recap skill
Report what the user's posts and comments did (impressions, reactions, comments, followers, who engaged), attach the numbers to the stories and lanes that produced them, and give one concrete recommendation. Use for "how did my post do", "daily recap", "weekly numbers", "who's engaging with me", and as the nightly scheduled task. Reads LinkedIn analytics and X analytics through the user's browser.
- GitHub stars
- 6
- Files
- 1
- Last commit
- Nov 10, 2026
- License
- MIT
- Allowed tools
- —
What does the recap skill do?
recap is an agent skill. Report what the user's posts and comments did (impressions, reactions, comments, followers, who engaged), attach the numbers to the stories and lanes that produced them, and give one concrete recommendation. Use for "how did my post do", "daily recap", "weekly numbers", "who's engaging with me", and as the nightly scheduled task. Reads LinkedIn analytics and X analytics through the user's browser.
How to install the recap skill
Each command copies the skill folder into the client's skills directory. Restart the client or start a new session to load it.Install the recap skill in Plugin
/plugin marketplace add Harmix/open-stanley /plugin install open-stanley@open-stanley
Install the recap skill in Claude Code
~/.claude/skills/recap/SKILL.md
$ npx degit Harmix/open-stanley/open-stanley/skills/recap ~/.claude/skills/recapInstall the recap skill in Claude Code (project)
.claude/skills/recap/SKILL.md
$ npx degit Harmix/open-stanley/open-stanley/skills/recap .claude/skills/recapInstall the recap skill in Codex
~/.agents/skills/recap/SKILL.md
$ npx degit Harmix/open-stanley/open-stanley/skills/recap ~/.agents/skills/recapInstall the recap skill in Gemini CLI
~/.gemini/skills/recap/SKILL.md
$ npx degit Harmix/open-stanley/open-stanley/skills/recap ~/.gemini/skills/recapInstall the recap skill in Claude.ai
npx degit Harmix/open-stanley/open-stanley/skills/recap recap && zip -r recap.zip recap # claude.ai → Customize → Skills → + → Create skill → Upload a skill → recap.zip
SKILL.md
View on GitHubRecap
The paid product's recap was "Today, 131 saw my content and 2 reacted." Numbers without a decision are noise. A recap here has three parts: what changed, who showed up, one thing to do.
1. Reconcile the ledger with the platform, then pull
The ledger only knows what went through publish. People post by hand, from their phone, from another tool; the first three weeks of real use had a post missing from the ledger the night after it went up, and the cadence check then undercounted. Start with stanley-vault drafts-due: scheduled drafts whose time has passed and that have no ledger row. For each, find the live post on the activity page by its first line and add it with the ready-made post-add --from-meta … --via reconciled command the tool prints, so the lane, shape and experiment tags come from the draft instead of being guessed. Then the general sweep: the first browser action of every recap is the user's activity page (/in/<handle>/recent-activity/all/; X: x.com/<handle>): every own post newer than the ledger's newest that is not in ledger/posts.jsonl gets post-add --via reconciled with its text, date (from the activity id: epoch_ms(id >> 22)) and --image if it carries one; every repost gets post-add --kind repost so presence is counted without polluting the shape statistics (stanley-stats ignores reposts). Say in the report how many posts were reconciled; more than zero means the user is publishing around the plugin, which is fine, but the ledger has to know.
1a. Pull
Through the browser: LinkedIn post analytics for every post in ledger/posts.jsonl younger than 7 days (impressions, unique members, reactions, comments, reposts, saves if shown), profile follower count, and the notifications page for new comments/replies. X: post analytics (views, likes, replies, bookmarks, profile visits) and follower count. If analytics are unavailable in a remote run, report what the ledger has and say so.
Write the pulled metrics into the ledger: skills/stanley/scripts/stanley-vault metrics-add --url <post url> --bucket 1h|24h|7d --impressions N --reactions N --comments N --reposts N [--saves N]. Then run skills/stanley/scripts/stanley-stats --vault <vault> (add --platform): it prints the median, lifts by lane / shape / hook feature, the top and bottom posts, and the status of every experiment declared in strategy.md. Read its warnings literally: under 8 posts per arm it says "directional only", and so should you.
2. What changed (three lines max)
Only deltas: which post moved, versus the median in strategy.md; new comments that need a reply (with the commenter's name and the comment's first line); follower delta over 24h and 7d.
2b. Conversations you are in
Two places, both every run. (a) Comments on the user's own posts: the notifications page plus each post younger than 7 days. (b) Replies to comments the user left on other people's posts: ledger/comments.jsonl for the last 7 days, and LinkedIn /in/<handle>/recent-activity/comments/ (X: x.com/<handle>/with_replies) for anything the ledger missed; open each thread and look for replies under the user's comment. The first recap only checked (a) and left a reply from a founder unanswered for a day; the user had to ask.
Triage every incoming comment and reply into three tiers, and report them that way. Most deserve a like, not a reply; a recap that drafts a reply for every comment buries the two that matter, and the first user had to say so.
- Reply: a real question, a disagreement with substance, someone building something next to the user's work, a larger account asking something.
- Like only: agreement ("so true", "I love this"), thanks, a story that needs no answer, one-liners. When unsure, like only.
- Ignore: spam, bots, pitches.
Before listing anything as unanswered, reload the thread right before reporting and look at its structure, not its text: a reply by the user sits nested under the comment with their name and an "Author" label. People answer by hand between runs; a text-only read listed a comment as unanswered that the user had answered hours earlier, and drafted a reply nobody needed. Check the user's own reaction the same way: a filled thumb with a count means they already liked it, and clicking it again removes the like.
