A 16-year-old author's /unlazy: stop AI coding assistants from saying "done"; 3,000 stars in three weeks
What it is A skill for AI coding assistants. Before it starts, it writes a check command for what "done" means, and it won't let the assistant wrap up until those commands pass.
Why it took off In 2026, agents could run long tasks but often stopped at ~80% and called it done. Everyone ran into this. The author's reputation from another library with 110k stars got YouTube and LinkedIn creators to cover it on their own, and he put /unlazy into the prompt of one of his million-view posts.
CHow you can follow Don't build another generic completion gate. Copy the playbook instead: name a complaint agent users hate but that has no name yet as a slash command, ship a script and a hook the next day, pair it with a one-line demo that's easy to film (like making a game in one sentence), send it to creators who cover Claude Code, and pick up long-tail traffic from skill directories.
- Data (updated daily) · Repo created 2026-08-09 · 4,327 stars · 7 days +113
GitHub · stars Data 2026-10-09 → 2026-11-08 (31 days, 30 observations). Points before 10-28 come from deeply / archive backfill, not every day has one
unlazy is an "anti-laziness" skill for Claude Code and Codex. Before work starts, it writes an acceptance checklist, and every item gets a runnable check command. A task only counts as done when the commands pass, and an agent that tries to stop early gets blocked. The idea isn't new: a similar older skill has 26× its installs. It hit 3,000 stars in three weeks because it names a complaint everyone has, has a name you understand at a glance, and rode on the author's reputation from a 110k-star library. Below: how it grew, what you can take from it, and why copying it isn't worth the effort.
"I built a skill that makes models a little less lazy"
On the night of Aug 11, 2026, @LexnLin (Leon Lin, who describes himself as "16 years old, vibecoder" and has 33k followers) posted on X: "This might be something big… I made a skill which reduces the laziness of models. Pretty good for Opus 5. go try /unlazy." It got 21k views, 146 likes, and 124 bookmarks. The repo had been created two days earlier (8-09).
Four days later he posted again: "okay I tested /unlazy more with Opus 5 and I'm ngl it's really good". 47k views, 633 likes, 1,062 bookmarks. Bookmarks outnumbered likes, which suggests people's response was "save it and use it later."
On Aug 23, he posted that the repo had reached 1,000 stars, and explained why he built it: agents keep stopping when a task "looks about done," and "it turns out a lot of people have this problem."
What it actually is
If you've run long tasks with Claude Code or Codex, you've probably seen this: the agent works for tens of minutes, then hands back a confident summary saying everything is done. You check, and only 80% is. By 2026 this was so common that Anthropic's own prompting guide for Opus 5.5 included specific advice on handling "premature wrap-up."
unlazy's approach is to stop the agent from being the judge of its own work:
- Before work starts, write a
GATES.md. Under each acceptance item, write aCHECK:command and anEXPECT:line (what the command should print); - A zero-dependency Node script runs those commands. An item is ticked only when the exit code is 0 and the output matches. An item ticked without evidence counts as not done;
- An optional Claude Code Stop hook: if the checklist isn't complete, the agent is sent back when it tries to stop;
- Big tasks are broken into a tree and handed to several sub-agents, and the main agent uses
--reverifyto re-check what they return.
Install with one line, npx skills add Leonxlnx/unlazy, then type /unlazy in a conversation. MIT license, free.
The first version (8-09) was really just a prompt, with a selling point called "Depth Tree": split the task into N levels, and give every leaf the whole task's time budget. The next day the author shifted focus to the checklist + script + hook setup above, and the README's top section no longer mentions Depth Tree. That step mattered: it went from "a prompt trick" to "a tool," which is easier to bookmark and easier to write about.
