https://github.com/esaldgut/ai-native-engineering-workspace
AI-native engineering workspace — 42 Claude Code agent skills, platform-base workflow docs, and a freshness system that re-verifies each pattern against vendor docs.
https://github.com/esaldgut/ai-native-engineering-workspace
agent-skills ai-native aws claude-code kotlin mcp swift workflow-automation
Last synced: 15 days ago
JSON representation
AI-native engineering workspace — 42 Claude Code agent skills, platform-base workflow docs, and a freshness system that re-verifies each pattern against vendor docs.
- Host: GitHub
- URL: https://github.com/esaldgut/ai-native-engineering-workspace
- Owner: esaldgut
- License: mit
- Created: 2026-06-05T02:04:03.000Z (19 days ago)
- Default Branch: develop
- Last Pushed: 2026-06-05T03:20:43.000Z (19 days ago)
- Last Synced: 2026-06-05T04:15:49.594Z (19 days ago)
- Topics: agent-skills, ai-native, aws, claude-code, kotlin, mcp, swift, workflow-automation
- Size: 345 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- Contributing: CONTRIBUTING.md
- License: LICENSE
- Roadmap: ROADMAP.md
Awesome Lists containing this project
README
# AI-Native Engineering Workspace
A public library of **42 reusable Claude Code Agent Skills**, the **platform-base workflow docs**
they extend, and a **freshness system** that re-verifies each pattern against the official docs it
cites.
Each skill is a generic extraction from production engineering, rewritten from scratch (not
copy-pasted) and verified against vendor documentation: Apple Developer, AWS, Android, Anthropic.
Every skill records its own provenance: which sources it was checked against, when, at which
version, and when to re-check it.
Scope: working patterns and conventions across Apple (Swift 6 / iOS 26), AWS (Lambda Go / CDK),
Android (Kotlin / Compose / KMP), and Claude Code / MCP. Not a tutorial collection — these are
the patterns an AI-native engineer applies, published for reuse.
---
## What's inside
| Area | Count | What it covers |
|------|------:|----------------|
| **`global-skills/apple/`** | 13 | SwiftUI iOS 26: Liquid Glass, adaptive layouts, native UX, chat / feed / marketplace / social-graph UI; Clean Architecture + error patterns + feature scaffolding; on-device AI (Foundation Models); an append-only Apple anti-pattern registry |
| **`global-skills/apple-auth/`** | 6 | Post-quantum CryptoKit, auth security audit + checklist, performance benchmarks, Swift Testing patterns — provider-agnostic (Cognito / Firebase / Auth0 as examples) |
| **`global-skills/aws-go/`** | 8 | AWS SDK Go v2 version policy, canonical error handling, Lambda cold-start segregation, refactor blast-radius audit, MongoDB TTL correctness, CDK drift detection, adversarial pre-build audit, provider-API property verification |
| **`global-skills/android/`** | 5 | Compose Clean Architecture + feature scaffolding, JUnit5/Turbine/MockK testing, security checklist (DataStore + Tink + Keystore), Kotlin Multiplatform extraction |
| **`global-skills/claude-code-workflow/`** | 6 | Project bootstrap, lesson-capture pipeline, build-sync pipeline, MCP orchestration + the Android-specific twins |
| **`global-skills/writing/`** | 1 | A canon for technical writing in Spanish (no spanglish, no voseo, technical terms kept in English) with a detection pass |
| **`global-skills/meta/`** | 3 | Skills *about* skills: extraction methodology, freshness audit, dossier-driven update |
| **`workflow-docs/`** | 13 | The generic platform base (Apple, Android, Expo, AWS, Next.js, MCP, Chrome, TypeScript, Shell, Kali, + the platform premise). Skills extend these. |
| **`reference-projects/`** | 2 | Sanitized architecture docs for a native iOS app and a native Android app — Clean Architecture, tech-research decisions, iOS 26 UX patterns |
**Total: 42 skills + 13 base docs**, roughly 14,000 lines of documented engineering.
---
## The idea: platform base + extensions
A project should never start from an ambiguous context. The flow this repo encodes:
```
workflow-docs/ (generic platform base) ← context loads here first
│ a project's CLAUDE.md @imports the relevant base docs
▼
project CLAUDE.md + docs/_EXTENSIONS.md ← project-specific detail lives here
```
The base docs are generic and reusable across any project of the same kind. A specific project
**extends** them — it never edits the base to inject its own bundle IDs, account IDs, or
architecture. The `project-init` skill automates this bootstrap; `PLATFORM_BASE.md` states the
full premise.
---
## The freshness system (why this stays trustworthy)
Technical patterns for fast-moving platforms rot in months. During this repo's own research,
five patterns were already stale or mythical at the moment of verification:
| Pattern | Naive assumption | Verified reality |
|---|---|---|
| CDK drift | "use `aws cloudformation detect-stack-drift` by hand" | `cdk drift` command (CDK CLI 2.1017.0, May 2025) |
| Android secrets | `EncryptedSharedPreferences` | deprecated → DataStore + Tink + Keystore |
| Hilt + Compose | `hilt-navigation-compose` | renamed to `hilt-lifecycle-viewmodel-compose` (Hilt 1.3.0) |
| Coroutine tests | `runBlockingTest` | deprecated → `runTest` (kotlinx-coroutines-test 1.6+) |
| Constant-time compare | `CryptoKit.timingSafeEqual` | **never existed** — use `HMAC.isValidAuthenticationCode` |
So every skill carries a `freshness` block in its frontmatter: cited sources with versions, a
verification date, a re-check trigger, and a status. The **`skill-pattern-freshness-audit`** skill
reads those blocks, re-verifies each cited API against current docs, and sets each status to
`current`, `needs-recheck`, `stale`, or `superseded` with the citation that proves it. The
**`dossier-driven-skill-update`** skill repairs what the audit flags. The contract is in
[`global-skills/FRESHNESS_SPEC.md`](global-skills/FRESHNESS_SPEC.md).
