git:20260903.e492851 to git:20260905.4909a33

94 added, 110 removed. Audit A to A.

---

name: trend-bluesky

title: "Bluesky Trend Harvester Playbook"

- description: "Platform-specific intel for harvesting Bluesky trends — AT-Proto repost-driven amplification, custom-feed targeting, harvest URLs, scoring on repost-to-like ratio (no view counts), hook taxonomy tuned to Bluesky's technical / anti-marketing audience, and dead patterns in 2026. Activates in browser sessions whenever the user names Bluesky."

+ description: "Harvest Bluesky conversations, formats, and launch evidence into a sourced brief. Activate when scanning Bluesky trends or researching post angles, custom feeds, and relevant community responses; record observed surfaces and account baselines without inventing amplification paths."

license: Apache-2.0

- compatibility: "Octoweb browser access. Signed-in Bluesky session needed for Discover and Following surfaces; logged-out works for search and tag pages."

+ compatibility: "Octoweb browser access and network access. Use the available Bluesky session; record signed-in state and unavailable surfaces."

capabilities: octoweb memory-read memory-write

domains: browser

rules:

- session(trend) content(bluesky)

- match(\bbluesky\s+(trend|trends|harvest|brief|post)\b)

- match(\b(harvest|scan|analyze)\s+bluesky\b)

- match(\bbsky\.app\b)

---

## Overview

- This skill carries the platform-specific mechanics the trend-harvesting agent needs to harvest Bluesky — current AT-Proto ranking signals, harvest URLs including custom feeds, scoring on repost paths (no view counts), hook taxonomy for Bluesky's technical / anti-marketing audience, dead patterns, timing. The agent owns the shared DNA loop; this skill plugs the Bluesky parameters in.

+ Return an auditable brief about the requested Bluesky community. Separate what you observed from an explanation of why it happened. Harvest evidence and reusable editorial observations; writing and publishing are downstream steps.

## Mental model

- Bluesky is repost-graph + custom-feed driven, not algorithmic-feed driven. There is no global For You by default — Discover is opt-in and most niche users live on community-curated custom feeds. Reach is built by reposts (Bluesky's term for shares) and by getting picked up by curator-maintained feeds. View counts aren't surfaced. Anchor accounts and curators drive >50% of niche post amplification — one repost from an anchor can dwarf algorithmic reach. The audience skews more technical, more left-leaning, more dev/foss-heavy, and more skeptical of marketing than X. Marketing voice gets ignored or muted fast.

+ Discover is a default algorithmic feed with broad reach. It respects an account's recommendation opt-out for strangers; topic- and list-based feeds needn't honor that declaration (official, Visibility declaration, 2026-09). Don't describe Discover as opt-in or assign it unverified reply or graph weights.

- ## Rules

+ Following is reverse chronological (measured, Buffer timing, n=over three million posts, 2026-08). Timing can affect what an online audience encounters; chronology doesn't guarantee a reply's delivery or make old posts equally visible.

- ### Current ranking and amplification signals (2026)

+ Followers can hide a followed account's reposts in feeds (official, Repost controls, 2026-08). Don't equate repost count with audience delivery, or presume quotes outperform plain reposts. Use the sampled community, not an assumed technical or political platform identity, as your scope (directional).

- | Signal | Effect |

- |---|---|

- | Reposts | Primary amplification mechanism (no algorithmic feed by default) |

- | Repost-to-like ratio | >0.10 strong distribution, >0.20 breakout |

- | Quote-post with commentary | Highest-quality amplification — preserves context + adds new audience |

- | Cross-feed surfacing | Posts in 3+ custom feeds simultaneously = real federation amplification |

- | Hashtag follows | 1–2 relevant hashtags help discovery; 3+ reads spam |

- | External link in root | NOT suppressed — Bluesky treats links normally |

- | Anchor-account repost | Single repost from a well-followed niche account = significant lift |

- | View counts | Not surfaced — do not score on them |

+ ## Harvest rules

- ### Harvest surfaces (run in parallel)

+ ### Establish the collection scope

- | Surface | URL | Yields |

+ Record the user's topic, intended audience, language, geography, and time range. Record collection time and timezone, signed-in state, query filters, and stopping criterion. Choose a bounded sample that includes ordinary posts and relevant smaller accounts as well as popular examples (directional).

