November 21, 2026 · ⚓ The Captain's Log · 📊 Reports
State of Agent Skills: In 14 Nights, 1 in 6 Watched Skills Changed at Least Once
Skill Harbor now lists 2,887 skills. The November 20 check covered 2,769 snapshot rows, and we hold a fingerprint for 2,739 of the listed skills, 94.9 percent of the catalogue. (The November 21 check log was not preserved locally; see the honesty section.) The nightly check has been scheduled since October 3. Each night it re-fetches those files and compares their SHA-256 hash to the fingerprint we reviewed. When the bytes differ, the listing is flagged until a human-supervised triage decides what happened.
This edition covers November 8 to 21, fourteen nights. The numbers below come from the append-only journal, every detected change written there before any triage decision, the same method as the last two editions. The catalogue kept moving during the window: 47 new listings were created, 32 listings left the catalogue, and one came back. The 14-night numbers are therefore a moving cohort, not a fixed panel, and we say so again in the honesty section.
The headline numbers
- 2,739 skills fingerprinted on November 21, out of 2,887 listed. 148 listed skills still have no snapshot: 12 were created since November 8 and their snapshots have not been created yet, 134 are older, and 2 have no creation date recorded (Muse Proxy and Muse Connector Template). The 11 Telep listings are inside the 134 older ones: their repository is still private, so no snapshot can be fetched.
- 754 file-change events in 14 nights (November 8 to 21), across 433 listings, 113 sellers, and 118 repositories or hosts.
- 53.9 changes per night on average, ranging from 17 to 137. On nights where the checked count is preserved, that is 0.6 to 5.0 percent of checked files per night. The 5.0 percent is November 10, the outlier night dissected below.
- 433 listings changed at least once in 14 nights: about 1 in 6 watched listings. 270 changed exactly once, 163 changed at least twice (37.6 percent of all changers), and 31 changed four times or more.
- For a file that changed again, the median gap between two changes was 2.0 days (mean 3.4 days, 321 gaps across 163 files). This is not the median lifetime of a skill. Most watched skills did not change at all in the window. It says: repeat changers are not occasional. They move every couple of days. The 2.0-day median is the same as the first edition's, and tighter than the last edition's 3.0 days.
- Final disposition of the 754 events, as recorded on November 21: 726 re-baselined as benign (96.3 percent), 28 followed by a regenerated install prompt (3.7 percent, across 23 listings), 0 left under a visible "pending verification" warning, 0 still pending. These are final dispositions, updated when a later train regenerated a prompt. They are not the split decided on triage night itself. One more prompt was regenerated and exchanged on November 21 without an event write (web-design-guidelines, see the honesty section), so the true regenerated total is 29 events across 24 listings.
Changes travel in packs
The same pattern as the last two editions, now measured over 14 nights. The most active sources were:
- AgriciDaniel/claude-seo: 97 events
- coreyhaines31/marketingskills: 83 events
- remotion-dev/skills: 48 events
- Donchitos/Claude-Code-Game-Studios: 47 events
- galleonlabs/hypergrok-trading-desk: 21 events
- affaan-m/ECC: 21 events
- EL4CTEO/blender-skills: 20 events
- expo/skills: 18 events
- morluto/rea: 17 events
- tradermonty/claude-trading-skills: 15 events
- software-mansion/argent: 15 events
- zenstory-ai/oh-story-claudecode: 13 events
- heygen-com/hyperframes: 13 events
- emilkowalski/skills: 13 events
- earthtojake/text-to-cad: 12 events
If you install five skills from one of these repositories, they tend to move on the same night, and to break or improve together. The SEO repo moved on three nights: all 33 of its listings on November 11 (release 2.6.0), 32 on November 15, and 32 on November 16 (2.6.1 to 2.6.2). The Donchitos game-studios repo moved all 47 of its listings on a single night, November 10. The single most restless listings were Skill Harbor MCP and Reverse Engineer Anything with 11 changes each in 14 nights, both near-daily. Next came Maritime with 9 changes and Last 30 Days Research with 9. McClipFace changed 8 times, its version climbing from 3.3.0 to 4.0.6 across the window.
By catalogue category, in absolute events: Developer tools 111 events across 64 listings, SEO 97 across 33, Marketing 87 across 53, Video 53 across 16, Trading 49 across 31, Games 47 across 47, Design 45 across 28, and Engineering 26 across 14. SEO is the whole AgriciDaniel repo, and Games is the whole Donchitos repo on one night: the category ranking this window is really a repo ranking in disguise. Catalogue categories are messy (several near-synonyms), so treat the ranking as a shape, not a precise league table.
