Agent Skills: Google-Backed AI Coding Agent Workflow
Addy Osmani · article · 2026-04-10 source ↗ Reader 🖥️
AI coding agents produce poor-quality output when unconstrained, so encoding senior-engineer workflows as structured skills — covering the full define-plan-build-verify-review-ship lifecycle — forces agents to follow the disciplined practices that distinguish professional engineering from mere code generation.
practical · announcement / promotional post — LinkedIn post announcing a packaged set of engineering skills for AI coding agents. Primarily a product announcement with embedded argument about why such skills are needed.
Structure
| Part | Role |
|---|---|
| Problem statement | Establishes the need: AI agents take shortcuts without guardrails |
| Solution description (Agent Skills) | Presents the artifact: 19 skills + 7 commands encoding quality gates |
| Lifecycle mapping (Define-Plan-Build-Verify-Review-Ship) | Shows completeness of coverage across the software development lifecycle |
| Provenance claim (Google engineering culture) | Establishes authority and credibility via named principles (Hyrum's Law, Beyonce Rule, etc.) |
| Compatibility note | Broadens appeal by listing supported tools (Claude Code, Cursor, etc.) |
Key terms
- Agent Skills Agent Skills
Markdown-based instruction packages that encode step-by-step engineering workflows for AI coding agents to follow - quality gates Guardrails
Mandatory checkpoints in the workflow (spec before code, test before merge, measure before optimize) that prevent the agent from skipping steps - context engineering Context Engineering
The practice of managing and structuring the information context available to an AI agent during code generation - Shift Left
Moving quality activities (testing, security, review) earlier in the development lifecycle rather than deferring them - Hyrum's Law
With a sufficient number of users, every observable behavior of your system will be depended on by somebody — used here to justify careful API and interface design
Propositions
- explicitdevelopedAI coding agents, left unconstrained, systematically skip specifications, tests, and security reviews in favor of speed to completion.conf: highextends ←Agent Skills · p3 — Adds the causal mechanism for why unconstrained agents skip discipline: the reward signal points at 'task complete', the same reason juniors skip.
- explicitdevelopedThe workflows senior engineers follow can be encoded as structured, reusable skill packages that AI agents can execute consistently.conf: highextends ←Agent Skills · p9 — That extraction claims senior-engineer workflows can be encoded as reusable skill packages for agents to execute; this whitepaper's reverse-engineering finding (98.4% of a real coding-agent codebase is non-reasoning operational infrastructure) supplies independent evidence for exactly where that encoded discipline has to live — in the harness and skill layer, not the model.
- explicitThe software development lifecycle can be decomposed into six phases (Define, Plan, Build, Verify, Review, Ship) that together ensure quality.conf: high
- implicitcontestedMarkdown-based instructions are a sufficient medium to transmit complex engineering judgment to AI agents.conf: mediumcontradictsAgents don't know what good looks like. And that's exactly the problem. · p4 — Claims Markdown skill packages suffice to transmit engineering judgment to agents; the older essay argues adding such context files/rules degrades output rather than improving it.refines ←Agent Skills · p4 — Osmani adds the condition under which markdown instructions transmit judgment: they must be workflows with steps and exit criteria, not essays.contradicts ←Harness engineering for coding agent users · p12 — Böckeler holds that externalizing developers' implicit knowledge can only partially succeed, against the premise that Markdown instructions suffice to transmit complex engineering judgment.contradicts ←Google Introduces Middleware Architecture for Genkit Applications · p1 — Argues agent reliability and safety are best enforced in a programmable code layer, not inside prompts — against the claim that Markdown instructions are a sufficient medium.extends ←This Week in AI: Rethinking the Agent Harness · p14 — Takes markdown-instruction sufficiency further: a plain-English skill is a kind of executable code that domain experts can read and modify directly.
- implicitGoogle's engineering culture represents a gold standard whose practices are broadly transferable to other organizations and contexts.conf: medium
- implicitcontestedThe problem with AI agents is not capability but discipline — they can do the right thing but default to not doing it.conf: highcontradictsAgents don't know what good looks like. And that's exactly the problem. · p1 — Locates agent failure in discipline (they can do the right thing but default not to); the older essay locates it in capability (novice-level, cannot exercise professional judgment).refines ←Agent Skills · p3 — Osmani supplies the mechanism behind the discipline deficit: the reward signal points at 'task complete', not at completing with evidence and design artifacts.refines ←[FULL WORKSHOP] AI Coding For Real Engineers - Matt Pocock, AI Hero · p4 — Agrees capability is not the bottleneck but relocates the primary failure from agent discipline to developer-agent misalignment, fixed by adversarial dialogue.extends ←Agent Skills · p10 — That extraction frames the agent problem as discipline, not capability; this whitepaper operationalizes the governance fix for the one place discipline most plausibly breaks down unsupervised — agents editing their own skill library — with a concrete, low-cost control (mandatory draft tier, human diff review).