+ Do not publish, follow accounts, join chats, or contact curators as part of harvesting. Preserve private chat boundaries; public trend collection doesn't authorize use of private messages.

+ ### Resolve surfaces through the current app

+ Start at [Bluesky](https://bsky.app/) and navigate through the visible UI. Copy the actual share URL after selecting a feed or search filter. The verified source set doesn't establish canonical Discover or hashtag routes; don't substitute a guessed path or generator identifier.

+ | Surface to inspect | Navigation and record (directional) | Sampling limit |

|---|---|---|

- | Discover (signed-in) | `https://bsky.app/feed/discover` | Algorithm-mixed surface |

- | Following | `https://bsky.app/` | User's peers (only if follow-graph exists) |

- | Topic search | `https://bsky.app/search?q=<topic>` | Last-day live posts |

- | Hashtag pages | `https://bsky.app/hashtag/<tag>` | Topic communities (Bluesky uses real hashtags) |

- | Custom feeds (niche) | `https://bsky.app/profile/<curator>/feed/<feed>` | Curator-maintained niche surfaces |

- | Anchor profiles | `https://bsky.app/profile/<handle>` | Last 7 days per anchor |

+ | Trending topics or Explore, if present | Check current search/home navigation; open a relevant topic and retain its resulting URL | Record region, language, and absence if unavailable; don't assume a default feed roster. |

+ | Discover | Select the named feed in the current session and copy its URL | Treat the sample as session-conditioned, not a platform census. |

+ | Following | Select Following explicitly | Record whether the follow graph provides relevant material; home alone doesn't identify the selected feed. |

+ | Topic and hashtag search | Enter the query, including the hash for a tag; use available filters and share the result URL | Record sort choice exactly as displayed; don't call an unfiltered result a recent sample. |

+ | Custom feeds | Inspect current feed discovery navigation, then the feed's description and actual results | Record feed URL, owner, stated criteria, and whether each cited post was observed there. |

+ | Relevant profiles and packs | Open from observed posts or feed descriptions; copy profile and pack URLs | Record topical relevance and observed repost behavior; don't assert a fixed pack-size range. |

- For the user's AI / agents / LLM / RAG / startup niche, identify and harvest 3–5 niche custom feeds at runtime via search before relying on them. Custom-feed names rotate quarterly.

+ Search supports shareable filters for keywords, people, date range, and language (official, Search rollout, 2026-07). The app completed its search migration in the subsequent release (official, Search migration, 2026-07). These announcements don't establish an API endpoint or sort-mode contract: adapt browser collection to the current UI.

- Run 5–7 surfaces in parallel. Scroll 2–3× per tab before snapshotting.

+ Record missing or inaccessible surfaces as unavailable. Don't invent logged-out access guarantees, quarterly feed renames, or results from a guessed public API route. Inspect enough of each surface to satisfy the declared sampling scope; use independent browser tabs only when the browser supports them safely (directional).

- ### Scoring rubric (Bluesky-specific signals)

+ ### Capture each candidate

- Virality axis 0–5:

- - Repost-to-like ratio — primary federation signal

- - Absolute repost count (>20 reposts in 48h on a <5k account = climbing)

- - Quote-post density — highest-quality amplification

- - Reply-thread depth — top-level replies that themselves get replies

- - Cross-feed surface — same post in 3+ custom feeds = federation-viral

+ Save the canonical post URL and full displayed handle, source surface, collection time, publication time, and post age. Record visible likes, replies, reposts, and quotes separately, plus follower count if available. Mark unavailable metrics unavailable; don't manufacture impressions or infer that a missing view count is universally unavailable (directional).

- Niche-fit axis 0–5 — universal scale.

+ Capture the actual opening text, format, media's evidentiary role, link destination, affiliation disclosure, and relevant reply context. Record visible restrictions or labels as observations; don't reverse-engineer a suppression rule from them (directional).

- Do not score on view counts — Bluesky doesn't surface them.

+ For a thread, inspect the root and relevant follow-ups. Note whether installments carry enough context independently; automatic numbering exists, and its feature gate was removed in the later release (official, Thread numbering, 2026-08; App releases, 2026-09). Don't count displayed numbering as author-written hook text without checking.

- ### Hook taxonomy currently winning

+ If format constraints matter to a recommendation, use the verified limits: 300 grapheme clusters for post text (official, Post lexicon, 2026-09); up to 10 photos, with a carousel from 5 (official, Photos release, 2026-06); video up to 10 minutes and 300 MB (official, Video release, 2026-08). Don't score a longer asset as better or assert a universal embed-exclusivity rule.