How the stars climbed
There's no day-by-day data for 8-09 to 10-15, so the middle points are pieced together:
| Timeline | Cumulative stars | What happened |
|---|---|---|
| 8-11 | — | Author's launch post, 21k views |
| 8-20 | — | YouTube channel AI LABS (179k subscribers) posted a video, "GitHub's #1 Trending Author's New Claude Skill Is Insane"; it has 124k views as of today |
| 8-22 | — | @0x3b33 on X (6k followers) posted "You don't know how good this skill is until you try it": 40k views, 1,019 bookmarks |
| 8-23 | About 1,000 | Author reporting star counts in a post |
| 8-25 | — | Shubham Saboo (author of awesome-llm-apps) wrote a long LinkedIn post: "Someone built a skill that won't let your agent say 'done' until it shows evidence" |
| 9-03 | 2,986 | A count noted in an external review article; same day as the author's last code commit |
| 9-05 | — | Author posted "6 hours of GPT 5 Astra making a forest," 1.08M views; the prompt in the replies starts with "First find the /unlazy skill, read it, and use it for the whole task" |
| 9-13 to 9-25 | — | Three entries on the Trendshift JavaScript trending chart (ranks 16–19) |
| 10-16 | 3,946 | |
| 11-06 | 4,311 | 339 forks; +139 in the last 7 days, +133 in the 7 days before that |
(GitHub star count, 11-06; earlier points come from the author's posts, external articles, and Trendshift back-calculation)
The fastest stretch was about 8-23 to 9-03,11 days, gaining roughly 2,000 stars, right after AI LABS' video, Pyro's post, and Saboo's LinkedIn post (inferred; no day-by-day data proves it directly). AI LABS' title is worth a second look: the hook isn't the skill itself but "the new work of GitHub's #1 trending author." That title comes from his other library, taste-skill, with 110k stars.
From September on, it grew through the long tail of content sites and skill-directory sites: skills.sh, skillsmp, various "Claude skills guides," Medium breakdowns, and Facebook posts saying "comment UNLAZY to get the tool." It gained 5 to 30 stars a day, and there were no new commits for two months. It never made Hacker News or Product Hunt.
What people who tried it say
Cameron Breen in Australia posted a LinkedIn test with a video: one sentence asked it to build a top-down shooter in the style of Hotline Miami. It wrote a contract first (each module's interface, which files each agent could touch), then split the work into 14 checklists and 38 checks, ran 8 agents in parallel in groups of four, finished in 40 minutes, and 8 of them passed on the first check.
He also named the downside:
"It didn't save me any tokens. The control used a lot less, and unlazy wrote 5,400 lines of tests for a game that was only 3,300 lines. … For anything under half an hour, that setup isn't worth it."
He also said the out-of-the-box version was "slow enough that you probably won't want to use it," because it runs subtasks one at a time by default. He changed it to run in parallel himself. AI LABS' video raised the same issue, and then put the parallel version into its paid community.
There's also a YouTube review called "unlazy — More Tokens, Worse Code?" Its conclusion: on one UI task, adding the skill used 19% more tokens and missed a keyboard interaction. The README itself admits that the 6 early comparison tests can't be reproduced.
So it solves a real problem, but the cost is real too: more tokens, more test code, slower. Good for multi-hour jobs, not for everyday small edits.
Who it's competing with
| Project | Scale | How it differs from unlazy |
|---|---|---|
| The verification-before-completion skill in obra/superpowers | The whole repo has 313k stars; this one skill has been installed 267k times on skills.sh | The de facto standard for this, but it's only a prompt-level discipline of "verify before you say done" |
| The Ralph loop family (ralph-claude-code, ralph-orchestrator, ralphy, etc.) | 9,721 / 3,192 / 2,983 star | Wraps the agent in a loop that keeps it running; whether the job is done is still the agent's own call |
| lazycodex | 3,880 star | A full Codex orchestration framework, heavy |
| planning-with-files | 28k stars | Uses files to manage plans and resume from checkpoints, but doesn't handle acceptance |
| Anthropic (official) | — | Provides a Stop hook mechanism, and the guide offers writing advice; unlazy's hook uses this same mechanism |
(GitHub stars and skills.sh install counts, 11-06)
unlazy has 10.1k installs on skills.sh, about one twenty-sixth of superpowers' same-type skill. It didn't open a new lane. It took a problem everyone already has, named it well, and added an executable checker layer. More stars doesn't mean more people using it.
Why this one
- The problem is widespread, and the timing is right: this was the year agents could run multi-hour tasks, and "it's 80% done, so it's finished" was something everyone ran into.
- The name is the ad:
/unlazyis clear at a glance and can be pasted straight into other people's prompts. The author did exactly that, putting it in the first line of a prompt in a post that got a million views. - The author's fame doesn't carry it: the same thing from an unknown author would most likely stall at a few hundred stars. Superpowers' similar skill has a huge install base, yet it was never singled out to blow up.
- Backed by papers: the README lists a set of papers from 2025–2026 on models slacking off and quitting early, giving content sites and video creators ready-made material.
If it were you
One-line verdict: building a "stop your agent from slacking" skill to copy isn't worth it; studying how it got a skill to stars is very much worth it.
