Posts
All the articles I've posted.
-
Precedents and Draft Slips: How to Turn Daily AI Decisions into a Simple Nod
Working with an AI assistant all day shouldn't drain your attention on tiny, repeated choices. Borrow common-law precedent and Ming dynasty draft slips so your AI logs past decisions and brings a recommendation every time it asks.
-
Why AI Writing Always Sounds Like a Machine: The Fix That Got It Right on the First Pass
Fed the exact same source material to write a video voiceover, the first AI draft was so defensive and preachy I rejected the whole batch. Changing how I briefed the task and switching writing models got it approved on the first pass in ninety seconds. Here is what changed—and four moves you can try right away.
-
Make Yourself a 'Human System Card': Have AI Read Your Personality Before It Talks to You
Have Claude read your past chat logs and write a 'human system card' covering your personality, decision logic, and communication preferences. Includes an excerpt of my own card and a four-step prompt to make yours.
- Updated:
AI Micro-Notes 2026: Thoughts Too Short to Trash
Short AI hot takes from 2026 onwards, accumulated from Threads and IG. Model roasts, dev pitfalls, industry observations, tool impressions — each no more than three lines.
- Updated:
AI Micro-Notes 2026: Chronological Archive
The more scattered, time-sensitive AI micro-notes from 2026, in chronological order. Curated notes (by theme) live on the main page.
-
Gemini Is Underrated Because Its Default Harness Sucks: My $140 Big Three Setup
A lot of people complain that Gemini isn't good, but Google's clunky web UI and native terminal harness are what drag it down. Once you hook it up to a proper interface via API, it easily outclasses other models in writing and multimodal work. Here is how I divide the workload across a $140-a-month Big Three setup.
-
Cutting Rules to Save Tokens: One Week of AI Harness Ablation
Most people keep piling instructions onto their AI, but removing rules is what actually saves tokens. Here are the one-week results of my ablation experiment and how Anthropic recommended the exact same thing.
-
A Weekly Checklist Run by AI Only Ever Grows, So I Added a Subtraction Step
When one person or a small team has an AI run routine work from a checklist, the list only grows. Why that happens, and the monthly subtraction review I added.
-
Turning Years of Teaching into a Knowledge Base AI Can Look Up: Catching AI's Fake Quotes with Code, and Settling My Own Contradictions
I gave AI over two hundred transcripts of recorded student interviews and had it build a teaching knowledge base it looks up when preparing lessons. The AI said every quote was exact; a small program found 6 of 44 did not match. This post covers how I caught the errors, how I settle it when my own advice contradicts itself, and a starter prompt you can paste straight to an AI.
-
My AI's Memory Files Kept Growing, So I Deleted 101 and Added a Gate
After a few months with Claude Code, the memory files it wrote for itself had piled up past four hundred, mostly running logs and client status. Here is how I sorted them, cut them to just over three hundred, and put a write gate in front to stop the regrowth.
-
Turning Lesson Transcripts into a Teaching Spec: One AI Drafts, Another Finds Faults, I Only Decide
My one-on-one lessons ran 51% over schedule. I went back to the lesson transcripts, had one AI write up my teaching rules and another hunt for faults, and only made decisions on a single checklist myself. The method, the numbers, what I actually told the AI, and a prompt you can paste.
-
GPT-6.1 Terra Hands-on: Fast and Cheap, but Glyphs and Chat Context Still Go to Sonnet
OpenAI just shipped GPT-6.1 Terra and says it matches Astra at a tenth of the price. I ran review, patching and screen tasks to see what you can hand it.
-
I Built an Exam From My Own Chat Logs and Retake It With Every New Model
I turned the work I hand to AI every day into a private exam, then used it over three weeks on Claude at different effort levels, Gemini, Codex and open models. How the questions are made, how each round runs, how scores become routing rules, and prompts you can copy.
-
How a Teacher Gives Spoken Instructions to AI: Notes From 30 Classes and 751 Commands
I went through 30 transcripts of one-on-one and company classes plus 751 commands I typed to AI in my own work, to see how I actually open a request, steer midway and close it. Includes 8 original lines and a starter prompt you can paste to an AI.
