Skip to content

Harmix/open-stanley / open-stanley/skills/scout

Scan the user's LinkedIn and X feeds (through their own logged-in browser, or any social MCP they have) for posts worth commenting on, and draft paste-ready comments in their comment voice. Use when the user asks "anything worth commenting on", "find posts to engage with", "who should I reply to", "what's trending in my niche today", or as the scheduled 10am/10pm scan. Also handles "draft a comment for this link".

Scout

The paid product's scan job was the feature the user kept and the one that failed most: wrong posts, no LinkedIn, duplicates, his own posts, comments that read like essays. Each step below removes one of those.

1. Read the feeds as the user

Use the browser capability, logged in as the user. Sources, in order:

  • LinkedIn home feed (scroll ~3 screens), then LinkedIn search for each lane's 2–3 key phrases sorted by recent, then the "Posts" tab of 5–10 accounts listed in strategy.md as people to be near.
  • X: home/For You (~3 screens), a list if the user has one, and search for lane phrases with min_faves:50 where the UI allows it.
  • If a post/social MCP with feed-reading tools exists, use it as a supplement, not a replacement (API feeds lack the ranking the user actually sees).

Capture for each candidate: URL (on LinkedIn, the post's own link from its "Copy link to post" menu item, read with the method in references/linkedin-permalinks.md; never a profile link, a guessed URL or a URN taken from nearby HTML), author handle and display name, follower count if shown, age of the post, reactions/comments, the text. Save the raw pull to inbox/scan-<date>-<am|pm>.md so a re-run never re-scrapes.

If no browser is available, say so and ask the user to paste links; do not fabricate candidates from memory.

2. Filter (hard)

Drop anything that is:

  • by the user (skills/stanley/scripts/stanley-vault is-self <url or handle>), or a repost/reshare of the user's post;
  • already in ledger/surfaced.jsonl (skills/stanley/scripts/stanley-vault surfaced-check <urls...>);
  • older than 24h on LinkedIn or 6h on X (comments on stale posts don't get seen);
  • a product announcement, a news roundup, a job post, or generic AI hype with no argument in it.

3. Score (soft) and pick 2–3 per platform

Score 0–5 on each, keep the top:

  • Lane fit: the post is about one of the lanes, or directly invites the user's thesis.
  • Adjacent-viral: high engagement on how work, teams, meetings, hiring, or org knowledge actually function, where the user's angle is a natural reply even though the post never mentions their field. This is the pattern the user pointed to as the one that was missed ("I need more posts like Matt's"). Weight it as much as lane fit.
  • Reach: author bigger than the user (5–20x on X) and the post is still rising.
  • Openness: the post asks something, admits something, or takes a position someone could add to. Skip announcements.
  • Freshness: under 2h is best on both platforms.

Explain the pick in one line each ("4.9k likes, meetings-are-broken take, your 80%-of-questions-were-answerable data is the reply").

4. Draft the comment

First decide where the user stands on the post, from their own posts, comments and stated views: do they agree, disagree, or have something to add? If they agree, the comment is short agreement plus at most one blunt opinion of their own, in their register (users often write "yep, ..." or one word, not a data point). If they disagree or can push on the claim, say what they would do instead. If they have nothing to add, show the target with no comment rather than inventing a position; the first user ignored three scout drafts in a row that put opinions in his mouth and wrote his own, shorter ones.

Read the comment exemplars in voice.md first, and the user's last approved comments in ledger/comments.jsonl; when four scouts in a row were rejected on voice for the first real user, every rejected comment was two tidy sentences with a built-in contrast or a fact bolted on. One sentence, under about 20 words, plain spoken words, is the shape that survives. It answers the specific claim in that post so it could not sit under any other post; no restating the author, no praise, no aphorism or balanced "X, Y" clauses, no correcting the author's framing with outside facts, no intent or experience the user never stated, and a stored receipt only when it answers the exact point. If there is no specific thing to say to that post, show the target with no comment and say so. A good comment here is 1–3 sentences: one observation from the user's own experience or data (with a receipt in stories/ if it's a number), no agreement opener, no restating the post, no wrap-up. It should read like the person typed it in 30 seconds because they had one specific thought. Two shapes that worked for this user:

  • an observation plus a pointed question back to the author ("how do you know which project context is still relevant? most projects are dynamic, and decisions change or conflict between meetings and emails");
  • a plain first-person data point ("we tested this: the agent answered 80% of the recurring questions; the other 20% were decisions that only existed in the meeting").

Audit each with skills/stanley/scripts/stanley-audit - --platform comment --vault <vault>.

5. Deliver and record

For each pick: the link, the author, one line of why, then the comment on its own lines so it can be copied. Then ask: post which? The user may answer with numbers.

  • Scanning is the highest-volume thing this plugin does, so the pacing discipline belongs here rather than around posting: the 25-posts-per-run cap, one page at a time, and stop-and-report on any captcha or rate-limit notice. ../stanley/references/platform-access.md has the rules and the reasoning.
  • A reply to someone who answered the user's comment is a different register from a first comment: see the Conversations section of the recap skill and the Reply voice block in voice.md; the no-agreement-opener rule does not apply there.
  • Approved → post through the browser (LinkedIn comment box; X reply box, which works for any account, unlike the API) or through a post MCP. Then reload the post and check it per skills/stanley/references/verify-live.md (the text found outside the comment box, under the user's name). Only then say it is posted and run skills/stanley/scripts/stanley-vault comment-add ... --verified "<what you saw>"; the script refuses a posted comment without it. A run once said "posted" because the unsent text was still in the box.
  • Skipped → comment-add --skipped --reason "<their words>"; the reason tunes the next scan.
  • Everything shown → skills/stanley/scripts/stanley-vault surfaced-add <urls> --author <handle> --platform <p>.

Also record what the user posted if they changed the text; the diff is voice data (diff-add --kind comment).

6. Widening when told

When the user says "you missed this one" and gives a link, read it, say in one line what feature made it a good target that your filter missed, add that feature to instructions.md with the date, and re-weight. Don't just apologize.

reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 467kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 319kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 319kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 319kUpdated Claude CodeCodexVerifiedInstall