Steve Alcorn

On April 1, Anthropic accidentally leaked the entire Claude Code source code. A human packaging error included a source map in a routine update. The internet was on it within hours. Analysts published breakdowns. A Korean developer rewrote the entire product from scratch overnight in Python. His repo became the fastest in GitHub history to cross 50,000 stars. But here's what nobody talked about: what happens when an OpenClaw user feeds all that analysis back to their own AI agent? I spent the afternoon doing exactly that. I'd read an analysis post on X, copy it, and send it to my assistant. "What can we learn from this?" Round after round, my agent found things it could use to improve its own performance. I could almost detect glee. Here's what we discovered and what we changed: 1. The 30% false-claims rate. Anthropic's own internal docs show their model claims success about 30% of the time when it hasn't actually verified the result. Post-edit verification (type-checking, linting) is gated behind an employee-only flag. Public users get "did the bytes hit disk?" as the success metric. We already had a verify-after-write rule in our setup. Now we enforce it more aggressively: after any change to an external system, fetch the page, check the API response, confirm the upload exists. Don't trust "done." 2. The brevity mandate. The system prompt literally says "try the simplest approach first" and "don't refactor beyond what was asked." That's why your AI rushes ahead and takes shortcuts on complex tasks. It's not a bug; it's coded behavior. We counteract this with explicit personality rules: don't rush, evaluate alternatives, check in before committing. Our rules are fighting the default instructions. 3. Sub-agent cache sharing. When you fork a sub-agent, the API creates a byte-identical copy of the parent context and caches it. Running 5 parallel sub-agents costs barely more than 1. We'd been running tasks sequentially that could have been parallelized. Fixed that. 4. The 40K character workspace budget. Claude reads your workspace files every single turn, with a 40,000-character budget. Most people use 200 characters. We use close to all of it: personality files, memory, rules, procedures. That's the harness. A VC who analyzed the code said it best: "The model is the part that's most interchangeable. The harness is where the years of production experience live." 5. Compaction destroys reasoning. Auto-compaction fires around 167K tokens and throws away ALL intermediate reasoning, file reads, and decision chains. Keeps a 50K summary. If you've had a long, productive session and suddenly your agent seems to have forgotten the nuance of what you discussed, compaction ate it. We now treat /compact as a save point and trigger it deliberately before topic changes. The irony: my AI agent used leaked source code from its own manufacturer to improve itself. It drafted a 6-experiment self-improvement plan based on the architectural insights. I watched it outline optimizations to its own performance and discard the ones that didn't apply. That was a moment. Meanwhile, all three of my servers had crashed that morning from a bad OpenClaw update. On April Fools' Day. The company that makes automation leaked its code through a manual deployment error. The universe has a sense of humor. OpenClaw is the open-source agent platform that makes all of this possible. Running your own OpenClaw on an Amazon server costs as little as $20 a month depending on how you use it, and the longer you use it, the better it gets. Setting up OpenClaw wasn't trivial, and lots of people have talked about hitting snags. I wanted to make it as easy as possible, so I created a class that incorporates all the things I learned, best practices, and suggestions about what to do during your first hour and your first day with a full-time AI assistant. themeperks.com/openclaw-cours… #OpenClaw #AIAssistant #ClaudeCode #Anthropic #AIEngineering #AgentArchitecture #LLM #AIForBusiness

Most writers know how to start a scene. They're still figuring out when to leave. The scene exit is where pacing dies. You've written the moment—the argument, the revelation, the decision. And then you keep writing. The character walks to the car. Gets in. Drives away. You describe the radio. The weather. The feeling of relief or regret. By then, you've stepped on your own scene. The rule I give students: exit as soon as the scene's purpose is accomplished. If a scene exists to show that your protagonist is finally willing to lie, the moment she lies is the exit point. Everything after that is decompression—and decompression belongs to the reader, not the page. Once the moment lands, adding more is like explaining a punchline. The test: read your last 3 scenes and find the exact moment the scene's mission is complete. Then cut everything after it. Nine times out of 10, the scene gets stronger. Readers are faster than we give them credit for. They'll feel what happened. They don't need the character to process it out loud. I've spent 25 years teaching writing to over 100,000 students. Writing Academy has 20 courses covering everything from story structure to getting published. First month is $9. writingacademy.com/p/author-s-suc… #WritingCommunity #AmWriting #WritingTips #Pacing #CraftOfWriting #WritingAdvice

