The ledger stores posts, never people. (infra/models/surfaced-posts.ts) A standing list of everyone who commented on a topic is a scraped list. The digest names engagers that week, quoted from a public comment, for a human to read.
• The search is the spend, so it stays in the model's config and capped. (infra/models/linkedin-posts.ts) Removing limit or widening datePosted re-buys last week's posts every Monday; the sync replaces the rows, so nothing is gained.
• The ledger is checked before comments are read, and written after the post. (infra/agents/listener.prompt.ts §2, §5, §6) Check late and every re-run re-pays for comments. Write early and a failed post marks posts as surfaced that nobody saw.
• Comments are read only on picked posts, under the cap. (infra/agents/listener.prompt.ts §4) The comment read bills per comment returned; reading every post's comments is the bill that grows without anyone choosing it.
• No invented employer. (infra/agents/listener.prompt.ts §4) An engager's company is what their headline states, or nothing. Looking people up elsewhere turns a digest into a dossier.
• channelId is locked. (infra/agents/listener.ts) The digest quotes people by name.

Done when

• node --import tsx evals/contract.mjs passes
• cargo-ai cdk plan reports the agent, the two models, the three connectors and the two folders
• the first sync landed rows in linkedin_posts, all from the past week, no more than limit
• one digest landed in the locked channel in the references/digest.md shape, every pick citing a context line and linking its post
• surfaced_posts has one row per pick plus one digest-<date> row, and no column holding a person
• a second run the same day posted nothing and read no comments
• the run transcript shows no LinkedIn action other than searchPostComments, and only on picks
• every quoted comment exists on the linked post, and every employer shown is in the commenter's own headline

What it costs

Read the live price of each paid step immediately before the plan, and say each one out loud:

• cargo-ai connection integration get linkedin: the fetchPosts extractor (per post returned, so the weekly ceiling is limit times that price) and searchPostComments (per comment returned).
• cargo-ai orchestration action list postMessage --kind connector --integration-slug slack.

The weekly cost is one search capped by limit, comment reads on at most five posts each capped by the prompt's comment limit, one Slack post, and the agent run billed as LLM tokens through the Anthropic connector. On the live run the agent run, not the search, was most of the bill: it reads the ICP, personas, objections and insights before it judges a post. A re-run that stops at the ledger costs a small fraction of a full run. Read the run's usage after the first week (cargo-ai ai message get) before you widen limit.

Composes into

account-scoring (the with-tiers variation puts engagers at A and B accounts first), web-capture (what the market says next to what competitors publish), and find-linkedin-url / enrich-linkedin-profile when a human decides one engager is worth a closer look.