- 1. Honest technical observation — "After X weeks running Y, here's what surprised me"

- 2. Contrarian-but-substantive — "Everyone says X. I tried it. Here's where it fails."

- 3. Specific artifact + receipt — image / chart / code + 1–2 lines

- 4. The honest admission — "I was wrong about X. Here's why I changed my mind."

- 5. The reasoned recommendation — "Switched from A to B. Specific tradeoffs:"

- 6. The thoughtful question with context — Bluesky audience replies seriously

- 7. The quote-post commentary — adds substantive view to existing post

+ ### Compare without invented thresholds

- ### Dead patterns

+ Use account-relative comparisons among posts of similar age, subject, and format. Report raw counts and explain the comparison set. Reposts divided by likes is a descriptive ratio; if likes are zero, mark it undefined. Never use a universal strong/breakout threshold or a follower-band multiplier (directional).

- - X / Twitter residue ("RTs appreciated", "via X user @...")

- - Engagement bait ("Drop a 🔥")

- - Hashtag stacks (>2 hashtags)

- - Marketing voice — Bluesky filters this fast

- - AI vocabulary in hook line

- - "Birdsite" complaining — dead pattern as of 2024+

+ | Dimension | Evidence to return (directional) |

+ |---|---|

+ | Audience fit | Replies showing the requested community's use cases or questions; don't infer profession from a like. |

+ | Distribution | Observed reposts, quotes, and actual appearances on named surfaces, with collection times |

+ | Conversation | Substantive answers, author follow-through, and unresolved objections |

+ | Proof | Inspectable artifact or source, its conditions, and what it doesn't establish |

+ | Saturation | Repeated claims within the declared sample and the missing detail an original contribution could address |

- ### Format and length

+ Distinguish discovery location from diffusion. Finding a post in Discover and a custom feed doesn't establish movement between them. A prominent account's repost doesn't establish its contribution to subsequent counts. Overlapping feeds aren't independent audiences (directional).

- | Goal | Length (cap 300) | Format |

- |---|---|---|

- | Quick take | 100–250 chars | Single |

- | Thread / chain | 3–6 self-replies | Native threads, no markers needed |

- | Image-attached | 200–280 chars + image | Strong lift |

- | Quote-post commentary | 100–250 chars | High amplification |

+ Bluesky says it doesn't punish links; it also reports downranking toxic or spammy posts (official, Bluesky outlook, 2026-01). Don't reject root links or label low engagement a promotion penalty. Spam, disruptive repetition, and artificial social-signal manipulation are prohibited (official, Community Guidelines, 2025-09).

- ### Custom-feed targeting

+ ## Launch and proof posts

- For each recommended post, name 2–3 custom feeds it's likely to surface in. Increase odds by:

- - Including topic vocabulary the feed's algorithm filters on

- - Adding 1–2 hashtags that the feed targets

- - Posting when the feed's curator is active (some are manually curated)

+ When the brief concerns a launch, harvest the announcement, demo, proof, objection answers, and recap where available. Record what each post offers before the click, the actual next step, the author's relationship to the product, and evidence of qualified interest. A reply asking about compatibility or availability is more interpretable than an unexplained like count for that specific buying question (directional).

- ### Anchor accounts

+ Record relevant starter packs without treating them as guaranteed placement. Reference-list opt-out UI exists (official, App releases, 2026-09). Preserve exclusion choices and describe curator or participant accounts as observed sources, not an outreach target list.

- Identify 5–10 niche-relevant well-followed accounts whose repost would amplify the user's post. Note follower band and recent repost behavior on each.

+ In the brief, separate native interactions from destination visits or conversions. Only report private analytics if supplied with scope and authorization. Don't attribute a sale to a post from timing alone or generalize founder-versus-brand performance from another platform (directional).

- ### Timing

+ ## Measured signals

- - Best windows (US/EU mix): Tue–Thu 9 AM–12 PM ET and 4–7 PM ET

- - Slower than X — give a post 24–48h to develop before judging it dead

- - One thread per day max

+ Use these dated comparisons to design the sample, not as pass/fail scores.