- Updated:
Having Claude Dig Through a Month of My AI Chats: Habits, Mods, and Auto Mode Rules
I asked Claude Code to sift through over 300 chat logs from the past month to extract four collaboration habits, turn my most frequent catchphrases into switchable modes, and refine auto-review rules using real rejection cases.
-
Switching Your AI's Working Mode with One Command: Building work-mode and a Week of Testing
I built work-mode, a Claude Code plugin with autonomous, delegate and wrap-up modes plus desktop/phone reply formats. The pain points, how it evolved, numbers from 136 sessions, and prompts you can copy.
-
The Inaugural Codex Defensive Writing Contest
Codex often slips into defensive mode, obsessing over not being wrong instead of figuring out what you actually need. Rather than arguing with it directly, let another model orchestrate the work, or assign it to adversarial reviews where its pedantry actually shines.
-
Measuring Again Three Months Later: From AI Adoption L2.5 to L3
In July I rated myself L2.5 on Boris's AI adoption table. Three months later I measured again: Claude can reach my knowledge base and old chats, code goes through Codex review first, schedules run everything, and 430 sessions in two weeks were started by code.
-
Burning and Saving AI Quota: How I Audited, Allocated, and Spent My Compute
An audit of seven AI quota pools: burning expiring credits on self-verifiable tasks while using surplus compute to build long-term token savings. Combines a terminal dashboard and routing scripts to dispatch tasks dynamically and refine rules from logs. Includes complete tables and reusable prompts.
-
The Sweet Spot for AI Video: Be on Camera Yourself, Hand Editing and Transcripts to AI
Record your own narration on camera and let AI help with the visuals; Astra plus DaVinci Resolve did the most tedious 90% of a six-day vlog edit; Qwen 3 ASR is enough for transcripts.
-
The Wall That Was On by Default: Codex Sandbox, ENV_SCRUB, Vercel Protection, Official Connectors
An office NAS that won't mount, AUTOMODE that won't come back, outside visitors getting a 401, a cloud schedule that can't send mail. None of them were bugs in my code; each was a protection switched on by default.
-
Cursing at It One in Eight Times: Yelling at AI vs. Cheering It On
Looking back at a month of chats with AI, I cursed at it one out of every eight times. But models are stateless and don't hold grudges—someone even moved the Riemann Hypothesis forward just by cheering one on. Instead of counting tokens on politeness, talk to agents the way you naturally speak to humans.
-
I Had a Second AI Blind-Solve My Entire Question Bank
My question bank is AI-generated, so I did not let the same AI check it. A second AI re-solved every sampled question without seeing the answer key, and people judged whatever did not match.
-
I Asked My AI "Now What?" 41 Times in Two Weeks
Letting an AI run long jobs in the background is easy. Knowing where it got to when you come back is not. I counted two weeks of my own messages and found I was missing a status board, not more monitoring.
-
Which Claude Code Plan to Buy: Pro, Max 5x or Max 20x
Who should buy Pro, who should buy Max 5x, whether Max 20x is worth it, and when to just use the API. Prices come from Anthropic's official pages; the usage judgment comes from running Claude Code every day.
-
A Rule Being Written Doesn't Mean It Got Read
Writing a rule or skill into a file does not mean the model opened it this turn. Two lesson-prep misses this week, plus magpie's lessons-file approach, show why the call site has to say "open and read this first."
-
The Dumb-User Benchmark: Why Lab Scores and Real Experience Feel Worlds Apart
Why can the gap between a model's benchmark score and actual developer feel be so staggering? Breaking down the "dumb-user benchmark" across three divides, the contrast between cleanroom tests and messy reality, and what defines a god-tier coding assistant.
-
Read the System Card Before Managing a Model: Personalities and Playbooks for Opus, Sonnet, Haiku, and Fable
Read the personality evaluations in the system card before using a new model so you can manage it like a teammate. A breakdown of Opus, Sonnet, Haiku 5.5, and Fable 5.2—their workplace personas, dispatch logic, and concrete communication playbooks.
-
Two Claude Mods I Built: a Recording Mask and a Warm-Cache Plugin
I built two Claude Mods: one that masks sensitive info while I record courses, and one that keeps the Claude Code prompt cache warm. Then I found Claude Code already has idle compact built in.