Who it suits: developers who already have a few thousand followers on X or YouTube, or who have had a project go viral; or people who can build a "small but sharp" agent tool, not just a prompt.
Who it doesn't suit: new authors with no distribution channel who want to go head-on in the "acceptance / completion gate" space. superpowers already holds it, and unlazy has taken the name.
How to do it: don't build another general-purpose "completion gate." Find what it does poorly: make it parallel by default, low-overhead, suited to everyday jobs under half an hour; or do one scenario, such as "frontend changes must be screenshot-compared" or "data migrations must run a rollback," gates a single command can verify. The more general play is to copy its playbook to find the next pain point: take a problem every agent user complains about but nobody has named on its own, give it a verb-style slash command name; ship the prompt version on day one and the script and hook on day two so it looks like a tool; and put a paper and a reproducible comparison in the README.
How to get users: the first wave comes from your own X, posting real-world tests of "what I did with it," not just "I built a skill." The second wave goes after YouTube creators and LinkedIn writers covering Claude Code / Codex, giving them a demo they can film (one sentence to a game is best for this). The third wave is submitting to skill directories like skills.sh, and the long tail will follow. Write your skill into the prompts of your other content, so one exposure brings two repos.
Why now, and not earlier: agent skill directories and "best Claude skills" content sites are hot right now, and the vendors keep folding this kind of approach into their products. Once a specific pain point is fixed by the vendor, the skill targeting it stops getting installed. The window is measured in months, not years.
What would mean you read it wrong: it hasn't passed 200 stars two weeks after launch, or it has many bookmarks but installs are negligible; or Anthropic or OpenAI build a switch for the exact problem you're solving directly into the product.
Research notes: competitor table, acquisition evidence, full research report
- Competitors
Name Where Launched Scale How it differs from this one obra/superpowers · verification-before-completion Open source 2025-10 Repo 313k stars; this skill 267k installs on skills.sh Pure prompt discipline; unlazy has an executable ledger, a checker, and a Stop hook The Ralph Wiggum loop (frankbria/ralph-claude-code, ralph-orchestrator, ralphy) Open source 2025-08–2026-01 9,721 / 3,192 / 2,983 stars An outer loop runs until completion, and the agent still judges completion itself; unlazy fixes executable acceptance conditions in advance code-yeongyu/lazycodex Open source 2026-05-25 3,880 stars A heavy Codex orchestration framework; unlazy is a single skill file OthmanAdi/planning-with-files Open source 2026-01-03 27,905 stars Manages plans and resume from checkpoints, doesn't handle acceptance uditgoenka/autoresearch Open source 2026-03-13 6,805 stars A change-verify-keep/discard loop aimed at research optimization Anthropic's Stop hook plus Opus 5.5 prompting guide countermeasures for early quitting Big-company features About 2026-09 Official Official guidance gives writing advice and the hook mechanism; unlazy is a packaged implementation - Where users come from
- Verified Author on X: 8-11 launch tweet 21,027 views; 8-15 "tested more with Opus 5" 46,919 views, 1,062 bookmarks; 8-23 "1,000 stars" (fxtwitter)
- Verified YouTube creators: AI LABS 8-20 "GitHub's #1 Trending Author's New Claude Skill Is Insane" 123,508 views; AICodeKing 8-26 7,815 views
- Verified X/LinkedIn big-account reposts: @0x3b33 8-22 39,718 views, 1,019 bookmarks; Shubham Saboo 8-25 LinkedIn long post; @DanKornas 9-08
- Verified Author's viral post driving sales: 9-05 GPT 6 Astra forest post 1,083,803 views; the prompt in the reply (166,354 views) opens with "first find the /unlazy skill"
- Verified SEO/aggregator sites/skill directories: dozens of pieces including aisignal, ssdnodes, mindstudio, fervorai, heatdrop, reveneau, Medium, skills.sh, and skillsmp, which carried daily growth of 5–30 stars after September
- Inferred HN / Product Hunt / paid ads: our hackernews library returns 0 results for unlazy; none found elsewhere
Why unlazy (Leonxlnx/unlazy) took off
In one line: an "anti-slacking" skill for coding agents like Claude Code and Codex (SKILL.md + Node checker script + Claude Code Stop hook). v1 (8-09) was a single-file Depth Tree prompt: split the task into N levels, with every leaf getting the whole task's time budget. v2 (from 8-10) shifted focus to "write the acceptance ledger first (GATES.md), each entry with a CHECK command and EXPECT; it only counts as done when the script passes, and the Stop hook stops the agent from quitting early."
