Anthropic posted the announcement at 2 AM UTC and it was everywhere by breakfast: starting July 20, Claude Fable 5 will be "included in all Max and Team Premium plans, at 50% of limits." Pro and Team Standard users stay on usage credits and get a one-time $100 credit. Twelve hours in, the post sits at 32,175 likes and 6.2 million views.
The upside needs no explanation. Fable 5 is the strongest coding model Anthropic ships, and from Sunday it comes with the plan you already pay for. The catch is the second half of the sentence: half limits, for the hungriest model you can point at a repository. The reply threads are already full of developers doing the math on what an agentic loop does to a half-size window.
That math has an answer you can measure in installs. Developers have been voting on it all week.
We ranked this stack 8 days ago. Here is what happened since.
When GPT-5.6 launched on July 9 and developers immediately hit their usage limits, we published a ranking of context-efficiency skills: the tooling that stretches a fixed usage window by cutting wasted tokens. Eight days later, that same stack has added 20,864 installs. The growth is lopsided, and the lopsidedness is the story.
| Skill | Jul 10 | Jul 18 | Change |
|---|---|---|---|
| subagent-driven-development | 140,798 | 149,562 | +8,764 (+6%) |
| improve (shadcn) | 20,365 | 24,497 | +4,132 (+20%) |
| ponytail | 12,544 | 16,937 | +4,393 (+35%) |
| context-engineering | 9,762 | 12,621 | +2,859 (+29%) |
| context-map | 9,357 | 9,406 | +49 (0%) |
| remembering-conversations | 8,891 | 8,974 | +83 (+1%) |
| claude-md-improver | 6,960 | 7,544 | +584 (+8%) |
To be clear about causality: none of these installs happened because of last night's announcement. They happened because GPT-5.6 taught everyone the same lesson in week one, and the lesson generalizes. Frontier models eat limits. The announcement just extended that lesson to every Max subscriber at once.
Look at where the growth concentrated. The three fastest risers (ponytail +35%, context-engineering +29%, improve +20%) all shape what the model does before it spends tokens: force the minimal solution, curate what the agent sees, produce a plan another agent can execute without extra context. The flat rows (context-map, remembering-conversations) accumulate more context into the session. Developers under limit pressure are choosing fewer, better turns over richer context. That is a preference you can read straight off the install counts.
The limit-stretching stack, re-ranked for July 20
Ranked by installs from the live catalog, refreshed daily from skills.sh, GitHub and MCP registries, ranked by real installs. Every count is from today.
1. subagent-driven-development (149,562 installs) (GitHub)
Executes an implementation plan by dispatching a fresh subagent per task, with diffs and reports kept in files instead of session history. The main context stays small while the work happens elsewhere. On half limits, context isolation is the single biggest lever you control.
2. improve (24,497 installs) (GitHub)
Surveys a codebase like a senior advisor and produces prioritized, self-contained implementation plans that other agents can execute without additional context. Self-contained is the key word: a plan that needs no follow-up questions is a plan that burns no follow-up tokens.
3. ponytail (16,937 installs) (GitHub)
Forces the simplest, shortest, most minimal solution at every step and strips unnecessary code, dependencies and features. The fastest-growing skill on this list, up 35% in eight days. Minimal output is cheap output, and a half-limit window turns every extra generated line into a real cost.
4. context-engineering (12,621 installs) (GitHub)
Addy Osmani's skill for deliberately curating what an agent sees, when it sees it, and how it is structured. The discipline of deciding what enters the window, rather than letting the session accumulate.
5. context-map (9,406 installs) (GitHub)
Generates a precise map of relevant files, dependencies, tests and patterns before any code gets touched. Flat this week, but still the right first move on unfamiliar code: a wrong-file exploration loop is the most expensive way to spend a premium model.
6. remembering-conversations (8,974 installs) (GitHub)
Recalls relevant past sessions when you hit a similar feature or a recurring bug, so you stop re-paying for context the model already had once.
7. claude-md-improver (7,544 installs) (GitHub)
Anthropic's own skill for auditing and improving CLAUDE.md files. Your CLAUDE.md is a fixed tax on every single session; this is the one-time cleanup that lowers it permanently.
8. NEW: strategic-compact (6,172 installs) (GitHub)
The new entrant since our July 10 ranking. Suggests manual context compaction at logical workflow boundaries instead of letting auto-compaction fire mid-task. Compaction timing sounds trivial until an auto-compact wipes the plan your agent was halfway through executing.
Two takeaways
1. Half limits on a stronger model is a trade you win by cutting retries, and the install data shows developers already optimizing for it. Plan-first and minimal-output skills grew 20 to 35% in eight days. Passive context tooling stayed flat. The community is converging on a workflow shape: decide more before the model runs, generate less when it does.
2. On July 20 the usage-limit conversation stops being an edge-case complaint and becomes the default condition of using the best model. Every Max subscriber gets Fable 5, and every Max subscriber gets it at 50% of limits. The developers who treat the window as a budget to engineer get roughly twice the model-hours out of the same subscription.
The full ranked shortlist, updated daily, is on skillselion.com, and the install-ranked leaderboard is at skillselion.com/leaderboard.
Disclosure: I run Skillselion, a directory of Claude Code, Codex and Cursor extensions. Install counts come from our live catalog, refreshed daily from skills.sh, GitHub and MCP registries, ranked by real installs.
Top comments (0)