Prompt Alchemy Is Dead: Agent Skills Enforce Engineering Discipline on AI

Stop Handwriting Thousands of Characters of Prompts: Agent Skills That Racked Up Tens of Thousands of Stars on GitHub Declare the End of Alchemy

Intro: From "Prompt Alchemy" to "Engineering Rationale"

Over the past two years, the entire tech community has fallen into a feverish "prompt alchemy."

Everyone's clipboard holds several system prompts thousands of characters long, stuffed with dense rules:

"You are a world-class senior architect... You must strictly follow Clean Code principles... Never omit code... Please think step by step deeply before outputting... If you perform well I'll give you a $200 tip..."

We tried to tame a black box of massive parameter count with these almost occult natural-language "incantations."

But the real engineering results are often brutally cruel: the longer the prompt, the more scattered the model's attention; instructions that worked in isolation get frequently diluted in a large context. Writing prompts feels more and more like praying to a deity, with luck outweighing science.

In the last few weeks, GitHub's trending list has undergone a landmark upheaval: obra/superpowers, started by open-source veteran Jesse Vincent, swept the web, while cloud giants like Cloudflare, NVIDIA, and AWS have all jumped in to open-source their official Agent Skills.

This collective shift sends an unmistakable signal: the "prompt engineering" built on natural-language piling is going bankrupt at an accelerating pace, while classic "software engineering discipline" is making a full comeback in the form of Agent Skills (agent skill-ification).

One: The Limits of Prompts: Verbal Admonition Is Never Institutional Guarantee

Why can't manually writing thousands of lines of prompts fundamentally fix the problems of AI randomly changing code or breaking logical flow?

Because in system engineering, "verbal admonition" is the least effective management method.

It's like hiring a genius newbie who has read extensively but lacks experience. You nag a thousand times at his ear, "Remember to write unit tests, don't break the existing interfaces, don't casually modify public config," yet once faced with tens of thousands of lines of real business code, he will still instantly mutilate the core logic.

Traditional prompt interaction has three fatal flaws:

Attention Dilution: stuffing thousands of characters of rules into the model causes it to gradually forget early constraints during the long forward pass;
Lack of State Machine and Deterministic Blocking: prompts cannot force steps. You tell the model "write tests first, then implementation," but it often writes both in one response, never verifying whether the test actually catches bugs;
Fragile Unmaintainability: prompts are fragile loose text. A single fine-tuning of the underlying model, or slight tweaks to the fine-tuning parameters, can silently invalidate previously effective prompts, making it impossible to bring them under serious team version control.

Two: The Essence of Agent Skills: Putting a "Deterministic Rein" on a Wild Horse

What exactly did skill-based frameworks like superpowers get right?

The core shift: it no longer tries to "teach" the model how to write code, but builds a "track that it cannot leave."

1. Enforce Hardcore Engineering Methodology (like TDD)

In superpowers' skill library, the most central capability is the test-driven development (TDD) specification. It's not an empty "please write tests first," but turns engineering discipline into a deterministic action contract:

First step: force the agent to write a failing test that precisely reproduces the bug (red light);
Execution gate: before the test actually runs and fails, the system's tool permissions are directly locked, forbidding any modification to production business code;
Second step: only allow writing the minimal code needed to turn the test green;
Third step: automatically run the full regression suite to ensure zero breakage of the existing chain.

Under this pipeline, the AI is no longer an improvisational speaker running wild, but a precise executor locked onto an industrial assembly line.

2. On-Demand Context Mounting: Goodbye to "Whole-Family Pollutant"

In the traditional mode, to make the AI understand the rules, we often had to feed a huge amount of architecture specs, design patterns, and lint rules at the beginning of the conversation.

Agent Skills adopt an extremely restrained dynamic mounting mechanism: normally, the core context is very clean, containing only the most basic task goal; only when the agent decides "we now need to enter security vulnerability investigation" does the system dynamically mount the corresponding security-audit.skill; once the audit finishes, the related context is immediately unmounted.

Load when needed, discard after use. This not only saves up to 70% of expensive tokens for long-horizon tasks, but also keeps the core reasoning chain at a high signal-to-noise ratio.

3. Sedimenting Alchemy into "Codified Assets"

A prompt is prose for the model to read, while a Skill is a structured engineering asset (standardized checklists, trigger conditions, verification hooks, helper scripts).

This means a team can manage every Agent Skill the way it manages production code: Git review, diff comparison, continuous integration, and canary releases. The determinism of software engineering has finally taken back control over AI's uncertainty.

Three: An Echo from History: Ancient Engineering Discipline Has Become the Lifeline of the LLM Era

Computer science has developed for over half a century. Those classic laws humans painstakingly summed up — test-driven, single responsibility, state isolation, defensive programming — were originally invented not to please computers, but to guard against human programmers' own carelessness and chaos.

What's remarkably dramatic is that, fifty years later, when we face a large model that computes tens of thousands of times faster but is also more prone to hallucination, the most effective weapon to tame it is still this ancient engineering wisdom.

Large models don't need you to use rhetoric in a prompt to teach it how to write Python; it has already memorized all the open-source code in the world.

What it lacks is not intelligence, but discipline.

From "handwriting thousands of characters of prompts and praying for luck" to "using standardized Agent Skills to build engineering pipelines," this trend unfolding on GitHub is not merely an iteration of developer tools; it is the coming-of-age ceremony of AI programming truly maturing.

💡 Follow [一缕82年的清风], gain insight into the technical bottom line and ecosystem evolution Feel free to discuss, like, and share in the comments.