Hacker News, distilled — No. 14.04.26
Weekly
Volume I

HN & Co.

Cover story — No. 01

Anthropic ships Routines for Claude Code — the next layer above skills, and the thing you’ve been trying to hack together for months.

01
AI Tools Creative Software Privacy Dev Tools Weird Science

Contents
ten stories.

For the curious,
the builder,
the suspicious.
No. 14.04.26 — p.02
Section — AI Tools
The Opening Directly Applies To You
01

Claude Code
Routines.

The CLI you use every day just grew a new verb — reusable, chainable, and finally worth building against.

Anthropic shipped routines to Claude Code: named, reusable sequences of prompts and tool calls that turn ad-hoc workflows into primitives. If your skills folder is already a graveyard of half-ideas, this is the abstraction above it. Read the thread for the gotchas — especially around arg passing and nested invocation.

claude.com 316 pts 199 comments
Submitted by matthieu_bl Read on HN
No. 14.04.26 — p.04
Section — Creative Software
The Dark Horse FREE · FULL
02

Blackmagic comes for Lightroom with DaVinci Resolve, Photo.

Resolve already owns indie color grading. Now the same non-subscription, professional-grade discipline arrives for stills — and the HN thread reads like a pent-up sigh of relief.

1,033
Points — today’s chart-topper
260
Comments
$0
License fee
Read
● ALERT — Surveillance Opt-Out Filed ● Dossier No. OPT-OUT-001 ● Response Time: Unknown ● Jurisdictions Affected: 4,000+ ● ALERT — Surveillance Opt-Out Filed ● Dossier No. OPT-OUT-001 ● Response Time: Unknown ● Jurisdictions Affected: 4,000+
No. 14.04.26 — p.06
Section — Privacy
OPT OUT
Case File 03 — Privacy

I wrote to Flock’s privacy contact to opt out of their domestic spying program.

One person tried to exercise their data rights against Flock Safety’s nationwide license-plate surveillance network. The answer — or the silence — is a case study in how opt-out mechanisms fail when the business model depends on you being in.

Read the full dossier
~/projects/hn-daily — jj shell — 80×24
04/10
$ hn --story 04 --category "dev-tools"
> fetching story… ok
> rendering…

jj — the CLI
for Jujutsu.

Steve Klabnik’s guide to jj, the version-control front end that’s been quietly eating Git mindshare. 473 points · 406 comments — today’s most-discussed thread. The usual split: half “Git is fine”, half “I switched and I’m never going back.” The interesting part is the second half sounds less theoretical every week.

$ jj log --stat
  points.........473
  comments.......406
  submitter......tigerlily
  source.........steveklabnik.github.io
$ hn --open
EXEC: OPEN ON HN
The HN Times
No. 14.04.26 · p.10
Monday, April 14, 2026
Story 05 of 10 Dateline: Mountain View
After a decade of abuse —

Google moves against “back-button hijacking.”

Google has published a new spam policy targeting sites that trap users in back-button redirect loops — one of the oldest and most widely tolerated dark patterns on the open web. The policy update, which landed with 807 points and 457 comments, signals that behavior long considered “just how SEO works” will now be treated as manipulative spam. Any property with aggressive interstitials, redirect chains, or forced history manipulation should read the policy page in full. The comments are where the real teeth are: engineers describing specific patterns they expect to see penalized, and SEO consultants quietly updating their playbooks in real time. This is the rare Google policy change that actually helps users.
807 points, 457 comments. The web has been begging for this for a decade.”
807 pts 457 comments
Continued on hn — Read the full story
HN Daily · Vol. I, No. 14.04.26 Research Brief — p.12
§ 06 — Weird Science / Research

Introspective Diffusion Language Models.

via introspective-diffusion.github.io · submitted by zagwdt · 215 points · 41 comments
— Abstract, as read by your editor

What if language models didn’t generate one token at a time? This paper applies diffusion-model thinking to text: refine an entire passage iteratively instead of autoregressing left-to-right. It’s early, speculative research — but the architecture is fundamentally different from every production LLM, and it’s the kind of paper that might look prescient in eighteen months. File this under “non-obvious” and revisit.

— Keywords diffusion non-autoregressive iterative refinement text generation architecture
cf. HN thread, 14 April 2026.
Read on HN
§ 07 — Actionable Google presents —
a low-code Chrome trick.

Turn your best prompts into one-click Chrome tools.

Save a prompt as an instant Chrome action — no extension, no API key, no plumbing. Spend five minutes checking which of your daily workflows compress into a single click.

blog.google 69 pts 36 comments by xnx
Try it
Formal methods promised —

Proved
correct.

A theorem prover signed off. Every lemma checked. Every invariant held. No warnings.

… and yet —

Bug.

Lean proved this program correct. Then the author found a bug anyway.

A fascinating post on where formal verification breaks: the gap between “the proof is correct” and “the program works”. It’s exactly where assumptions, specifications, and real-world I/O go to die. 375 points, 167 comments.

kirancodes.me by bumbledraven
Read on HN
§ 09 — Drawing No. MN-001 · Rev A Relevant To Your Agent Work
MODEL A MODEL B MODEL C MODEL D … N TOOL 1 TOOL 2 TOOL 3 TOOL 4 … M edge case M tools N models
Fig. 09 — the combinatorial problem
M tools × N models = chaos.

The M×N problem of tool calling and open-source models.

As tool calling becomes standard across LLMs, the lack of a universal interface means every model–tool pair is a special case. If you’re building agent workflows with MCP servers and custom skills, this post articulates the exact friction you’re already navigating — and it’s a worthwhile map of where the standardization battles are about to happen.

thetypicalset.com 120 pts 41 comments by remilouf
Read on HN
The Finale — No. 10 — and now, the closing dispatch.

PUBLIC PUBLIC

customer files · fiverr · searchable

A Tell HN post reveals every file delivered through Fiverr’s system was left indexable — contracts, logos, source code, all searchable.

A masterclass in how “uploaded to the cloud” and “secured in the cloud” are very different claims. If you’ve ever delivered work through Fiverr — or used it to receive work — assume it’s in the wild and search accordingly.

Tell HN (self-post)
209 points · 30 comments
submitted by morpheuskafka
Read on HN