The detect-then-repair loop makes drift a queryable signal rather than something a reader finds in
production.
---
## What keeps this library dynamic: `lessongate`
A skills library that's a snapshot rots. This one is fed by an agent.
[**`lessongate`**](https://github.com/esaldgut/lessongate) is a local Go runtime agent that watches a
private project's merged pull requests, asks Claude which of the lessons captured from them are
**generalizable** (a reusable pattern, not business-specific), sanitizes each through an NDA gate,
and opens a **draft pull request** to this repository for human review. The library upstreams from
production engineering instead of being hand-curated once.
The relationship is the point: one repo is the **agent** that produces knowledge; the other is the
**library** it feeds. That separation is also what makes the flow demonstrable — an agent that opens
PRs to a sibling repo, not a repo editing itself.
The agent is built to a production bar, and its design choices are the same ones this library
documents:
- The confidentiality control is a **deterministic gate** (deny-list + structural regex + a template
allow-list, tested against a golden corpus with a seeded canary), not an LLM. The Claude verify
pass is extra recall only — it can downgrade *safe → unsafe*, never the reverse.
- Sanitization runs **before** anything leaves the process; the agent consumes curated lesson text,
never raw diffs.
- The dependency stack was **verified against live sources**, not assumed — which overturned several
defaults (e.g. `go-github` v88's `NewClient` now returns an error; Claude Opus 4.8 is
adaptive-thinking-only, so determinism comes from forced structured output, not a temperature knob).
- Built across six TDD phases; the security-critical path runs deterministically offline in CI.
It's the working proof of the premise this repo states: that AI-native engineering means building
the tooling that keeps your knowledge current, not just writing the knowledge down.
---
## How a skill is structured
Every `SKILL.md` follows the same shape — a focused, single-job Agent Skill:
```yaml
---
name: swift-liquid-glass-design-system-ios26
description: >-
Apply iOS 26 Liquid Glass to custom SwiftUI views the canonical way … Use when building
or restyling SwiftUI surfaces that target iOS 26+ …
version: "1.0.0"
freshness:
verified_against:
- source: "Apple Developer — View.glassEffect(_:in:)"
url: "https://developer.apple.com/documentation/swiftui/view/glasseffect(_:in:)"
version: "iOS 26.0"
verified_on: "2026-06-03"
recheck_after: { trigger: "WWDC26 keynote", or_date: "2026-12-01" }
decay_risk: medium
status: current
---
```
…followed by: when to invoke (with trigger signals), a table of **verified** APIs, the
load-bearing rules with code, a canonical example using generic names (`MyApp`,
`com.example.app`), a decision aid (when *not* to use it), cross-links, sources, and a
"Last verified" footer.
[`global-skills/apple/swift-liquid-glass-design-system-ios26/`](global-skills/apple/swift-liquid-glass-design-system-ios26/SKILL.md)
is the reference example for the format.
---
## Quality bar
Each skill meets every one of these:
- **Verified, not assumed.** Every API in a "canonical APIs" table was confirmed against the
vendor's live docs. That verification caught more than a dozen errors during authoring: wrong
version numbers, renamed artifacts, a nonexistent crypto symbol (`CryptoKit.timingSafeEqual`),
and an API URL that returned 404.
- **At least two cited sources**, one of them a primary vendor or standards doc, each with a
resolving URL and a version.
- **Anti-sterilization.** Every body names at least one concrete public API, which anchors the
pattern and gives the freshness audit a literal symbol to re-check.
- **One skill, one job.** No mega-skills.
- **Honest about canon.** Where a domain is still emerging (Claude Code / MCP), skills mark which
patterns are documented-canonical versus opinionated convention.
- **Zero confidential context.** No client or employer identifiers anywhere; these are generic
patterns, not an extract of any private codebase.
---
## Using these skills
Drop a skill folder into a project's (or your user-level) `.claude/skills/` directory. Claude Code
auto-invokes it when its `description` matches the situation, or you can invoke it explicitly with
`/`. The `freshness` and `version` keys are inert to the loader (it reads only `name`
and `description`) — they exist for humans and for the audit tooling here.
To check whether a skill has drifted before you rely on it:
```
/skill-pattern-freshness-audit # e.g. apple, aws-go, android
```
---
## Layout
```
ai-native-engineering-workspace/
├── README.md ← you are here
├── LICENSE ← MIT
├── workflow-docs/ ← 13 generic platform-base docs + PLATFORM_BASE.md
├── global-skills/
│ ├── FRESHNESS_SPEC.md ← the provenance contract every skill follows
│ ├── meta/ ← 3 skills about skills (extract / audit / update)
│ ├── apple/ ← 13
│ ├── apple-auth/ ← 6
│ ├── aws-go/ ← 8
│ ├── android/ ← 5
│ ├── claude-code-workflow/ ← 6
│ └── writing/ ← 1
└── reference-projects/ ← sanitized iOS + Android architecture docs
```
---
## Notes
- **Bilingual.** This README and the skills are in English; some platform-base workflow docs keep
their original Spanish body. Each skill's `description` (the part Claude reads to decide whether
to fire) is English.
- **Process.** The repository follows Git Flow with branch protection and a Conventional Commits
history — see [`CONTRIBUTING.md`](CONTRIBUTING.md).
- **License:** MIT — reuse freely.
- **Provenance dates** throughout reflect when each pattern was last verified; run the freshness
audit to re-confirm against current docs.