State of Agent Skills: In 15 Nights, 1 in 7 Watched Skills Changed at Least Once

Skill Harbor now lists 2,872 skills. On the night of November 6 to 7, we checked 2,739 snapshot rows and held a fingerprint for 2,726 of the listed skills, 94.9 percent of the catalogue. 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 October 24 to November 7, fifteen 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 edition's clean 12-night series. The catalogue kept growing during the window: 106 new listings were created, and the checked rows rose from 2,686 to 2,739. The 15-night numbers are therefore a moving cohort, not a fixed panel, and we say so again in the honesty section.

The headline numbers

• 2,726 skills fingerprinted on November 7, out of 2,872 listed. 146 listed skills still have no snapshot: 52 were created since October 24 and their snapshots have not been created yet, 92 are older, and 2 have no creation date recorded (Muse Proxy and Muse Connector Template). The 11 Telep listings are inside the 92 older ones: their repository is still private, so no snapshot can be fetched.
• 635 file-change events in 15 nights (October 24 to November 7), across 409 listings, 129 sellers, and 134 repositories or hosts.
• 42.3 changes per night on average, ranging from 8 to 72. On nights where the checked count is preserved, that is 0.3 to 2.6 percent of checked files per night.
• 409 listings changed at least once in 15 nights: about 1 in 7 watched listings. 296 changed exactly once, 113 changed at least twice (27.6 percent of all changers), and 19 changed four times or more.
• For a file that changed again, the median gap between two changes was 3.0 days (mean 3.4 days, 226 gaps across 113 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 few days.
• Final disposition of the 635 events, as recorded on November 7: 595 re-baselined as benign (93.7 percent), 33 followed by a regenerated install prompt (5.2 percent, across 31 listings), 6 left under a visible "pending verification" warning (0.9 percent, across 2 listings), 1 still pending (0.2 percent). These are final dispositions, updated when a later train regenerated a prompt. They are not the split decided on triage night itself.

Changes travel in packs

The same pattern as the last edition, now measured over 15 nights. The most active sources were:

• remotion-dev/skills: 36 events
• coreyhaines31/marketingskills: 35 events
• earthtojake/text-to-cad: 28 events
• EL4CTEO/blender-skills: 22 events
• affaan-m/ECC: 21 events
• heygen-com/hyperframes: 21 events
• software-mansion/argent: 20 events
• addyosmani/agent-skills: 20 events
• cloudflare/skills: 18 events
• JimLiu/baoyu-skills: 18 events
• morluto/rea: 17 events
• aklofas/kicad-happy: 16 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. Remotion's 36 events came from 12 listings moving on the same version bumps. The single most restless listing was Reverse Engineer Anything: 12 changes in 15 nights, on 12 of the 15 nights. Next came Maritime with 11 changes in 11 nights, still close to its daily rhythm from the last edition. Grandet API Price Finder produced 8 events across its three files, including two endpoints moving in opposite directions on the same night. Answer Me with HTML changed 7 times, McClipFace 6 times.

By catalogue category, in absolute events: Developer tools 145 events across 90 listings, Video 48 across 21, Engineering 45 across 19, Marketing 43 across 30, Design 37 across 26, Media 34 across 20, Productivity 32 across 21, and Trading 27 across 21. Video is the mover this window: Remotion, HyperFrames, and Higgsfield bumps together. Finance went quiet, 3 events across 3 listings, after 23 events in the last edition. Writing stayed restless by repetition: 14 events across only 6 listings. Catalogue categories are messy (several near-synonyms), so treat the ranking as a shape, not a precise league table.

One night, dissected: October 30

On October 30, 72 files changed, the busiest night of the window. The triage that morning auto re-baselined 67 routine changes and flagged 5 for a human look. The 5 included: Grandet, twice, its two endpoints contradicting each other, latest.json rolling back from build 20261018b to 20261017 while price-finder.json advanced from 20261017 to 20261018b on the same night; Entra Permissions Management, massively rewritten, the EPM service retired on 2025-10-01 and the skill reoriented toward Defender CSPM and Delinea; and Meshy 3D Generation and Meshy 3D Printing, whose CLI jumped from 0.4.3 to 0.7 with a Node 22.12+ requirement while our prompts still pin 0.4.3. The routine 67, by contrast, were prose, version bumps, and small fixes, including the third consecutive re-baseline of the earthtojake CAD files as cadgen moved from 0.8.2 to 0.8.3.

The same day, on Max's word, the follow-up landed: the Entra prompt was regenerated with disclosure and exchanged, the two Meshy listings were re-baselined keeping the tested 0.4.3 pin (the jump to 0.7, and now 0.8, is left as a separate tested regeneration), and Grandet was left parked with its badge kept. By the cutoff, October 30's 72 events stand at 69 re-baselined, 1 prompt regenerated, and 2 under a pending-verification warning, both Grandet's.

Other non-routine changes in the window, each dated:

• October 25: two prompts regenerated, Tutorial Engineer and Portfolio Manager.
• October 26: ten prompts regenerated in one train: McClipFace, Neon AI Gateway, Paper slide deck, Longbridge, five ECC security skills (Delivery Gate, Open-Source Pipeline, Perl, Django, Laravel), and Datawright web tools. Grandet's October 26 event is the one event still pending at the cutoff.
• October 27: Big Arrow on the Screen gained a new Cards feature, a floating card for values, secrets, and questions, with secrets handled through pass, keychain, op, env, or file, never on the command line. It was flagged for a human look, not auto re-baselined. The same night, Grandet moved to build 20261018b, RunPod renamed its hosted endpoint from mcp.getrunpod.io to mcp.runpod.io, earthtojake migrated cadgen from 0.7.26 to 0.8.0, and twelve Remotion skills bumped from 4.0.538 to 4.0.539.
• October 28: four prompts regenerated: BeCreative, Higgsfield video explainer, Parametric CAD modeling, and IB 0DTE Credit Spreads.
• October 31: four watch items examined in depth, all re-baseline suffices: two Bankai skills whose install script was hardened (truncated-download guard, version validation, atomic install), HyperFrames Media Use with a new sign-in flow, and a Salesforce policy skill rewritten from CLI sf to MCP tools in v2.0. The prompts pin the old versions, so none of it makes a prompt false. Grandet was flagged again the same night.
• November 1: the quietest night of the window, 8 changes. The first check run died before recording anything and a rerun completed the night; all 8 events come from the rerun. The red-line detector also tripped on Graphify in that rerun.
• November 2: five prompts regenerated: McClipFace again, and the text-to-cad SDF, SRDF, and URDF files.
• November 3: 44 events, all re-baselined. Seven examined, including three MediaKit skills, Baoyu XHS Image Cards, and the two Meshy listings with an explicit reserve: the re-baseline is not an upgrade to 0.8.
• November 4: the McKinsey prompt was regenerated: the public descriptions still said "67 methods, 12 categories" where the upstream now has 62 methods, 8 categories, version 2.4.1, and a preferred pip install mck-ppt. Graphify's red line tripped again, and its suspension is dated November 4 in D1.
• November 5: GPT Taste was rewritten from v2 to 2.5.0 and re-baselined after a full read found no suspect content. Six prompts regenerated: Higgsfield Generate (default model now GPT Image 2.6 Sunburst), Aeon, PennyLane, PyHealth, scikit-learn, and SDR master prompts.
• November 6: Git Commit re-baselined after a rewrite that adds conflict checks. Two Huawei prompts regenerated and exchanged: FunctionGraph trigger-create (now also OBS event triggers, not just TIMER) and OBS bucket-create (commands now wrapped in skill-quality-cli with telemetry, opt-out available).
• November 7: 55 events, all re-baselined. Ten watch items examined, including Cloudflare One (a new section sending CF-Access service-token headers), four Argent skills condensed to roughly half their size, three Momentic skills noting a hosted-MCP instruction that only matters if the pin ever advances, and Cloudflare replacing C3 with cf init. All re-baseline suffices, because every prompt pins the reviewed commit.

Gone, and still listed

Five listings are now unlisted: MuseFi, Manjangilchi, the Shipshitdev fullstack workspace initializer, the Terrylica calendar event manager, and, new since the last edition, Myk Pono's ultimate-seo-geo, removed under the permanent opt-out decided October 31. 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 146 uncovered listings above, not hidden.

On the morning of November 7, 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

Six things we got wrong, missed, or cannot prove cleanly. They stay in the published version if this draft is published.

1. The night of November 1 ran twice. The first check run at 07:39 UTC died before recording anything: its local log holds only the "checking 2,719 snapshots" line. A rerun at 07:46 completed the night and found 8 changes, the quietest night of the window. The table holds exactly the rerun's 8 changes, stamped 07:56 UTC, so nothing was double-counted and nothing was lost. The 8 is a real quiet night, not a gap, but the first run's failure is part of the record.

1. Two check logs were not preserved locally. The local log files for the October 25 and November 6 checks are missing, so the per-night checked totals for those two nights are not available from logs. The event counts themselves are verified complete from the triage records: October 25 closed at 24 re-baselined plus 2 regenerated, exactly the table's 26, and November 6 at 34 re-baselined plus 2 regenerated, exactly the table's 36.

1. Two November 4 snapshots were stored truncated. Book to Skill and KiCad Project Analysis hit the 50 KB storage cap and their stored copies end mid-word. Live re-checks confirmed the upstream files are intact: the sections that looked deleted exist in the full files. Both were re-baselined after that verification, with a backup kept, under the rule decided November 4 (verify the full file live, archive, then re-baseline). Detection itself hashes the full 200 KB fetched, so no change was missed; only the stored copy is short.

1. Graphify's suspension has two dates in the record. The November 1 rerun log records the red line arming the suspension, and D1 records the current suspension as dated November 4. Both are true: the check script overwrites the suspension date with the current date on every trip, and the line tripped again on November 4. Max decided on November 5 to keep the suspension. The detector's earlier false positive on coreyhaines31-attribution did not recur in this window: attribution had zero change events, so nothing tripped it. Whether the detector itself was fixed was not verified for this edition.

1. The uncovered count rose, and its by-cause split was not recomputed. Listed skills without a snapshot went from 87 to 146. Most of the rise is new listings: 52 of the 146 were created since October 24 and their snapshots have not been created yet. The last edition's split (34 with no SKILL.md to fetch, 36 ambiguous, 17 fetch failures) was not recomputed for the 146, so no by-cause split is claimed here. The 92 older uncovered listings include the 11 Telep listings and 2 with no creation date recorded. A smaller residual is unattributed: the listed count rose by 110 while 106 listings were created in the window, and the table has no listed-at date to say which 4 listings account for the difference. The last edition carried the same kind of small residual (6).

1. The first week is still not a clean series, and neither is this window a fixed panel. The caveats from the last edition stand: the October 3 to 11 calibration week is reported separately and no single total is presented as if the method had been constant. In this window, coverage grew from 2,686 checked rows on October 24 to 2,739 on November 7, and 106 new listings arrived mid-window. The 15-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 (8 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, 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 15 nights, 1 watched skill in 7 moved, and more than a quarter of the movers moved again within days, a median 3.0 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 November 21, should be able to say whether the 3.0-day repeat gap holds, whether Reverse Engineer Anything's near-daily rhythm is an outlier or a class of its own, 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.

---

---

🤍 0