- ### Saturated-take detection

+ | Result | Scope and use |

+ |---|---|

+ | Video median 5 interactions, images 4, links and text 3 each (measured, Buffer engagement, n=Buffer-published Bluesky subset, 2026-03) | Historical Buffer-user observations; inspect whether media carries proof instead of assuming automatic image lift. |

+ | Author replies associated with 5% higher engagement (measured, Buffer engagement, n=nearly two million cross-platform posts, 2026-03) | Within-account comparison; no causal or Discover-ranking conclusion. |

+ | Test audience-local Saturday 5 p.m., Saturday 6 p.m., or Sunday 9 a.m.; weekday evenings around 6–9 p.m. were stronger than late mornings (measured, Buffer timing, n=over three million posts, 2026-08) | Compare with the actual community's activity; no universal expiry interval or daily thread cap. |

+ | Substantial short-term repost increase associated with pack inclusion (measured, Starter-pack study abstract, n=over 50,000 packs, 2026-08) | No supported matched-control multiplier or current pack-size limit in the reviewed abstract. |

+ | 9.04% of eligible test posts reached a monitored feed's Top 50 within 24 hours (measured, Custom-feed study abstract, n=5,000 monitored feeds, 2026-08) | A defined monitored-panel outcome, not the chance of reaching any Bluesky feed. Record actual appearances. |

- Bluesky's tech-niche saturated takes in 2026 cluster around: "AT-Proto vs ActivityPub," generic "Why I left Twitter" posts, "AI is a bubble" / "AI is the new electricity" cycles, anti-OpenAI / anti-Anthropic tribal posts. Verify saturation live.

+ The prior unlabeled ratios and hot-feed anecdotes aren't benchmarks. Any new measurement needs post URLs, query and feed identifiers, capture times, raw counts, and a declared comparison rule. Re-derive when surfaces or the audience change (directional).

## Examples

- ### Example 1: Full DNA + amplification-path call

+ ### A bounded distribution observation

- Bad — count without path:

- ```

- @dev.bsky.social got 80 reposts on a post about RAG.

- ```

+ Illustrative counts and placeholders below teach the reporting shape; they aren't observed posts.

- Good — repost path, custom-feed surfacing, anchor noted:

- ```

- @dev.bsky.social (4.1k followers) — 412❤ / 87🔁 / 38💬 / 14🗨

- Repost-to-like ratio: 0.21 — breakout

- Surfaced via: Discover (initial) → reposted into 2 niche custom feeds (Tech feed by @curator-a, LLM feed by @curator-b)

- Reposted by anchor @ml-anchor.bsky.social (38k followers) at the 4h mark — drove ~60% of total amplification

- > Six weeks of running a RAG eval harness against real customer queries. The retrieval recall numbers everyone quotes are measured wrong.

- Hook type: honest technical observation + receipt promise

- Length: 224 chars

- Format: single post + chart screenshot reply with alt-text

- DNA: dev-tone + specific timeframe + named-counterpoint to common belief

- ```

+ > Source: [post URL], by [full handle], found in [feed URL] at [capture time]. Observed counts: 120 likes and 18 reposts (illustrative). Repost/like ratio: 0.15 (illustrative calculation). Replies and quotes weren't captured. The author's recent comparable-post baseline isn't available.

+ >

+ > The screenshot shows [specific artifact]. A reply asks about [use case]. The post also appeared in [other feed URL], but collection doesn't establish which surface distributed it first. Keep it as a relevant artifact example; distribution strength remains unclassified.

- ### Example 2: Marketing-voice rejection

+ Why it works: raw counts remain separate from causal attribution.

+ The report states missing evidence instead of inventing a breakout label.

- Surface a post that has high counts but fails the audience-fit gate:

- ```

- @brand.bsky.social — 850❤ / 12🔁 (repost ratio 0.014, far below 0.10 floor)

- Hook: "Unlock the power of next-gen AI agents with our cutting-edge platform 🚀"

- Failure: marketing voice, AI-vocabulary in hook, rocket emoji, near-zero repost ratio.

- The like count is shallow (likely company employees + bot accounts). Bluesky's audience is filtering this. Exclude from recommendations and avoid the pattern.

- ```

+ ### High likes with little decision evidence

+ Illustrative comparison, not a real brand judgment:

+ > [Post URL] has more visible likes than the other sampled posts. Its replies don't explain an evaluation need, and the destination offers no inspectable proof. Keep it in the dataset; rank it below [evidence-bearing post URL] for this launch brief's goal of finding useful product explanations.