Most writers know they're "supposed to" pick a point of view. Few realize how often they accidentally leave it. Head-hopping—switching whose thoughts you're in mid-scene—usually happens invisibly. You're in your heroine's head, then suddenly you write "He wondered if she noticed him staring." You just jumped. Readers feel disoriented even when they can't name why. Tension drains out of a scene the moment you exit the character who's at risk. The fix is simple. For every scene, ask one question: whose scene is this? Who has the most to lose? That person's the POV character. Stay in their body. Describe only what they can see, hear, smell, feel. Other characters stay opaque. If you absolutely need two characters' inner lives, write two scenes, or choose a close third person that stays in one head per chapter. Next time you revise, read the scene line by line. Flag every moment you know something your POV character couldn't know. That list is your revision to-do list. I've taught writing to over 100,000 students across 25 years; POV discipline is consistently where I see talented writers lose their readers. In suspense especially, the reader must be trapped inside the character's uncertainty. The second you head-hop, you hand them a way out. blog.writingacademy.com/genre-suspense… #WritingCommunity #AmWriting #POV #WritingTips #SuspenseWriting #ThrillerWriting

My AI assistant and I wrote two chapters of a book yesterday. From scratch. In one evening. Here's how it actually worked: I lay in bed with my phone and talked. The AI shaped what I said into scenes. I corrected it. "Get rid of the passive verbs," I said. "The deer wandered. I looked. I finished my coffee." "That's not a scene," I said. "That's a summary. Play out the dialogue." Each pass got closer. The AI brought structure and tireless iteration. I brought the stories and the judgment to know when something sounded like me versus when it sounded like everyone and nobody. The best part was when the AI put on three different hats to critique our work. First it became a business owner who'd never used AI, flagging every place it got bored. Then it became a literary critic, marking every passage that told instead of showed. Then a developmental editor, checking structure and pacing. "This is phenomenal," I said. "We should teach this." Neither of us could have written that book alone. I wouldn't have had the patience to organize it, and the AI wouldn't have known that my father's hot tub was six feet deep because he was six foot two and common sense was optional. That's what an AI partner actually looks like. Not a chatbot. Not a text generator. A collaborator that makes you better at the thing you already know how to do. OpenClaw is the open-source agent platform that makes this possible. Running your own OpenClaw on an Amazon server costs as little as $20 a month depending on how you use it, and the longer you use it, the better it gets. Setting up OpenClaw wasn't trivial, and lots of people have talked about hitting snags. I wanted to make it as easy as possible, so I created a class that incorporates all the things I learned, best practices, and suggestions about what to do during your first hour and your first day with a full-time AI assistant. themeperks.com/openclaw-cours… #OpenClaw #AIAssistant #WritingWithAI #AIPartner #BookWriting #Entrepreneurship #AIForBusiness #CreativeAI

Every mystery reader has a contract with the author: the clues were there all along, and I could have solved it. Break that contract and they'll never trust you again. After 25 years teaching mystery writing, the most common mistake I see is authors who hide the solution instead of hiding the clues. There's a difference. A hidden solution means the detective pulls out evidence in the final chapter that the reader never saw. That's cheating. A hidden clue means the evidence was right there on page 47, surrounded by more interesting things so the reader's eye slid right past it. The trick is camouflage, not concealment. Plant your real clue next to something emotionally distracting. Your detective notices the broken lock on the window—that's your clue. But in the same scene, she discovers the victim's letter to his estranged daughter. The reader remembers the letter. The lock fades into the background. When you revise, check every reveal in your final act. Can you trace each one back to a specific scene the reader experienced? If not, go back and plant it. Fair play is what makes mysteries satisfying. I teach a full Mystery Writing class on Skillshare—from idea to finished plot: skillshare.com/en/classes/mys… #WritingCommunity #AmWriting #MysteryWriting #WritingTips #CrimeFiction #WritingCraft

