标签：#Engineering

Why Prompts Matter More Than You Think The difference between a great AI response and a mediocre one isn't always the model. It's the prompt. Experience this: You ask ChatGPT a vague question and get a vague answer. You ask the same AI a perfectly crafted prompt and get something incredible. The skill gap is massive. Companies are paying prompt engineers $150K+ because mastering prompts directly impacts: Response quality Token usage (costs) Speed of inference User satisfaction The Science of Better Prompts Rule #1: Be Specific, Not Vague BAD : "Write me something about AI" GOOD : "Write a technical explanation of how transformer attention mechanisms work, suitable for a developer with 2 years of ML experience" Specificity reduces hallucinations and increases relevance by 10-50x. Rule #2: Use Roles & Context You are an expert senior software engineer with 15 years of experience. You specialize in system design and scalability. Respond in a way that balances technical accuracy with accessibility. Target audience: Mid-level engineers. How would you design a real-time chat system for 10 million concurrent users? Role-based prompting improves response depth and tone. Rule #3: Provide Examples (Few-Shot Prompting) Classify the sentiment of these reviews: Example 1: "This product is amazing!" → Positive Example 2: "Terrible experience, would not recommend" → Negative Example 3: "It's okay, nothing special" → Neutral Now classify: "The service was slow but the staff was friendly" Examples guide the AI toward your exact expectations. Rule #4: Break Complex Tasks Into Steps Instead of: "Analyze this code and find bugs" Use: "1. First, read through this code carefully Identify any logical errors Check for performance issues List potential security vulnerabilities Provide a summary of findings with severity levels" Step-by-step prompts (Chain-of-Thought) improve reasoning by 20-40%. Rule #5: Specify Output Format Respond in JSON format: { "summary" : "brief explanation" , "key_

A while back I was researching a topic I didn't know much about — the kind of casual, late-night "let me just ask the AI a few questions" session. A few messages in, I asked a follow-up that only made sense in the context of what we'd just been talking about. I didn't restate the subject, because... why would I? We were three messages into the same conversation. The answer came back completely off-topic. It had lost track of what "it" referred to, latched onto the wrong noun, and confidently explained something I hadn't asked about at all. Not a small tangent — a whole paragraph about the wrong thing. My first reaction was annoyance at the model. My second, more useful reaction came a bit later: I'd been treating it like a person who remembers what we were just discussing and fills in the gaps naturally. It doesn't do that the way a human conversation partner does. If I don't restate the subject, it's genuinely not there for the model — it's not being lazy, there's just nothing to work with. So I started over-specifying. Every follow-up got longer: restate the subject, restate what I actually wanted, restate the constraint I cared about. It worked, but some days I didn't have the energy for it — I'd just take the mediocre answer, say "ok thanks," and move on. Which meant I was quietly leaving useful answers on the table half the time, just because typing out the full context felt like a chore. Eventually I stopped thinking of it as "the AI being difficult" and started treating it as a simple rule: if I want it to know something, I have to say it. It won't infer the unstated stuff the way a person would , no matter how obvious it feels to me. Once that clicked, a few concrete habits followed. Restate the subject, every time Not "what about the second one" — the actual name of the thing. It costs three words and removes an entire failure mode. Say what you actually want, not just the topic "Tell me about X" and "I'm trying to decide whether X is worth the switching co

The biggest losses in software history were, with one deliberate exception, not attacks. They were silent, correlated, self-inflicted — and they teach the exact risk autonomous AI agents are about to make expensive again. At 9:30 in the morning on August 1, 2012, Knight Capital Group was one of the largest trading firms in the United States, executing a sixth of all the volume on the New York Stock Exchange. By 10:15 it was, for practical purposes, finished. In those forty-five minutes a piece of its own trading software (not a hacker's, its own) fired more than four million unwanted orders into the market, accumulating roughly $7 billion in positions the firm never meant to hold and a loss of about $440 million by the time humans understood what their machine was doing. The cause, documented in the SEC's administrative proceeding, was almost insultingly small: a deployment that updated seven of eight servers. The eighth still carried a dormant piece of code called Power Peg, retired years earlier, and the new release reused the old feature flag that woke it up. No one attacked Knight Capital. The market data was accurate, the exchange functioned perfectly, and every system reported itself healthy while the company bled ten million dollars a minute. That shape (no adversary, no alarm, one change propagating everywhere at once) turns out to be the shape of almost every entry on the list below. We've written before about the biggest bug-bounty payouts in history , the ledger of what it costs when someone does attack. This is the other ledger, the bigger one: what software has cost when nobody attacked at all. Every figure below states what it counts, and comes from a primary or authoritative source (inquiry boards, SEC filings, statutory inquiries) linked at the end. The ledger 1. CrowdStrike outage (2024) — roughly $5.4 billion in direct losses to Fortune 500 companies alone (estimate). One faulty content update to the Falcon Sensor security agent blue-screened Windo