-
How I use herdr: three workspaces for standing, today and long-term
After Opus 5.5 analyzed 30 days of my session logs, I split herdr into standing, today and long-term workspaces. Four status lights keep the standing agent aware, and Opus can jump into the next pane to rescue a task.
-
Model Vendors Take Turns Stumbling: Why Teams Need Built-In Redundancy
Commercial AI vendors take turns dropping the ball—OpenAI slashes speed and quotas, while Anthropic wrestles with outages and capacity crunches. Building reliable agent workflows means planning multi-model failovers and tuning out reflexive defensive tone.
- Updated:
My Pre-Trip Remote Checklist: Leaving the Mac Mini as Home Base
Eight things I checked before flying out, to keep a home Mac Mini reachable as the base for Claude and Codex: Tailscale, Herdr, caffeinate, auto-reboot, and remote file access.
-
Can Haiku 5.5 Take Over Sonnet's Work? I Tested It on My Own Tasks
Instead of official benchmarks, I built a test from 761 sub-agent jobs I actually delegated over 45 days. On seven checking tasks Haiku 5.5 matched Sonnet and Opus at about a tenth of the cost; long transcript extraction is still not complete enough.
-
159 Confirmation Prompts in 14 Days, and About 70% Were False Alarms
I audited 14 days of PreToolUse hook ask prompts: 159 of them, roughly 70% false positives. A guard hook should match the command position, not the whole string, and you can measure the fix from hook events in the transcript without replaying anything.
-
Keys Leak Where You Feel Safe
Three near-leaks of API keys in one week, and none came from pasting one somewhere public: skill packaging, a masking command that failed, and a build config. Plus the rule I use to decide whether to rotate.
-
Reviewing a City Government's AI Rollout: My Four Suggestions
I was invited to evaluate a city government's AI rollout. Here is what I saw and the four suggestions I gave: people own accountability, order by risk, fix the surroundings, judge quota by efficiency.
-
Are AI Courses Just a Cash Grab: How I Rebuilt a Solid Teaching Workflow
Many assume that because AI answers questions for free, teaching an AI course is just a cash grab. After 15 years teaching the GMAT, my first AI classes still ran hours over. Here's how I overhauled the workflow so students get a practical system that actually works.
-
Why GPT 6 Sol / Luna Feel So Dumb
A small job, transcribing consulting videos: the smart model builds a pipeline, the dumb one does every step by hand. After Opus 5.5 took over Codex's mess, it found four failure modes in 6 Sol / Luna, and why Cost per Task now needs a time variable.
-
Opus Commands, Sonnet Implements, Fable Advises: How I Split the Work Across Three Models
With Opus 5.5, Sonnet 5.5 and Fable 5.1 all out, my split is: Opus leads, verifies and deploys, Sonnet implements as a subagent, and Fable gets called in now and then as an advisor.
- Updated:
The Economics of AI Usage: Quota Black Boxes, Subscription Math, and Calming Token Anxiety
A breakdown of AI subscription economics: from 50x official subsidies and five-hour versus weekly quota gaps to prompt caching, the night-market steak framework, and automated off-peak scheduling.
- Updated:
Do Not Grade AI by Its Own Summary
I hit the same failure in several forms this week: an agent denied changing files, a commit message overstated the diff, and a transcript reassigned a speaker ID halfway through. A later batch added checks that never ran and still reported clean.
-
Auditing Workflows After a Model Upgrade: How to Rethink and Trim Your AI Skills
When a smarter model lands, how should you prune the bloated skills written to babysit older ones? Through video editing and cross-language translation tests, here's how to run zero-skill baseline tests, cut obsolete guardrails, and turn core rules into clean specs.
- Updated:
Most of the Gains From AI Adoption Have Little to Do with AI
In practice, most of what an organization gains from adopting AI comes from fixing permissions, data governance, buy-in, and bloated processes along the way, not the AI itself.
-
Running 5+ Claude Code/Codex at Once: My Five Fixes for the Attention Bottleneck
Once I have more than 5 Claude Code/Codex windows going, I hit an attention bottleneck — constant context switching is genuinely exhausting. Five things that get me past it: a good multi-window manager, voice input, an overnight schedule, a reviewable skill, and regular pruning.