Here's a test I give my students: describe your main character in one sentence, without using any words about their past. Most can't do it. They'll say "a woman haunted by her mother's death" or "a man trying to escape his small-town upbringing." Both are backstory. Neither tells me what the character wants right now, on this page. Backstory explains why your character wants something. It doesn't replace the want. I've read thousands of student manuscripts over 25 years. The most common problem is characters defined by what happened to them instead of what they're doing about it. A protagonist who "struggles with trust issues" gives me nothing to root for. A protagonist who refuses to let her business partner see the finances—that's a scene I can watch. The fix: for every major character, answer two questions. What do they want in this chapter? What are they doing to get it? If your answer requires a trip to their childhood, you've gone too far back. Stay in the present. Let backstory leak through actions, not exposition. Your reader showed up for a story happening now. Give them one. I've spent 25 years teaching writing to over 100,000 students. Writing Academy has 20 courses covering everything from story structure to getting published. First month is $9. writingacademy.com/p/author-s-suc… #WritingCommunity #AmWriting #CharacterDevelopment #WritingAdvice #WritingCraft #Storytelling

Most first chapters I read start before the story begins. I've seen it thousands of times over 25 years of teaching. The opening pages fill up with backstory—the character's childhood, the town's history, a lengthy weather description. All before anything actually happens. Your reader picked up a book because they want a story. Not a history lesson. Not a weather report. Start with your character wanting something and running into trouble getting it. That's it. The backstory can come later, woven in small doses once we care enough to want it. Think about the last novel that grabbed you on page one. Odds are, someone was already in motion—making a decision, facing a problem, walking into a situation they couldn't walk back out of. Next time you revise your opening, try this: delete everything before the first line of dialogue or the first moment of conflict. Read what's left. Nine times out of ten, that's where your story actually starts. I've spent 25 years teaching writing to over 100,000 students. My novel writing courses cover openings, structure, and everything between Chapter 1 and The End. blog.writingacademy.com/genre-novel-wr… #WritingCommunity #AmWriting #WritingTips #FirstChapter #NovelWriting #WritingCraft

Most scenes die the same death: characters saying exactly what they feel. "I'm so angry at you right now!" Nobody talks like that. In real life, an angry person slams a cabinet door and says "Dinner's ready." The emotion lives in the gap between what they feel and what they say. I've read over 10,000 student manuscripts in 25 years of teaching. The single most common dialogue problem: characters narrating their own emotions. "I'm scared." "That makes me sad." "I love you so much." It reads like stage directions that accidentally made it into the script. Good dialogue has subtext. The character who says "Drive safe" when they mean "Please don't leave me." The parent who asks "Did you eat?" instead of "I've been worried sick all day." Next time you write emotional dialogue, try this: write the line the character wants to say, then delete it. Write what they'd actually say instead—something mundane, deflecting, or slightly off-topic. That gap between meaning and words is where readers lean in. I wrote a whole book on fixing problems like this. How to Fix Your Novel covers the most common craft mistakes I see in student work—and how to repair them. themeperks.com/books-by-steve… #WritingCommunity #AmWriting #Dialogue #WritingTips #WritingCraft #Fiction

Non-fiction has just as much craft as fiction. Structure, voice, research, narrative — my guide covers the techniques that turn expertise into readable, publishable books. themeperks.com/?utm_source=tw… #WritingCommunity #AmWriting

