The Apple HIG, compressed to a third so agents can actually read it
Coding agents guess at Apple's interface rules because the full HIG won't fit in context — this skill distills all 157 references to a third of their size and adds a guided workflow for the folding iPhone Duo.

What it does
This independent, Apple-unaffiliated project packages the Human Interface Guidelines as an agent skill: 157 source-reviewed references spanning iOS, iPadOS, macOS, tvOS, visionOS and watchOS, built from Apple’s September pages. Ask an interface question and the agent routes to the relevant references, quoting the operative clause — conditions, recommendation strength and exact specs intact — before applying it. A /duo workflow takes an existing iOS app through fixed stages (assess, match, decide, build, verify) to adapt it to the folding iPhone Duo, editing your actual SwiftUI or UIKit code rather than sketching mockups. An opt-in enforce mode adds two independent reviewers that score an interface against the HIG while the agent repairs what they find, for up to three cycles.
The interesting bit The unglamorous part is where the value sits: Apple’s 261,621-word guidance rewritten down to 90,027 words — 65.6% shorter — while keeping exact values, conditions, how firmly each rule is phrased, API names and scoped exceptions, with every topic passing independent source review. It’s a context-budget problem solved by careful rewriting and review, and the README is candid that the ~75% compression target lost to fidelity.
Key highlights
- Covers all six Apple platforms, including the new iPhone Duo page on displays, poses, reserved regions and side-mounted bars.
- Answers cite the operative source clause before applying it — the agent quotes the rule rather than paraphrasing it into mush.
/duokeeps aduo-notes.mdin your project so later sessions can resume, and reports which checks it couldn’t run.- Enforce mode passes only at 90/100 with complete evidence and no unresolved material finding; the diagnostic approach is informed by shadcn/ui lint.
- The full corpus sits around 163k tokens (o200k_base), with the 16 foundation files at roughly 27k — about 17% lighter than the June release.
Caveats
- Token counts are static file sizes, not measured per-request loads; source review doesn’t establish runtime behavior or native-device certification.
- Compression stopped at 65.6% against a ~75% goal — short, rule-dense topics only got 57–68% shorter.
- Enforce mode needs subagents, Python 3.10+ and a way to observe the real artifact; automatic skill discovery hasn’t been tested in every host.
Verdict If you build Apple-platform interfaces with an agent in the loop — especially with a folding iPhone landing on the support matrix this fall — this is worth a close look. Everyone else can admire the compression discipline from a respectful distance.
Frequently asked
- What is justinwetch/HIGAgentSkills?
- Coding agents guess at Apple's interface rules because the full HIG won't fit in context — this skill distills all 157 references to a third of their size and adds a guided workflow for the folding iPhone Duo.
- Is HIGAgentSkills open source?
- Yes — justinwetch/HIGAgentSkills is open source, released under the MIT license.
- What language is HIGAgentSkills written in?
- justinwetch/HIGAgentSkills is primarily written in Python.
- How popular is HIGAgentSkills?
- justinwetch/HIGAgentSkills has 841 stars on GitHub and is currently cooling off.
- Where can I find HIGAgentSkills?
- justinwetch/HIGAgentSkills is on GitHub at https://github.com/justinwetch/HIGAgentSkills.