Picture a power plant burning fuel to spin a turbine. It is tempting to assume that with enough engineering — better seals, smoother bearings, cleaner combustion — the plant could be pushed toward converting nearly all its heat into useful work. It cannot. A large modern thermal power station turns only something like 40 to 45 percent of its fuel energy into electricity, and the missing majority is not lost to sloppy design. It is lost to a law of physics. That law sets a ceiling on every device that turns heat into work, from a car engine to a steam turbine to a jet. The ceiling is called the Carnot efficiency, and the remarkable thing about it is how little it depends on. Not on the working fluid, not on the mechanism, not on the cleverness of the builder — only on two temperatures. This article explains where that limit comes from, how to compute it, and why it reshapes how engineers think about efficiency. Why this calculation matters The Carnot efficiency is the benchmark against which every real engine is judged. When an engineer reports that a gas turbine runs at 38 percent efficiency, that number means little on its own. Compared against the Carnot limit for the same hot and cold temperatures, it suddenly tells you how much room is left — whether the design is already near the physical wall or still has slack worth chasing. It also redirects design effort toward the things that actually matter. Because the Carnot limit depends only on the ratio of cold to hot absolute temperatures, the single most powerful way to raise the ceiling is to raise the temperature at which heat enters the engine, or lower the temperature at which it is rejected. This is why turbine inlet temperatures have climbed for decades, pushing the limits of metallurgy and cooling. Polishing internal friction yields small gains; raising the hot-side temperature raises the ceiling itself. The core formula Sadi Carnot, in 1824, imagined an idealized engine running on a perfectly reversible cyc

How you split your code into repositories seems like a plumbing decision, but it quietly shapes how your team collaborates, ships, and reasons about the system. A monorepo keeps everything in one repository; a polyrepo gives each service or app its own. Neither is universally right, and the loudest opinions online usually ignore your actual stage and team size. Here's how to think about it clearly. What a monorepo buys you A monorepo puts your web app, mobile app, backend, and shared libraries under one roof. The advantages are real, especially for smaller teams: Atomic changes. Update a shared type and every consumer in the same pull request. No cross-repo coordination dance. One source of truth for tooling. A single lint, format, and CI config instead of drift across a dozen repos. Effortless code sharing. Shared TypeScript packages are just imports, not published versions you have to bump and reinstall everywhere. Easy refactoring. You can find and fix every caller of a function because it's all in front of you. Tools like Turborepo and Nx make this practical by caching builds and only running work for the parts that actually changed. What a monorepo costs The trade-offs show up as you grow. Build and CI times can balloon without smart caching. Access control is coarser — it's harder to give a contractor one service without the whole codebase. And a naive setup rebuilds and tests everything on every change, which gets slow fast. Good tooling mitigates all of this, but you have to invest in it deliberately. What a polyrepo buys you Separate repositories give each service hard boundaries . A team owns its repo end to end, deploys on its own schedule, and can't accidentally reach into another team's internals. Access control is naturally granular, CI for each repo is small and fast, and the blast radius of a bad change is contained. The cost is coordination. A change that spans services becomes multiple pull requests across multiple repos that must land in the right

Microservices have a marketing problem: they're associated with the engineering cultures of Netflix and Amazon, so ambitious teams assume adopting them is what serious companies do. But those companies moved to microservices to solve problems of enormous scale and huge headcount — problems you almost certainly don't have yet. For most products, splitting too early is one of the most expensive mistakes you can make. Here's the honest trade-off. What a monolith actually gives you A monolith is one deployable application. That simplicity is a feature, not a limitation, especially early: One codebase, one deploy. No orchestration, no service mesh, no distributed tracing just to understand a request. Simple debugging. A stack trace crosses your whole request. You're not correlating logs across five services to find one bug. Fast local development. Run the whole app on your laptop and iterate. Easy transactions. Data consistency is a database transaction, not a distributed saga you have to design and get right. The modern version isn't a big ball of mud. A modular monolith enforces clean internal boundaries — separate modules with clear interfaces — giving you much of the organization of microservices with none of the network overhead. What microservices actually cost Splitting into services doesn't remove complexity; it moves it from your code into the network, where it's harder to see and reason about. You inherit a long list of new problems: Distributed systems failure modes — partial failures, retries, timeouts, and eventual consistency become your daily reality. Data consistency across services — no more easy transactions; you're designing sagas and compensating actions. Operational overhead — every service needs deployment, monitoring, logging, and on-call. Slower local development and debugging — reproducing a bug can mean running half your architecture. For a small team, this overhead can consume the very velocity you were trying to gain. When microservices genuin

