I almost tweeted the Fable 5 launch post yesterday afternoon. Karpathy called it a “major-version-bump-deserving step change forward.” Alex Albert, Boris Cherny, Thariq, Garry Tan — all in. The first two stories on Hacker News were Fable derivatives. The timeline was begging me to do it.

Then I opened the GitHub monthly leaderboard.

mattpocock/skills added 56k stars in a month and cracked the top five.¹ Leonxlnx/taste-skill (a skill that teaches your AI taste) and hardikpandya/stop-slop (a skill that removes AI tells from prose) hit the same monthly chart the same month.² ³ Not one of them has “Fable” in the name. None of them are riding the model hype. They all do the same thing: make what the AI writes faster to review.

That cold-showered me. It also finished a thought I hadn’t closed.

The question I was asking wrong

The post I almost wrote was titled “Should you switch to Fable 5?” The question itself is broken.

The hidden assumption in “should I switch models” is: a smarter model means I get more done. But the real shift in Fable 5 is not “the model is smarter.” It’s “the model can run longer.” Longer tasks, more files, cross-session, twenty-to-thirty-minute runs.

Long-horizon agents don’t die because they can’t write. They die because you can’t review what they wrote. A model that can run for 20 minutes hands you 50 PRs a day. Each one is fine. Stacked together, you can’t keep up. By Friday’s merge, three have bugs, two went off-piste, one finished a thing you forgot to ask for.

Switching to Fable 5 doesn’t fix that. Installing a review loop does.

This sits on the same vine as yesterday’s anthropic-three-fronts-june-2026 post⁴ — that one argued strategically that “Anthropic is grabbing the harness layer.” Today’s post is the tactical half: where is your harness.

The contrarian core: what Fable 5 actually changed is review, not prompting

Three layers, walking up.

Layer 1: The Claude Code author himself says Fable 5 changed verify loops, not prompts

Boris Cherny (Claude Code’s main author) posted on June 10 with 8,000+ interactions.⁵ The headline is about Fable 5, but the sentence in the middle is the one worth re-reading:

“Fable 5 is the biggest upgrade since Opus 4.5 … what changed is self-verification loops for long-horizon agents.”

self-verification loops. Not my framing — the actual phrase, posted by the person who wrote the tool, in 8,000 interactions worth of attention.

Swyx, the same day, gave the practical version:⁶

“In the window before Fable goes usage-priced, running review my code for issues in Claude Code is alpha — you’ll scare yourself with what you’ve shipped.”

Boris tells you why. Swyx tells you how to use it. Both point at the same place: the actual new thing Fable 5 surfaces is “the model can review its own output.” The moment you can’t review, the moment you let the model review is the moment you actually catch the upside.

Layer 2: Today’s GitHub monthly top three are all review accelerators

If Boris is just an internal Anthropic voice, the leaderboard is community consensus:

Repo What it does Monthly stars Plain English
mattpocock/skills A real engineer’s .claude directory +56k Feed the model skills so its code lands “right”
Leonxlnx/taste-skill Teaches your AI taste +24k Stop the model writing greasy, generic slop
hardikpandya/stop-slop An AI-tell removal skill +6.4k Make the model’s writing pass for human

The names sound like “style / taste / de-AI.” What they all do is shorten your review loop on a PR.

  • taste-skill installed means the model picks names, defaults, and API shapes with a human voice. You don’t have to stop every line and ask “why this?”
  • stop-slop installed means the model’s PR descriptions and docs skip “moreover / furthermore / it’s important to note.” You don’t have to re-humanize the prose after every read.
  • mattpocock/skills installed means the model knows you’re a real engineer, skips tutorial-grade code, and stops writing things like “first, let’s create a function that …” You don’t have to grade it like an intern during review.

These three skills don’t make the AI smarter. They make your review bandwidth go further.

Layer 3: HN’s front page is telling you this is the real problem

The night Fable 5 went viral, the top of Hacker News was almost all Fable-adjacent. Story #1 (165 points when I read it; 205 by the time I wrote this):⁷

“Cybersecurity researchers aren’t happy about the guardrails on Anthropic’s Fable”

Story #2: the 30-day data retention policy, 170 points, pointing at Anthropic’s official support page:⁸

“Anthropic requires 30 day data retention for Fable and Mythos”

And sitting at #28, 285 points, “Anthropic’s model naming, extrapolated"⁹ — a snark piece whose comment-section consensus was “the actual Fable 5 problem is that nobody can explain what it is vs Mythos 5.”

Get the pattern? HN’s front page isn’t arguing about how strong the model is. It’s arguing about what happens around the model. Guardrails, data retention, naming — these are all “things that show up after the model runs in longer loops.” Fable 5 isn’t a new model. It’s a new scale. And the review loop is what breaks first.

Where the three review-loop skills go, and what changes after

No code tutorial. Just where to install, and what your review flow looks like afterwards.