Likes on the like-only tier follow likes: in strategy.md. With likes: auto the recap likes them itself: click by coordinates from a screenshot (ref clicks on LinkedIn's reaction button sometimes do not register), confirm the filled thumb with a count after, and if the reaction picker opens, move away without picking. With likes: ask (the default until the user says otherwise) list them and like after a yes, and the first time ask once whether to do it automatically from now on. A like is the only thing the recap does without a yes; never text.
For the reply tier, draft the response. The register is a conversation, not a post: it is fine to open with "Yeah" or "Thanks"; no benchmark numbers, receipts or story lessons unless the person asked for them; no closing question by default. When the person is building something adjacent to the user's work, offer the next step (a call or DM) in plain words. Never put a commitment in a reply that the user has not made ("adding a kill date to every bet"): only facts from the ledger, the stories and the user's own words. Audit with stanley-audit --platform comment. In the report, show each reply-tier draft under the commenter's full text (and, for a reply under the user's own comment, the user's original comment above it), because a user cannot approve a summary; keep the main report short and put these below it. Then wait for a yes; post through the browser, verify per skills/stanley/references/verify-live.md, and comment-add --verified with --reply-to <url>. Record the draft→final diff when the user rewrites it.
3. Who showed up
People who engaged twice or more in the window, with what they do (from their headline) and which post they reacted to. This feeds the monthly Orbit ritual. Never suggest engaging with them unless there's an actual thread to reply to.
3b. Cadence before shape
Before any recommendation about shapes or lanes, compare own posts (kind post, after the reconcile step above) published in the last 14 days with the cadence in strategy.md; report reposts separately. If the ledger shows fewer than half the planned posts, the recommendation is about publishing — which draft in drafts/ is closest to ready, what is blocking it — not about which shape works. Three weeks of runs that produced drafts and two published posts is a publishing problem, and no amount of shape analysis fixes it.
When reading a post page, record the reaction breakdown (hover the reaction count: like, celebrate, support, love, insight, funny) with metrics-add --insight N --like N .... Insight share is the best predictor of reposts in the 226K study and a like share above ~85% means people scrolled past; stanley-stats reports both once a few posts carry them.
4. The one recommendation
Pick one, with evidence from stanley-stats, never from memory. A lift on fewer than 3 posts (the table marks them n<3: not a signal) is not evidence for or against a shape; say "one post, no read" rather than calling a shape the user's worst. Examples: reply to X's comment now; the number-in-hook posts are running 3x, do another; you haven't posted in 9 days and the streak matters less than the compounding; the X account has 5 followers, so comments on larger accounts matter more than original posts there this month. When two posts in the same lane are underperforming, say the lane may be wrong, not the timing.
4b. The share card (opt-in, weekly)
If strategy.md has share_card: yes, the Sunday recap renders one card from the ledger with stanley-chart --kind bars: the week's posts ranked by impressions, the title stating the week's number ("11,204 people saw my posts this week, 3 posts"), and the source line "made with Open Stanley · github.com/Harmix/open-stanley". Hand it to the user with a two-line draft in their voice; they post it or not. Never without the opt-in, never with a number not read from the platform this week, and the card is cleaned (image-clean) before upload. This is the loop the paid tools do not have: a user's own numbers, shareable, with the tool named once.
4c. Milestones
strategy.md may list milestones: 100000, 250000 (90-day impressions) and milestone_lane: <lane>. The top-posts page does not show the 90-day total, so every recap also opens https://www.linkedin.com/analytics/creator/content/?metricType=IMPRESSIONS&timeRange=past_90_days, reads the cumulative "Impressions" number and the "vs. prior 90 days" percentage, and runs stanley-vault milestone-check --impressions-90d <number>. Decide from the script, not from notes in strategy.md: in real use a note reading "100K on 2026-11-07, one post for 100K" made a recap report "250K is not close" and skip the 100K post.
When it prints CROSSED (exit 2), say so first, screenshot the analytics card (the receipt), and draft the milestone post through write: what the number is, what the plugin did and did not do, the one lesson, the link in the first comment, and the screenshot cleaned (image-clean) as the image. Save it in drafts/ with its .meta.json and hand it over in this run. Then run stanley-vault milestone-check --mark-done <M> so it fires once. Never mark a milestone done without a draft saved in drafts/.
5. Weekly and monthly
Weekly (Sunday): same shape over 7 days, plus the stanley-stats experiment table. When an experiment reaches min_n per arm, state the result in one line and move it to "Resolved bets" in strategy.md with the numbers and date; propose the next experiment (one at a time, alternating variants post by post). This is the whole experimentation method: declare, alternate, wait for n, resolve, write it down. Monthly: rewrite the "what good looks like" section of strategy.md with the new median and top three, and hand the orbit list to rituals.
Keep the whole message under 12 lines. If nothing was posted and nothing moved, say that in one line and give the one recommendation anyway.
Files in this skill
1recap skill questions
How do I install the recap skill in Claude Code?
Run npx degit Harmix/open-stanley/open-stanley/skills/recap ~/.claude/skills/recap. Claude Code loads skills from ~/.claude/skills (all projects) or .claude/skills (one project) and uses a skill when its description matches the task.
Which clients can use the recap skill?
Any client that reads the Agent Skills format: Claude Code, Codex, Gemini CLI and claude.ai (upload as a ZIP). Scripts in the skill run only where the client allows code execution.