I am interested in agentic coding for the same reason I care about good engineering process in general: I want work to move forward in a way that is inspectable, repeatable, and resilient once the task gets messy. A lot of AI-assisted coding still feels like improvisation. You ask for something, get a result, adjust the prompt, try again, and hope the useful reasoning is still somewhere in the scrollback. That can work for tiny edits. It gets much less convincing when the task starts touching architecture, tests, review, or pull requests. What I want instead is a workflow where the model helps me think and execute, but inside a structure I can inspect afterwards. I want artifacts, gates, and something I can resume tomorrow without reconstructing the entire mental state from memory. That is why I use po8rewq/agentic-skills . It gives me a practical way to do agentic coding as an engineering workflow rather than as a long sequence of chat turns. A task moves through requirements, architecture, implementation, checks, review, and pull request creation. Each stage leaves something I can read, verify, and challenge. What makes this interesting to me The interesting part is not just that there is a CLI. Plenty of tools have a CLI. What matters to me is that it turns AI-assisted coding into a staged system: requirements force the task to become explicit architecture makes risks visible before code is written implementation happens against a plan instead of against a vague prompt checks and review happen as part of the flow, not as an afterthought runs are resumable, so interruptions do not destroy context That changes the feel of the work quite a bit. Instead of asking "what should I prompt next?", I am usually asking "what stage is this task in, and what should exist before I move on?" Where this really clicked for me was when I noticed I was spending less energy trying to preserve context in my head and more energy evaluating actual outputs. What the repository actually

Explore how the Aspire team turns merged product changes into SME-reviewed docs pull requests, closing the gap between release and documentation. The post Automating cross-repo documentation with GitHub Agentic Workflows appeared first on The GitHub Blog .

Itamar Friedman discusses how architects and engineering leaders can break through the AI productivity ceiling using adaptive multi-agent systems. He shares insights on moving past simple autocomplete to resilient workflows by integrating autonomous testing, intelligent code review, and robust arbitration. Learn how to govern agent communication and build a context-driven SDLC that scales. By Itamar Friedman

Enterprises are now talking about AI coding agents in a very predictable way. The first question is usually: "How much more code can it help us generate?" It is not a wrong question. But if that is the only question, the ROI calculation will probably be wrong. Because enterprises are not really buying "more code." They are buying: faster delivery less rework lower maintenance cost better developer experience more stable software quality more controllable security and compliance risk faster translation from product capability to business value Code generation is an input. It is not the outcome. That distinction matters. An AI coding agent can help developers write functions, fix bugs, add tests, generate documentation, understand codebases, and refactor legacy systems. That sounds powerful. But the enterprise question is not: "How many lines of code did it generate today?" The better question is: Did that code reach production faster? Did incidents go down? Did the team spend less time on repetitive work? Did customers get value sooner? If the answer is unclear, generating 100,000 lines of code a day may simply mean producing technical debt faster. The short version: AI coding agent ROI does not end inside the IDE Many teams start measuring AI coding tools with the most obvious numbers: code suggestion acceptance rate lines of code generated number of active users number of prompts time saved on individual tasks These metrics are useful. But they mostly show that the tool is being used. They do not prove that the enterprise is getting value. Enterprise ROI has to be measured across software delivery, quality, risk, and business outcomes. In other words, an AI coding agent is not just a point solution for individual efficiency. It affects the entire software value stream: Request -> Design -> Coding -> Review -> Testing -> Deployment -> Monitoring -> Feedback -> Business outcome If you calculate value only inside the "coding" box, you miss the bigger picture. Why "amo