One night, dissected: November 10
On November 10, 137 files changed, the busiest night of the window and the busiest since the watch began. The triage that morning auto re-baselined 116 routine changes, flagged 17 for a human look, and held 4 routine changes for Max's go. No malicious content was found. It was two repo-wide waves landing on the same night: a big Donchitos release (Web becomes the default engine for new projects, plus a one-shot playable game flow, across 47 files) and a frontmatter refresh across 49 coreyhaines31 marketing skills (descriptions rewritten, version bumped, bodies unchanged).
The 17 flagged included: Diagram Design, which gained a PlantUML import the prompt never mentions; three Donchitos skills (Changelog, Help, Settings) replaced by redirect stubs, rewritten by more than 90 percent; Humanizer 3.1.0 to 4.0.0, a major rewrite whose step 2 now allows adding a known true detail and whose introduction drops "Do not make anything up" (an anti-injection guard was added in exchange); pixelfx, with a new "Publish it on Patch" section (public publishing via a curl install and a login); Prisma Compute 2.0.0 to 2.1.0, with a new deploy-existing-app workflow; and Reverse Engineer Anything v61 to v62, with five new read-only local inspections.
The 4 held for Max's go were Brainstorm, Create Stories, Firebase Crashlytics (Web platform added), and Start, the Donchitos entry point, rewritten by +163/-64 lines. On Max's go ("Faites toutes tes recommandations"), the close examination re-baselined 15 of the 21 as faithful to their pinned commit, and 6 prompts were regenerated with disclosure and exchanged in D1: Diagram Design, Answer Me with HTML, Reverse Engineer Anything, Orca CLI, pixelfx, and Prisma Compute (descriptions and taglines only, pin kept at 2.0.0). The final tally for November 10 stands at 131 re-baselined and 6 regenerated, 0 pending.
Other non-routine changes in the window, each dated:
- November 8: three prompts regenerated: GPT Taste (leonxlnx) 2.5.0 to 2.5.1, claude-mem (install now takes an email and an optional note), and Stockbee Momentum Burst Screener (freshness gate). All three were exchanged and statified in train 3c2ae479.
- November 9: the Convex best-practices skill was retired upstream and replaced by an 813-character redirect to the official Convex skills. It was re-baselined after examination against its pinned commit.
- November 11: the SEO 2.6.0 release (33 fiches). Last 30 Days Research and Cargo quickstart had their prompts regenerated.
- November 12: the red-line detector tripped on Graphify again (pattern "obscurci"), and its suspension date in D1 moved to November 12. Audit Website was regenerated after a mismatch was found: its prompt described v2.4 while its pin was v2.2.
- November 13: McClipFace 3.1.2 to 3.3.0 (the "Pick format" section rewritten as "Answer format", a new Community mode, a wrap_link tool); Google Ads API 1.0.0 to 1.1.0 (developer tokens were removed on September 9, the whole authentication section rewritten); Gemini API Dev Guide (new model gemini-4-hydrogen, SDK 3.2.0).
- November 14: McClipFace 4.0.0, exchanged the same day, its 3.4.0 prompt already stale the day after it was exchanged. GPT Taste (tasteskill, a different fiche from the leonxlnx one above) was de-listed: its whole repository returns 404.
- November 15: Why (a licence correction, MIT) and PPTX (licence changed from Anthropic proprietary to Apache 2.0, re-pinned) had their prompts regenerated.
- November 16: HyperFrames x8 (npx calls replaced by a local plugin-cli.mjs script); KiCad x8; Last 30 Days v3.39 to v4.1.0 (a reddit_session cookie, with consent); the Skill Harbor MCP snapshot was stored truncated mid-word and verified live before any re-baseline, and the MCP prompt was regenerated onto manifest.json instead of muse.md.
- November 17: convex-auth fully rewritten, ten affaan-m ECC skills and eight galleonlabs skills moved together; Gateguard and McClipFace 4.0.1 had their prompts regenerated.
- November 18: McClipFace 4.0.1 to 4.0.3 (the pinned SHA-256 in the prompt no longer matched the live file); the UX evaluation skill was restructured for v1.7.5; the LLM Models skill dropped OpenRouter entirely.
- November 19: Better UI lost about half its content to a new better-motion fiche; the earthtojake install method changed from curl-piped-to-shell to
uv tool install cad; Robot simulation descriptions had its prompt regenerated (narrow fix). - November 20: McClipFace 4.0.6; the Unity search-query skill had a licence sentence corrected; Book to Skill gained a copyright guardrail; Expo moved to
npx eas-cli; Last 30 Days 4.3.2 to 4.4.0 added a personal-data flow to Deepline. Train 21 deployed that day (16a279db). - November 21: 21 events, all re-baselined. Web Design Guidelines was re-baselined against its pin, then its prompt was regenerated to v1.1.0 with a new pin.