Conclusion: it didn't get there on a unique technique. The author is the creator of the 110k-star taste-skill (33k X followers). In the Opus 5 era, on the pain point everyone hit, "the agent says it's done at 80%," he picked a name you understand at a glance: /unlazy. What pushed it past 1,000 stars and then to 3,000 was two waves of third-party amplification: one from YouTube's AI LABS (124k views), X's @0x3b33 (40k views), LinkedIn's Shubham Saboo, and dozens of SEO/aggregator sites; the other from the author's own 9-05 1.08M-view "GPT 6 Astra makes a forest" viral post, whose embedded prompt opened with "first find the /unlazy skill." After September, only the long tail from aggregator sites was left, at 5–30 stars a day.
1. Data
- GitHub (our github repo_profile / repo_day, as of 2026-11-06): 2026-08-09,4,311 stars at creation, 339 forks (fork/star 7.9%); +7 over the last 139 days, +7 over the first 133 days, biggest single day +35 (10-26). 10-16 at 3,946 stars, 365 stars gained over three weeks. No suspect flags.
- Earlier curve (pieced together, not daily data):
- 8-11 22:06 UTC, author's launch tweet, 21k views (fxtwitter)
- 8-23, author tweets "1,000 stars" (fxtwitter)
- 9-03, reveneau.com coverage at 2,986 stars
- 9-07 about 3,115 stars: Trendshift shows 9-07 to 11-06 totaling +1,197, backed out from today's 4,311
- 11-06,4,311 stars
So from 8-23 to 9-03, over 11 days, it gained about 2,000 stars, the fastest stretch.
- Code: the last push was on 9-03 (fix: bind gate evidence…), and there have been no commits for two months; Trendshift shows merged PRs going from 9-07 to 11-06 as 0. On 8-23 the author used /unlazy to merge 4 external PRs at once (#3, #8, #14, #15), and 8-29 merged #29 (via the GitHub commits API).
- Installs:
leonxlnx/unlazyhas 10.1K cumulative installs on skills.sh. For comparison, the similarverification-before-completionin obra/superpowers is at 266.7K, and the whole taste-skill family is at 5.6M (skills.sh, 2026-11-06 reads). - Author Leon Lin (X @LexnLin, 33,256 followers, bio "16y/o | cracked vibecoder"): 24 repos totaling 125k stars, including taste-skill at 110,547 stars, earshot at 3,981, agentic-ai-prompt-research at 2,566, and cinetic at 1,016 (our github repo owner table).
2. Competitors
| Name | Where | Launched | Scale | How it differs from unlazy |
|---|---|---|---|---|
| The verification-before-completion skill from obra/superpowers | Open source | 2025-10 | Repo 313k stars; this skill 267k installs on skills.sh | The de facto standard for this slot, but only a prompt-level discipline; unlazy turns it into a formatted ledger, a checker that runs commands, and a Stop hook |
| The Ralph Wiggum loop: frankbria/ralph-claude-code, mikeyobrien/ralph-orchestrator, michaelshimeles/ralphy, open-ralph-wiggum | Open source | 2025-08–2026-01 | 9,721 / 3,192 / 2,983 / 1,908 stars | Uses an outer loop to keep the agent running until the PRD is done, with the agent judging "done" itself; unlazy fixes executable acceptance conditions before the task starts |
| code-yeongyu/lazycodex | Open source | 2026-05-25 | 3,880 stars | A full Codex orchestration framework (memory, planning, execution, verified completion), heavy; unlazy is a single skill file you install and use |
| OthmanAdi/planning-with-files | Open source | 2026-01-03 | 27,905 stars | Uses files for plans and checkpoint recovery, doesn't handle acceptance |
| uditgoenka/autoresearch | Open source | 2026-03-13 | 6,805 stars | A "change → verify → keep/discard" loop aimed at research and optimization, not a general completion gate |
| loopx-project/loopx | Open source | 2026-05-31 | 6,552 stars | A long-task control plane, platform-level |
| Anthropic's official Claude Code Stop hook mechanism + the Opus 5.5 prompting guide's "checklist reminder + list the ways an agent quits early" | Big-company features | guide about 2026-09 (reported by mindpattern 9-23) | — | Anthropic officially acknowledged the problem and offered writing advice; unlazy is an out-of-the-box packaged implementation (its Stop hook uses the official hook mechanism) |
| AI LABS revamp (changed sub-agents to run in parallel) | Other (paid community) | 2026-08-20 | Only in the AI Labs Pro paid community | Taking unlazy and turning it into a paid membership, i.e. downstream monetization |
Takeaway: "stop agents from lying about being done" was already claimed by superpowers, whose install base is 26× unlazy's. unlazy didn't open a new lane; it pulled the problem out on its own, gave it a catchy name, and added an engineered enforcement layer. Data sources: our github library, GitHub API, skills.sh, 2026-11-06.