Most "Debezium vs managed tool" articles get the question wrong. They frame it as a product bake-off, feature grid included, and declare a winner. But if you've actually run change data capture in production, you know the real decision isn't which tool captures a transaction log better. They mostly read the same logs the same way. The real decision is who operates everything that sits around the capture, and whether that work is a good use of your team's time. That's a build-vs-buy question, not a product question. This post is a framework for answering it for your own situation. First, let's kill an outdated assumption A lot of Debezium criticism floating around is two or three years stale, and if you repeat it in 2026 you'll get corrected fast. So let's set the record straight before we compare anything. Debezium is no longer just “the thing you run with Kafka Connect.” In recent Debezium 3.x releases, the project has become much more flexible than the old tutorials suggest. Today, you have several deployment options: Kafka Connect , the classic setup, which gives you the Kafka ecosystem, distributed fault tolerance, durable schema history, and access to Kafka Connect sink connectors. Debezium Server , a standalone application that streams changes to systems like Amazon Kinesis, Google Cloud Pub/Sub, Apache Pulsar, Redis Streams, or NATS JetStream without requiring Kafka. Debezium Management Platform , which builds on Debezium Server and the Debezium Operator to provide a higher-level way to configure and manage CDC pipelines in Kubernetes-style environments. Embedded usage , where you run Debezium Engine inside your own application. Recent Debezium releases also added framework support such as the Quarkus extension. A few more things are worth knowing so the comparison is fair: Kafka 4.x runs in KRaft mode, and ZooKeeper mode has been removed. “You need to babysit ZooKeeper” is no longer true for a modern Kafka deployment. Debezium's default remains at-least-once

Software architecture debates have a problem that most other engineering disciplines don't: the alternative was never built. When a bridge fails, the failure is physical, attributable, and measurable against every other bridge that didn't. The engineering decisions that caused it can be isolated, traced, and corrected — not just in theory, but in the next bridge, because the material itself produces feedback that no amount of professional opinion can override. Steel deflects. Concrete cracks. Physics doesn't care what the architect believed. Software produces no equivalent feedback. A system built around the wrong abstractions compiles, runs, ships, and passes its tests just as readily as one built around the right ones. A bug introduced by a misaligned domain model looks identical, from the outside, to a bug introduced by a typo. A feature that took three times longer than it should have, because the structure made it harder than the business logic warranted, produces no artifact that distinguishes it from a feature that was simply difficult. The cost is real. The cause is invisible. This is the unfalsifiability problem, and it runs deeper than "we can't measure everything." It means that when a system becomes expensive to change, the diagnosis almost always lands on the wrong variable. The domain is complex. The requirements changed. The previous team was careless. Almost never: the structure was wrong, and the structure was wrong because nobody ever built the other version of it to compare against. That version doesn't exist, it never will, and every architectural argument in the industry is conducted in its absence. This would be a purely philosophical problem if there were nothing to do about it. There is something to do about it — but it requires accepting that the standard metric for software quality, whether it works, is measuring the wrong thing entirely. The Metric That Hides the Problem The natural substitute for "is this good engineering" is "does it wor

Why HDI PCB Manufacturing Starts Long Before the First Hole Is Drilled When people think about PCB manufacturing, they usually imagine drilling, plating, imaging, etching, solder mask, and surface finishing. For conventional PCBs, that assumption isn't too far from reality. For HDI (High Density Interconnect) PCBs, however, manufacturing actually begins long before any physical production starts. The success of an HDI project is often determined during engineering review rather than on the factory floor. Manufacturing Starts with Design Decisions A PCB layout may pass every design rule check inside CAD software while still being difficult to manufacture efficiently. Typical examples include: unnecessary stacked microvias excessive sequential lamination extremely aggressive trace and space dimensions unrealistic copper balancing inefficient stack-up planning None of these issues are fabrication defects. They are engineering decisions. The earlier they are identified, the lower the overall project cost becomes. The Stack-Up Is More Important Than Many Engineers Expect One of the biggest misconceptions is that increasing the layer count automatically solves routing problems. In reality, a carefully planned stack-up usually provides greater benefits than simply adding more copper layers. A good stack-up improves: signal integrity impedance consistency EMI performance power distribution thermal behavior More importantly, it creates a PCB that is easier to manufacture repeatedly with stable quality. HDI Is a Balance Between Performance and Manufacturability Many first-time HDI designs focus only on routing density. Experienced engineers usually focus on manufacturability. For example: Should this microvia really be stacked? Can staggered vias achieve the same result? Is another lamination cycle actually necessary? Can the BGA fan-out be optimized differently? Each decision influences fabrication complexity, yield, lead time, and production cost. Why DFM Matters More for H