1. mattpocock/skills → ~/.claude/skills/ for Claude Code

mattpocock/skills is not another prompt collection. It’s Matt Pocock’s actual .claude directory, public.¹⁰ Drop the skills/ folder into your own ~/.claude/skills/, restart Claude Code. From then on, the model follows Matt’s TypeScript habits, testing habits, review habits.

What this actually changes for your review flow: PR descriptions start saying things like “this changes the public API, call out the breaking change” instead of “This change adds a new feature …” boilerplate. That alone halves the time you spend scanning for “is the model being lazy.”

2. taste-skill → your global system prompt

taste-skill’s README opens with: “gives your AI good taste. stops the AI from generating boring, generic slop.“¹¹

Concretely: with it loaded, when you ask the model to design an API, name a variable, or write a comment, it actively avoids filler adjectives, “untitled-1”-style defaults, and “a powerful, flexible, and extensible …” stacks.

Real impact on review: you stop breaking flow every 30 seconds to ask “is this name too generic.” Your review pace picks up.

3. stop-slop → anything where the model writes for humans

stop-slop targets the seven AI tells: 3-in-a-row parallel structures, em-dash abuse, filler words, “It is important to note” sentences, over-summarization, and two more.¹²

If you, like me, let the model write PR descriptions, technical docs, and changelogs, this one shaves 20% off your review time on its own — because you stop having to re-polish every doc the model hands you.

The actual contrarian take: I suggest you hold off on Fable 5

If you haven’t installed the three skills above yet, or if you installed them halfway and never tuned them, Fable 5 genuinely doesn’t matter for you.

The math is simple:

  • Fable 5 is ~30% stronger than Sonnet 4. Your review load goes up ~30%.
  • If you couldn’t review 5 PRs/day on Sonnet 4, you still can’t review 6.5 PRs/day on Fable 5. Absolute output didn’t move.
  • The three skills above, installed, cut your per-PR review time by 20%-40% — that upside is available on Sonnet 4 today.

One sentence to summarize the call:

Don’t ask “should I switch to Fable 5.” Ask “after I switch to Fable 5, how many PRs a day can I actually review?” If you can’t answer, it’s not time.

Three things to do tonight

  1. 5 minutes: Clone mattpocock/skills, copy the skills/ directory into ~/.claude/skills/, restart Claude Code, and check whether your next PR description includes “calling out breaking change”-style framing. If yes, the skill loaded.
  2. 10 minutes: Read Leonxlnx/taste-skill’s README, paste its recommended taste prompt into your global system prompt. Ask the model to design an API. Watch whether it avoids “powerful, flexible, extensible.”
  3. 20 minutes: Read hardikpandya/stop-slop’s “7 AI tells.” Once you’ve read them, you will start seeing them in every AI-written doc you open from tomorrow. You will not be able to unsee them.

Whether to upgrade to Fable 5 depends on whether you’ve done these three. Without them, Fable 5 just gives you a smarter intern whose output you still can’t review. With them, Sonnet 4 already buys you 1.5 extra mergeable PRs a day.

Closing with the line I keep coming back to — Aaron Levie, same week, same HN cycle:¹³

“There’s no amount of intelligence that can get packed into AI models that replaces the need for context.”

In review terms: no model is smart enough to skip the review. The only question is “5 minutes” or “30 minutes.”

Your call.

References

  1. mattpocock/skills on GitHub — https://github.com/mattpocock/skills
  2. Leonxlnx/taste-skill on GitHub — https://github.com/Leonxlnx/taste-skill
  3. hardikpandya/stop-slop on GitHub — https://github.com/hardikpandya/stop-slop
  4. Anthropic’s three-front night (2026-06-10, prior post) — https://cuigh.com/posts/anthropic-three-fronts-june-2026/
  5. Boris Cherny, “Fable 5 is the biggest upgrade since Opus 4.5 … self-verification loops” — https://x.com/bcherny/status/2064431111154053187
  6. Swyx, “In the Fable pre-usage-pricing window, running review my code for issues is alpha” — https://x.com/swyx/status/2064492823781789969
  7. HN #1: “Cybersecurity researchers aren’t happy about the guardrails on Anthropic’s Fable” — https://news.ycombinator.com/item?id=48478969
  8. Anthropic’s 30-day data retention policy (official support page) — https://support.claude.com/en/articles/15425996-data-retention-practices-for-mythos-class-models
  9. HN #28: “Anthropic’s model naming, extrapolated” — https://news.ycombinator.com/item?id=48480852
  10. mattpocock/skills README (raw) — https://raw.githubusercontent.com/mattpocock/skills/main/README.md
  11. Leonxlnx/taste-skill README (raw) — https://raw.githubusercontent.com/Leonxlnx/taste-skill/main/README.md
  12. hardikpandya/stop-slop README (raw) — https://raw.githubusercontent.com/hardikpandya/stop-slop/main/README.md
  13. Aaron Levie, “Context > intelligence” — https://x.com/levie/status/2064186766907887941