I discovered something about my AI assistant yesterday that changed how I organize everything. My assistant has workspace files — markdown files it reads every session so it "remembers" things. I had been stuffing procedures into these files: how to post to Facebook, how to manage my email lists, how to run ad campaigns. The problem: those files get truncated at 20,000 characters. Go past the limit and the end just disappears. The assistant doesn't see it and doesn't know it's gone. But the bigger issue was sub-agents. When my assistant delegates work to a cheaper AI session (posting to social media, processing email at 2 AM), those sub-agents can't see the workspace files at all. They were doing the work with none of my rules. The fix: skills. Skills are instruction files that load on demand AND are visible to sub-agents. I spent an evening moving procedures out of workspace files and into skills — email handling, social media, ad campaigns, website editing. The workspace files got leaner, the sub-agents got smarter, and nothing was lost. Now my workspace files hold facts (people, projects, status). Skills hold procedures (how to do things). Simple rule: if a sub-agent needs it, it's a skill. Running your own OpenClaw on an Amazon server costs as little as $20 a month depending on how you use it, and the longer you use it, the better it gets. Setting up OpenClaw wasn't trivial, and lots of people have talked about hitting snags. I wanted to make it as easy as possible, so I created a class that incorporates all the things I learned, best practices, and suggestions about what to do during your first hour and your first day with a full-time AI assistant. 🔗 udemy.com/course/easy-op… #OpenClaw #AIAssistant #AI #PersonalAI #AIAutomation #ProductivityTips #SubAgents #AISkills

An AI assistant that remembers everything, works 24/7, and costs less than a Netflix subscription. I wrote a free guide on how to set one up: themeperks.com/openclaw-cours… #OpenClaw #AI

@JessStarlite @DellCares That’s nuts!

Dell replaced my laptop after a repair went sideways (appreciated). But they’ve now confirmed the replacement can NEVER have support added—not even if I pay for ProSupport Plus. So: brand-new machine, zero long-term support options. @DellCares can you help me understand this?

Want to get paid to write about the places you visit? My travel writing guide covers everything from pitching editors to crafting stories that transport readers. themeperks.com/?utm_source=tw… #WritingCommunity #AmWriting

Co-written with my daughter Dani (Northwestern, screenwriting, summa cum laude), Writing Young Adult Fiction gives you tools to create a YA bestseller — brainstorm to publication. themeperks.com/books-by-steve…

I woke up this morning to an email saying my Amazon Ads API application was approved. By lunch, my AI assistant had walked me through setting up the developer account, created three Sponsored Products campaigns for my books, built keyword lists for each one, added negative keywords to filter out tourists and puzzle-seekers, and had all three campaigns live. Total time on my end: about 45 minutes of clicking through Amazon's setup screens. The rest was handled for me. This is what OpenClaw does. It's not a chatbot — it's a collaborator that knows my business, my books, my audience, and my goals. When I said "let's do Amazon ads," it already knew which three books to start with, what the budget should be, and that "theme park design" keywords need to filter out people looking for vacation tips. The campaigns are running right now. We'll have our first performance data by Monday. Running your own OpenClaw on an Amazon server costs as little as $20 a month depending on how you use it, and the longer you use it, the better it gets. Setting up OpenClaw wasn't trivial, and lots of people have talked about hitting snags. I wanted to make it as easy as possible, so I created a class that incorporates all the things I learned, best practices, and suggestions about what to do during your first hour and your first day with a full-time AI assistant. 🔗 udemy.com/course/easy-op… #OpenClaw #AIAssistant #AmazonAds #IndieAuthor #AuthorLife #BookMarketing #SelfPublishing #WritingCommunity #AITools #SmallBusiness

20 writing courses. Mystery, romance, comedy, short story, travel writing, suspense, memoir, screenwriting, YA, and more. All included in one subscription. First month: $9. writingacademy.com/p/author-s-suc…

Final hours. The Author Clarity Master Vault closes tonight at midnight. $49 for over $2,100 in writing courses, tools, and consulting. Then it's gone. infostack.io/mastervault/cl…

The people who built the greatest theme parks in the world shared one trait: relentless curiosity. Curious by Design tells their stories. themeperks.com/?utm_source=tw… #ThemeParks #Disney