Part 1 of "Trust the Machine" -> a series on building AI infrastructure that is secure, compliant, and governable by design. Most organizations can produce an accurate catalog of the web services they operate. Far fewer can produce an equivalent catalog of the AI systems they run — the models, fine-tunes, retrieval pipelines, agents, and third-party AI APIs now embedded throughout their products and internal tooling. This asymmetry defines the state of AI security in 2026. Adoption has outpaced oversight. Industry reporting this year has described a surge in enterprise AI activity on the order of 83% year over year, with governance and visibility lagging well behind. The consequence is a large and only partially mapped attack surface — one that many organizations cannot fully enumerate, let alone defend. Every mature security program rests on a single first principle: you cannot protect what you cannot see. Artificial intelligence is no exception. Before threat-modeling an agent or authoring a guardrail, an organization must be able to answer a deceptively difficult question: what AI is running across the environment, and who is accountable for it? This post examines how to build that answer. The rise of shadow AI Shadow IT — the unsanctioned adoption of tools outside official channels has been a recognized challenge for decades. Shadow AI is its faster-moving successor, and it appears in more forms than most inventories are designed to detect: Embedded API calls. A product team integrates a hosted model in a few lines of code and an API key, with no formal review. Copilots and assistants enabled across existing SaaS platforms, frequently activated by the vendor rather than the customer. Fine-tunes and adapters trained on internal data and stored in locations that fall outside standard scanning. Agents and automations that have incrementally acquired the ability to act—filing tickets, sending communications, initiating transactions—one permission at a time. Model de

Reading Lilian Weng's harness engineering survey as a reliability engineer — what self-improving harness papers actually show, and the three invariants every working loop converges on.

We’ve all been there: staring at a stack of printed lab results or a folder full of cryptic report_final_v2_NEW.pdf files, trying to remember if our cholesterol was higher or lower two years ago. For developers, this isn't just a filing problem—it's a data engineering challenge. In the world of healthcare, data is messy, siloed, and often locked in "unstructured" formats. To build a truly personal Electronic Health Record (EHR) system, we need more than just a folder; we need a RAG (Retrieval-Augmented Generation) pipeline that can parse PDFs, map them to the FHIR (Fast Healthcare Interoperability Resources) standard, and provide natural language insights. In this guide, we’ll leverage Unstructured.io , Milvus , and DuckDB to turn chaotic medical PDFs into a queryable, structured knowledge base. The Architecture: From Raw Pixels to Structured Insights Before we dive into the code, let’s look at how the data flows from a messy lab report to a structured answer. graph TD A[Unstructured PDF Reports] --> B[Unstructured.io Partitioning] B --> C{Data Split} C -->|Textual Context| D[Milvus Vector DB] C -->|Tabular Data| E[DuckDB Structured Storage] D --> F[LangChain RAG Engine] E --> F G[User Query: Is my glucose trending up?] --> F F --> H[FHIR-Formatted Response] Why this stack? Unstructured.io : The gold standard for handling "ugly" PDFs (tables, headers, and nested lists). Milvus : A high-performance vector database built for scale. DuckDB : Perfect for running complex analytical SQL queries on the extracted "structured" parts of our medical data. FHIR Standard : To ensure our data follows global healthcare interoperability rules. Prerequisites Make sure you have your environment ready: pip install langchain milvus unstructured[pdf] duckdb openai Step 1: Extraction with Unstructured.io Medical PDFs often contain complex tables. Standard PDF parsers usually fail here. We’ll use unstructured to partition the document into logical elements. from unstructured.partition.pdf