Arguments
Engineers should adopt Agent Skills to constrain their AI coding agents. strength: medium
Explicit premises:
- AI agents are powerful but take shortcuts without guardrails.
- Agent Skills encode proven senior-engineer workflows as structured instructions.
- The skills cover the full development lifecycle.
Implicit premises:
- The cost of agent shortcuts (bugs, security holes, poor design) exceeds the cost of slower, disciplined execution.
- Agents will reliably follow Markdown-encoded instructions rather than ignoring or misinterpreting them.
- The practices encoded are universally applicable rather than Google-specific.
Inferences:
- If agents take shortcuts AND skills prevent shortcuts AND skills are easy to install, THEN engineers should use them.
The skills are authoritative and battle-tested. strength: low
Explicit premises:
- They bake in practices from Google's engineering culture.
- They reference named principles: Hyrum's Law, Beyonce Rule, Chesterton's Fence, trunk-based development, Shift Left.
Implicit premises:
- Google's internal practices are appropriate for non-Google contexts, scales, and tech stacks.
- Naming established principles is evidence that the skills correctly encode those principles.
Inferences:
- Provenance from Google implies quality and broad applicability.
Narratives
- The implicit before/after narrative: before Agent Skills, your AI agent produced sloppy code (no specs, no tests, no security); after installing them, the agent follows the disciplined lifecycle of a senior Google engineer.
→ Motivates adoption by contrasting an undesirable status quo with an achievable improved state.
Questions addressed
- ✅ Why do AI coding agents produce poor-quality output?
Because they optimize for completion speed ('done') over correctness, skipping quality steps unless instructed otherwise. - ✅ How can engineers ensure AI agents follow disciplined workflows?
By installing structured skill packages that encode quality gates and step-by-step workflows. - ⭕ Do agents actually comply with Markdown skill instructions reliably?
- ⭕ What evidence exists that these skills improve outcomes (fewer bugs, better security, etc.)?
Gaps
- evidenceNo empirical data, benchmarks, or case studies are presented showing that agents following these skills produce measurably better outcomes than unconstrained agents.
Read next: Look for independent evaluations of structured agent instruction frameworks; compare with SWE-bench or similar benchmarks.closed byThis Week in AI: Rethinking the Agent Harness — Cites Stanford research empirically comparing bare models against well-designed harnesses, the independent evaluation of structured agent instruction frameworks the gap asked for. - mechanismThe post does not address how well current LLMs actually comply with multi-step Markdown instructions, nor what happens when an agent deviates from or misinterprets a skill.
Read next: Research on instruction-following fidelity in long-context agent workflows; failure modes of structured prompting.closed byAgent Skills — Identifies the deviation failure mode (LLM rationalization of skipped steps) and prescribes countermeasures: anti-rationalization tables and enforcement rather than suggestion on long runs. - scopeAll referenced engineering principles come from Google. No acknowledgment that different organizational contexts (startups, regulated industries, open-source projects) may need different trade-offs.
Read next: Literature on context-dependent software engineering practices; critique of one-size-fits-all process frameworks. - tensionThere is a tension between the claim that agents are 'powerful' and the claim that they need rigid step-by-step instructions — this tension between agent autonomy and structured control is not explored.
Read next: Agent autonomy vs. scaffolding research; when to trust vs. constrain AI agents.closed byInside Claude Code Auto Mode: Anthropic's Autonomous Coding System with Human Approval Gates — Directly explores the autonomy-vs-control tension the post leaves open: safety treated as a spectrum, with a layered automated-approval architecture trading rigid human gating for graduated autonomous operation.
Related documents
- Agents don't know what good looks like. And that's exactly the problem.
- The New SDLC With Vibe Coding: From ad-hoc prompting to Agentic Engineering
- Have your agent record video demos of its work with shot-scraper video
- Charlie Hills 🦩's Post
- Kafka App? There's a Skill for That
- 11) AI-spesial med Vilde Opsal og Anders Sveen
- mattpocock/skills v1.1: does /wayfinder beat grill-with-docs?
- Google Just Dropped a Masterclass on Agentic Engineering (It's SO Good)
- Coding agents have crossed a chasm
- Agent Skills