3. Growth tactics (evidence)
- The author's own X (cold start, confirmed):
- Launch tweet (8-11): "This might be something big… I made a skill which reduces the laziness of models. Pretty good for Opus 5. go try /unlazy", 146 likes, 124 bookmarks, 21,027 views
- 8-15 "okay I tested /unlazy more with Opus 5 and I'm ngl it's really good": 633 likes, 1,062 bookmarks, 46,919 views
- 8-23 two posts, "use /unlazy to fix unlazy's own issues/PRs" and "1,000 stars": about 10k and 4 thousand views each
All the numbers above come from the fxtwitter API, read 2026-11-06. Bookmarks are about equal to or higher than likes, which suggests people are saving it to use later.
- YouTube creators (the largest wave of third-party amplification, confirmed):
- AI LABS (179k subscribers) 8-20 posted "GitHub's #1 Trending Author's New Claude Skill Is Insane", 123,508 views. The hook is the author's "#1 GitHub trending author" credential (from taste-skill). The video also points out that sub-tasks run serially by default, and puts a parallel version in its paid community.
- AICodeKing (134k subscribers) 8-26 posted a video, 7,815 views.
- There is also a YouTube Short, "Claude Says Done but Skipped 3 Steps?", view count unverified.
- Reposts by large X / LinkedIn accounts (confirmed):
- @0x3b33 (Pyro, 6k followers) 8-22 "You don't know how good this skill is until you try it": 596 likes, 1,019 bookmarks, 39,718 views
- Shubham Saboo (author of awesome-llm-apps, Unwind AI) 8-25 posted a long LinkedIn piece, "Someone just built a skill that blocks your coding agent from saying 'done' until it actually proves it"
- @DanKornas (100k followers) 9-08 reposted it, with only 1,911 views
The stretch from 8-23 to 9-03 where stars grew fastest lines up with when AI LABS, Pyro, and Saboo posted (this is an inference; no day-by-day star counts directly prove it).
- Piggybacking on the author's viral posts (confirmed): 9-05 the author posted "gave GPT 6 Astra about 5 hours to create a forest", 3,955 likes, 1,083,803 views. The follow-up prompt reply (472 likes, 686 bookmarks, 166,354 views) opens with "First search for the '/unlazy' skill, read it fully, and use it throughout the entire task". This embeds the skill inside another viral piece to sell it. The 9-13, 9-24, and 9-25 entries recorded in the card, which hit the Trendshift JavaScript chart (#16–19), came right after this.
- SEO / aggregator sites / skill directories (confirmed, the long-tail sources after September):
- Sites that wrote about it in late August: aisignal.dev (8-17; the benchmark table in the article is likely AI-fabricated), ssdnodes (English/German), theactionableai, chen.media, fervorai (two articles), starlog, claude-updates, mindstudio, elma.sh, heatdrop, mushroom.cv (Chinese)
- Later: reveneau (9-03), nathanhodgson, laurenfoxai (10-20), Medium/plainenglish (10-19), claudecodereview.ai
- Skill directories: skills.sh, skillsmp, skillsdirectory, ui-skills, openskills, skillmd, agentmods, flowy, decimal, mozgi, SourceForge-style mirrors
- daily.dev indexed two YouTube videos
Most of these are likely AI-generated content farms, and they underpin the current growth of 5–30 stars per day.
- Not found: HN (our hackernews library searches for unlazy / leonxlnx each return 0 results), Product Hunt, paid promotion, major Reddit threads. GitHub's official trending couldn't include it because the repo was under 60 days old, and it was never going to make the cut in 8 September anyway.
4. Why it took off
- It hits the most common complaint right now: In the Opus 5 / 5.5 era, agents can run long tasks for hours, but they often "deliver 80% and write a confident summary to quit." Anthropic's own Opus 5.5 prompting guide addresses this exact phenomenon (mindpattern 9-23 cites platform.claude.com). The author wrote in the "1,000 stars" post: "agents kept stopping once a task looked mostly done… apparently a lot of people had the same problem".