You asked your AI to help you plan a trip. It gave you a paragraph about packing layers and booking early. You needed a checklist, a hotel shortlist, a flight window, and a rough daily schedule. What you got was a thoughtful non-answer dressed up as advice. That gap — between what AI tells you and what it could actually do for you — is the gap agentic AI is designed to close. And most people don't know it exists. The Difference Between Answering and Acting Standard AI models are trained to respond. You send a prompt, they generate a reply. The entire interaction lives inside a single text exchange. Agentic AI operates differently. Instead of producing one answer, it takes a goal and breaks it into a sequence of steps — then executes them, one after another, checking its own output along the way. It can look things up, organize information, write to a document, revisit a step if something doesn't look right, and deliver a final result that's actually usable. The travel example makes this concrete. A conversational model tells you to pack a rain jacket. An agentic setup builds you the trip: it pulls destination weather data, generates a packing list specific to your travel dates, identifies hotels in your price range, and drops everything into a structured itinerary. Same goal. Completely different level of output. Author's note: The word "agentic" has been overloaded to the point of meaninglessness in tech marketing. For our purposes here, it means one specific thing — an AI that runs a loop: think, act, observe the result, decide the next action. If it's not doing all four of those things in sequence, it's not really an agent. It's just a chatbot with extra steps. Why This Loop Changes Everything The reason agentic AI feels qualitatively different isn't magic — it's architecture. The core mechanic comes from a framework called ReAct (short for Reasoning and Acting), introduced in a 2023 paper by Yao et al. and now foundational to most production agent systems. The l

Quick answer: The PromptEval Prompt Quality Report scored over 1,000 real prompts across 12 use cases. The average was 52 out of 100, and only 8% reached "good" (75+). The strongest single predictor of a good prompt is whether it defines its output format, worth 27 points on average. In 9 of 10 prompts, the weakest dimension was robustness. This is the PromptEval Prompt Quality Report . Over 1,000 real prompts have been scored on PromptEval , submitted by real users across use cases from customer support to healthcare to code. Each was scored from 0 to 100 on four structural dimensions: clarity, specificity, structure, and robustness. Every figure below comes from that set. No prompt text is stored; the analysis is anonymous and aggregate. Only 8% of the 1,000+ scored prompts reached "good" (75 or higher). Fewer than 1% reached "excellent." Source: PromptEval Prompt Quality Report, 2026 How the scores break down Here is how the scores spread across the set. The bar for "good" is 75, the point where a prompt is clear, specified, and holds up under variation. Score range Share of prompts 0 to 40 (failing) 25% 41 to 60 (below par) 31% 61 to 74 (functional but mediocre) 36% 75 to 84 (good) 8% 85 to 100 (excellent) under 1% Roughly 92% of prompts never reach "good," and almost none reach "excellent." This includes prompts from people who clearly know the tools. The gap is not talent. It is a few missing pieces that repeat. What separates a good prompt from a bad one For each structural element, we compared the average score of prompts that had it against those that did not. These are averages across the set, not a controlled experiment, so read them as correlation. But the gaps were large and consistent. The prompt... Avg with Avg without Point gap Defines the output format 58 31 +27 Has explicit constraints (what not to do) 63 41 +22 Assigns a role or persona 57 42 +15 Includes at least one example 64 51 +13 Prompts that define their output format score 27 points higher

A lot of AI memory systems start with the same temptation: "Just save the useful thing." That sounds harmless until the knowledge base becomes a junk drawer. Half the notes are too specific, a few are duplicates, some are obsolete, and nobody knows which ones the agent should trust. In ai-assistant-dot-files , the memory system is deliberately slower. It uses a promotion lifecycle: Capture -> Candidate -> Audit -> Approve -> Index -> Retrieve -> Expire That lifecycle is documented in docs/runbooks/memory-engineering.md , and the important word is not "capture." It is "candidate." Nothing writes directly to memory The framework has a durable memory layer: Knowledge Items in shared/knowledge/ , ADRs in docs/adrs/ , the domain dictionary, team topology, a feature archive, and a registry at shared/memory-registry.json . But a lesson from a delivery does not jump straight into shared/knowledge/ . It first becomes a Candidate Record. That record has required fields: Source Type Evidence Tags Expiration condition Then memory-engineer audits it: Is it reusable? Is it already covered? Is it too speculative? Does it belong as a Knowledge Item, or should it become a rule change, prompt edit, or ADR instead? Only after that does a human approve the destination. The design is intentionally similar to code review. Durable memory changes future behavior, so they deserve a paper trail. Rejection is a feature One of my favorite parts of the memory runbook is that it has explicit rejection rules. Do not promote a memory when it is: a one-off already covered too speculative That makes "zero candidates promoted this cycle" a healthy result, not a failure. This is where memory engineering starts to look less like note-taking and more like gardening. The point is not to preserve every leaf. The point is to keep the soil useful. Expiration matters The lifecycle also includes expiration. A Knowledge Item can become stale when the underlying code, agent, or pattern changes. It can be supers

I hesitated to write this. Not because I don’t have an opinion about AI in software engineering, but...