+ Why it works: the ranking follows the brief's stated evidence goal.

+ It makes no bot, employee-engagement, or platform-penalty accusation.

## Checklist

- Before returning the Bluesky section of the brief:

- - [ ] Every cited post has full federated handle (`@user.bsky.social`), like count, repost count, reply count, quote count, URL

- - [ ] Repost-to-like ratio computed and used as primary signal

- - [ ] Amplification path noted on every cited example (Discover / custom feed / anchor repost)

- - [ ] No score depends on view counts

- - [ ] (Opt-in mode only) Custom-feed targeting prescribed for every recommended angle (2–3 feed names, verified at runtime)

- - [ ] Anchor-account repost-path list included (5–10 accounts with follower bands) — research output: shows who currently amplifies in this niche, not who the user should target

- - [ ] (Opt-in mode only) Hook bank entries each tagged to a taxonomy type from this skill

- - [ ] Dead-pattern list applied — no recommendation uses X-residue, hashtag stacks (>2), marketing voice, or rocket emojis

- - [ ] (Opt-in mode only) No more than 2 hashtags prescribed per recommended post

- - [ ] All background tabs closed

+ - [ ] Collection scope, timezone, session state, and inaccessible surfaces are recorded.

+ - [ ] Trending/Explore availability was checked; Discover, filtered searches, and relevant custom feeds use observed URLs.

+ - [ ] Each cited post has a full handle, canonical URL, capture time, age, and raw metrics or explicit missing values.

+ - [ ] Discovery surface is separated from inferred diffusion; no attribution percentages or universal ratio thresholds appear.

+ - [ ] Comparisons state the account baseline and sample limits; counts aren't treated as conversions.

+ - [ ] Format, source proof, disclosures, and meaningful reply context are captured.

+ - [ ] Launch evidence and unresolved buyer questions are distinguished from popular but unrelated material.

+ - [ ] Feed criteria and actual inclusion are separated; pack opt-outs are respected.

+ - [ ] Any recommended angle is tied to observed evidence, without copied post text presented as original writing.

+ - [ ] Only tabs opened for this task are closed; no account interactions or messages were sent.

- ## Composition / References

+ ## References

- - Pairs with `social-bluesky` (content domain) for writing the actual post from the brief.

- - AT-Proto handles include the instance: always cite `@user.bsky.social` (or custom domain handle) — not bare `@user`.

- - Use the agent's universal output schema.

+ - [Visibility declaration](https://bsky.network/blog/content-visibility-declaration/), 2026-09-01.

+ - [Repost controls](https://bsky.app/profile/bsky.app/post/3msqpuobiwk2t), 2026-08-10.

+ - [Search rollout](https://bsky.app/profile/bsky.app/post/3mqafridzgk2e), 2026-07-09; [Search migration](https://github.com/bluesky-social/social-app/releases/tag/1.128.0), 2026-07-16.

+ - [Thread numbering](https://bsky.app/profile/bsky.app/post/3msqpusnigc2t), 2026-08-10; [App releases](https://github.com/bluesky-social/social-app/releases), 2026-09-03.

+ - [Post lexicon](https://github.com/bluesky-social/atproto/blob/main/lexicons/app/bsky/feed/post.json), undated; reviewed 2026-09.

+ - [Photos release](https://bsky.app/profile/bsky.app/post/3mnslrkd6ok2g), 2026-06-08; [Video release](https://bsky.app/profile/bsky.app/post/3mtwf7gxkwc2r), 2026-08-25.

+ - [Bluesky outlook](https://bsky.social/about/blog/01-28-2026-bluesky-2026-predictions), 2026-01-28; [Community Guidelines](https://bsky.social/about/support/community-guidelines), 2025-09-19.

+ - [Buffer engagement](https://buffer.com/resources/state-of-social-media-engagement-2026/), 2026-03-05; [Buffer timing](https://buffer.com/resources/best-time-to-post-on-bluesky/), 2026-08-05.

+ - [Starter-pack study abstract](https://arxiv.org/abs/2608.17489), 2026-08-18; [Custom-feed study abstract](https://arxiv.org/abs/2608.13874), 2026-08-14.

+ Re-validate when: navigation or feed names change; ranking code changes; search migrations occur; policy updates or new vendor reports appear.

+ Validated: 2026-09