GitHub AI Daily · 2026-10-19 (Monday)

Picks (5)

1. MadsLorentzen/ai-job-search — the agent's first white-collar workflow hit

Repo: https://github.com/MadsLorentzen/ai-job-search ｜ +2,232 stars in a day, 53.1k total

Not a product but a forkable Claude Code workflow: read the job description → score it against your experience → tailor the résumé → draft a cover letter → prepare an interview question bank. An entire hiring funnel split into reproducible agent steps that the user owns after forking.

2. android/skills — a platform vendor starts defining "skills" itself

Repo: https://github.com/android/skills ｜ Kotlin, ★8.8k, +375 in a day

Google's Android team packaged platform development capability into Agent Skills that Claude Code, Codex and Cursor can call. Community skill libraries answer "does it exist"; an official one answers "is it safe to use in production".

3. Aperivue/medsci-skills — medical skill libraries go from broad to deep

Repo: https://github.com/Aperivue/medsci-skills ｜ ★344, updated 10-17

One job only: turn the path from research question to submission into executable skills — literature search, study design and reporting standards, citation checking, statistical analysis, figures, submission materials — compatible with Claude Code, Codex, Cursor and Copilot.

4. xinxuxin/keystone-bench — clinical evaluation asks a different question

Repo: https://github.com/xinxuxin/keystone-bench ｜ ★808

Instead of "is this clinical answer correct", it asks "when the evidence changes, does your action change": give the same patient scenario with evidence of varying quality and direction, then watch whether the recommended workup, prescription or referral moves accordingly.

5. thetahealth/mirobody — the unified-platform route for health data is validated

Repo: https://github.com/thetahealth/mirobody ｜ ★1.4k, updated 10-18

A self-hosted AI health data engine that ingests lab reports, wearables and genetic data into traceable answers, with a built-in MCP server, Docker deployment plus your own LLM key, or a fully local llama.cpp run.

Trends

1. Official skill supply appears. android/skills is the first time a platform vendor defines agent skills from its own repository. Community skills solve "does it exist"; official skills solve "dare we ship it".
2. Agent engineering pivots to reliability. The infrastructure list — task-state-guard (state reconciliation), cayu (crash recovery and approvals), maka (append-only audit), astrid (capability sandbox) — all answer one question: what happens when an agent dies halfway through. That is the debt every demo must pay before production.
3. Medical AI evaluation enters an "evidence sensitivity" stage: from correct (HealthBench) to explainable (auditable) to "does your action track the evidence" (keystone). The evaluation axis is converging on how clinical decisions actually work.
4. Nutrition projects converge on local-first plus judgement. fud-ai, nutritrace and others are zero-telemetry and self-hosted, expanding from calorie logging to additive scoring, allergen detection and supplement evidence. Privacy is no longer a selling point but an entry ticket.
5. Noise worth ignoring. The nutrition and healthcare-ai topics now carry many low-quality SEO repositories — the same pubmed-mcp-server copied several times in a day, template sites like NourishAI-Forge. Look at commit history and issue quality, not star counts.