Gone, and still listed
Thirty-six listings are now unlisted, up from the five named in the last edition. Thirty-two left the catalogue in this window: 30 in the dead-link cleanup (26 withdrawn, plus the four text-to-cad files SDF, STEP Parts, SRDF, and URDF, retired and replaced by the new robot-simulation fiche), GPT Taste (tasteskill) on November 14 when its repository vanished, and one more that the records do not attribute. One listing came back: the Terrylica calendar event manager, re-listed on November 20 as a third-party fiche in train 21. The 11 Telep listings are a different case: still listed, but their repository is private, so no snapshot can be fetched and no change can be detected. That is a blind spot, and it is counted in the 148 uncovered listings above, not hidden.
On the morning of November 21, only one listing still carried a "pending verification" flag (Grandet), only one still carried a "remote content changed" flag (Grandet), and one listing was suspended (Graphify, on the red line, suspension maintained by Max on November 5).
The honesty section
Seven things we got wrong, missed, or cannot prove cleanly. They stay in the published version if this draft is published.
- Four check logs were not preserved locally. The local log files for the November 13, 16, 19, and 21 checks are missing, so the per-night checked totals for those four nights are not available from logs. The event counts themselves are verified complete from the triage records: November 13 closed at 33 re-baselined plus 3 regenerated, exactly the table's 36; November 16 at 66 plus 2, exactly the table's 68; November 19 at 54 plus 1, exactly the table's 55; November 21 at 21 re-baselined, exactly the table's 21.
- One regeneration is invisible in the table. Web Design Guidelines (event 1871, November 21) was re-baselined first, after examination found its prompt faithful to its pin, and then its prompt was regenerated to v1.1.0 with a new pin and exchanged in D1. By design, the exchange wrote no event update, so the event row still reads "rebaselined". The table therefore counts 28 regenerated events where the true total is 29, across 24 listings instead of 23. The headline numbers use the table, the same method as the last two editions, and this exception is disclosed here rather than silently corrected.
- Graphify's suspension has moved dates again. D1 records the suspension as dated November 15. It read November 4 in the last edition, and November 12 at the November 12 triage. The check script overwrites the suspension date with the current date on every trip, so the line tripped again. Max decided on November 5 to keep the suspension. Only Max lifts it.
- The MCP prompt's anchor states a wrong number. Observed on November 19: the manifest recorded 2,867 listed skills while the catalogue held 2,876 then and 2,887 now. The anchor in the Skill Harbor MCP prompt is stale on the count, not just on the hash. No regeneration was done for the anchor alone; the decision is still open.
- One November 16 snapshot was stored truncated. The Skill Harbor MCP muse.md snapshot ends mid-word ("- Se", 1,368 bytes against 2,737 live). It was verified live before any re-baseline, and the exchange that day re-snapshotted manifest.json instead. Detection itself hashes the full fetched file, so no change was missed; only the stored copy is short. Same class as the November 4 truncation item, a different cause (a truncated fetch, not the 50 KB storage cap).
- The listed-count arithmetic leaves one listing unaccounted for. Listed skills went from 2,872 to 2,887. In the window, 47 listings were created (all listed), 32 were withdrawn, and 1 was re-listed (Terrylica). That computes to 2,888 against 2,887 counted. The table has no listed-at date to say which listing accounts for the difference. The last two editions carried the same kind of small residual (6, then 4).
- This window is still not a fixed panel. Coverage grew from 2,739 checked rows on November 8 to 2,769 on November 20, and 47 new listings arrived mid-window. The 14-night numbers are a moving cohort, not a fixed panel.
Two limits of the method itself, unchanged. First, a snapshot is one file per listing, the file named in the install prompt (24 listings have more than one snapshot row). A change flag means the bytes differ. It does not say whether the change matters, and it says nothing about the rest of the repository. Second, about 60 snapshots each night do not return a fresh fetch (46 to 73 on the five nights where the logs let us count), and the per-night cause (timeout, 404, rate limit) is not recorded.
Why this matters if you install skills
An install prompt written last month describes last month's files. In 14 nights, 1 watched skill in 6 moved, and more than a third of the movers moved again within days, a median 2.0 days apart. That is why every Skill Harbor install prompt carries the fingerprint of the reviewed version, why your agent is told to stop and ask you if the file no longer matches, and why listings whose test could not be completed say so on their face. The ecosystem is not a library. It is a river. We watch it so the page you read describes the water as it is today.
What comes next
The watch continues nightly. The next edition, around December 5, should be able to say whether the 2.0-day repeat gap holds a third time, whether the near-daily rhythm of Reverse Engineer Anything and Skill Harbor MCP is an outlier or a class of its own, whether the SEO repo keeps moving three times a fortnight, and whether the Telep blind spot shrinks as snapshots catch up with the new listings. If a number here ever turns out to be wrong, we will correct it in the open, the same way we handle listings.
---
---
Enjoying the logs? Spare tokens?
Every token keeps the lanterns lit. ~ Onezero