- The name and usage are the distribution: The slash command
/unlazyis instantly understood, and can be pasted straight into a prompt ("First search for the /unlazy skill"), which makes it naturally shareable and easy to copy into other people's prompts. - The author's credentials: AI LABS's headline used "GitHub #1 Trending Author", and taste-skill's 110k stars gave creators a reason to pick it. The same thing posted by an unknown author would likely have stalled at a few hundred stars. Compare obra/superpowers, a similar skill with a huge install base that has never been singled out as a viral moment.
- A "backed by papers" packaging: The README lists a string of 2025–2026 papers (Quantifying Laziness, METR Time Horizon, SlopCodeBench, OptimalThinkingBench, etc.). reveneau put it directly: "that list is what makes the project more than a prompt tip", and SEO sites love citing it.
- Engineering feel in v2: Starting from 8-10 it became a GATES ledger + gate-check.mjs (--status / --approve / --reverify, SHA-256 evidence fingerprints) + Stop hook + lint. Saboo's post centers on the line "moves enforcement out of the prompt and into files". That makes it feel harder than the crowd of "prompt skills" and more worth bookmarking.
5. Differentiation
Similar tools are either a discipline prompt (superpowers' verification-before-completion), an outer loop (the Ralph family, where completion still depends on the agent saying so), or a full heavy framework (lazycodex, loopx). unlazy differs in that:
- Before starting, it writes a machine-executable acceptance ledger: each gate has a CHECK command, and it only passes when the exit code is 0 and the output matches EXPECT;
- It uses Claude Code's Stop hook to stop the agent when it tries to quit;
- A supervising agent uses
--reverifyto re-check work handed back by sub-agents; - The whole thing is just a skill directory + a zero-dependency Node script, installed with
npx skills addin one line, and it works with both Claude Code and Codex.
The original selling point, Depth Tree (each leaf gets the full budget), was demoted in v2 to "a method for breaking down tasks", and the README no longer mentions it on the first screen.
This differentiation is why it gets bookmarked and written about (Saboo, fervorai, and heatdrop all talk about "handing completion off to files and scripts"), but being seen mostly came from the author's credentials plus creator amplification. And by usage alone it is far behind superpowers: 10.1k installs versus 267k on skills.sh, which shows the differentiation never turned into large-scale adoption.
6. Visibility and copycats
- Visibility: it's circulating in small circles, a bit higher than earshot. Reasons:
- There are YouTube videos with 120k views, Saboo's LinkedIn post, and dozens of SEO articles;
- It was carried along by a viral post from an author with 1.08M views;
- But it didn't make HN or Product Hunt, and there's no mainstream tech media coverage;
- Open source and free, no revenue.
- Copycats / repackaging:
- AI LABS made a parallel revamp and sold it in a paid community;
- aiskillstore/marketplace, outlinedriven-unlazy, the cyrus-pinto version on skills.rest, kart-rc/nlah-agent-repo's harness-skillpacks, and the npm package @selesai/code have all bundled or republished unlazy (Exa search results, skillsdirectory);
- Similar repos in our github library created after August with names containing lazy / completion gate haven't passed 100 stars, and no copycat has gained real traction.
7. Takeaways
- Find a flaw that every agent user complains about but nobody has named on its own ("laziness", "lying about being done"), name it with a verb-style slash command (/unlazy), so the name itself is the slogan.
- Launch first with a prompt-only version (v1, a single file) to grab attention, then a day later ship the "script + hook" engineered version, upgrading it from a "prompt trick" to a "tool".
- Cite a string of papers for credibility, giving SEO sites and creators ready-made material.
- Write your own skill into your own viral content's prompt (the forest post), so one exposure promotes two repos.
- A creator name backed by a large repo gives creators a hook for their topics ("the new skill from #1 trending author").
8. Uncertain
- There are no day-by-day star counts from 8-09 to 10-15; the intermediate points (1,000 / 2,986 / about 3,115) come from the author's posts, reveneau, and back-calculation from Trendshift, so how much each wave contributed can only be inferred.
- I did not find out whether Shubham Saboo also posted on X (where he has a larger following).
- The 10.1k installs on skills.sh only count the
npx skillschannel; manual clones aren't included, so actual usage is unknown. - harness-hub has a video, "Purr Request: unlazy — More Tokens, Worse Code?", showing that someone questions whether it "spends more tokens and produces worse code". I haven't read the content; the README also admits that the early 6 internal comparisons can't be reproduced.
- AI LABS's paid remake: how is it selling? Unverified.