- Updated:
Cheap Models for Grunt Work, Flagship Models for Review: A Durable Division of AI Labor
Offload time-consuming browser scraping and bulk data entry to budget models, while saving flagship models for spec review. From pulling bank statements and hunting award seats to tagging customer profiles, here is a durable framework for dividing AI labor.
-
Before Career Coaching: Have AI Grill You First
Before meeting with a career coach or making a major move, have AI grill your updates into a focused cheat sheet. Let your coach target your blind spots during the session, then use AI to turn those takeaways into concrete next steps.
-
Where JEV Fits: Four Ways I Use It and Three Criteria
JEV is a new model that only does probability judgments and formatted output, fast and cheap. I swapped it into the semantic gates I used to run on luna and flash-lite, plus harness slimming, news filtering, and exam-question QA.
- Updated:
Letting an Agent Tune a Local Video Model Overnight: My Three Gates
Three things I learned from letting an agent run local video-model parameter research overnight: hard constraints in CLAUDE.md, a gate script watching SSD writes, and a 30-minute check-in loop.
- Updated:
Zero Findings: Nothing Wrong, or the Checker Isn't Checking
Four times this week a "check passed" turned out to be a broken checker, plus four real bugs where every layer looked fine on its own. Before you trust a zero, prove the checker is actually comparing something.
- Updated:
When to Call In Fable: Three Moments, and Don't Turn On ultracode
The most expensive model is not meant to run the whole way. Three moments where calling in Fable feels right to me, plus why I cap Opus 5 at med, and the weekly cycle I fell into: efficiency king Monday to Thursday, Fable mode Friday to Sunday.
- Updated:
The Skills and Hooks I Use Every Day, Now Open Source
I open-sourced the Claude Code skills and hooks I actually use every day. No technical skill required, pure logic, built for knowledge workers with no coding background.
- Updated:
GPT-image-2 Wants Fewer Constraints — Plus a Consistency Drill
The "ugly crayon doodle" prompt that went viral overseas works because over-constraint kills GPT-image-2's creativity. Here is what I observed this week — including why routing prompts through Claude makes things worse, and why "consistency" matters more than raw image quality at work.
- Updated:
Why Obsidian is something special to me, and why I only made this video now
I'd been wanting to make this video for a long time, but kept putting it off. Obsidian's flexibility is so high that it grows into a different shape in every person's hands. The payoff isn't immediate either—it's only after the vault has grown for a while that you realize you can't go back.
- Updated:
Your Own Machine Is the Attack Surface: Leftover Credentials, Permission Flags, Untrusted Input
Everyone talks about AI security as whether the model will say the wrong thing. What actually bites you is the plaintext keys left on your disk, the permission flag you waved through, and the untrusted input you fed it.
-
How to Write a Gate That Actually Blocks
The gate was running and stopped nothing. Four rules: bind it to a field that cannot survive a no-op, counter-test both directions before trusting a zero, exit non-zero or it is not a gate, and never treat wording the prompt never specified as a contract.
-
Proving Absence Is Much Harder Than Proving Presence
Four times this week my agent concluded something did not exist. The worst: it ruled that 16 emails were never sent, when they had all gone out days earlier.
-
When the Model Thinks Nobody Is Watching
I had Astra and Fable 5.1 each run an ablation study on my business process. One cut it down past the point a human could run it. The other left something I could still operate.
-
Persuade the Jury, Not the Opponent
How I stay happy online: I moved to Thailand, I separate whose problem is whose, and if I really want to reply I reply once, to the bystanders.
-
Why Client Conversations Drift
Notes from a client diagnosis: three root causes, four prescriptions, plus the main/sub agent pairing run backwards and the rambling only a hook can cure.
-
Don't Vibe Code for the Money
Six things I jotted down after the product I vibe coded made it out of Taiwan and started pulling steady subscribers at home and abroad.
-
Civil Servants Can't Keep Up With AI? Wrong
I reviewed the AI training project demos at a municipal government's research and evaluation office and wrote ten pages of written comments. Veterans have experience and judgment, and with the right guidance it goes a long way.
- Updated:
My Client's Chat Blew Up: Three Questions in One Prompt, and Claude Just Drifted Along With Him
Asking about revenue, metrics, and decisions in a single prompt, or never persisting conclusions after brainstorming. Two real-world cases of client drift with AI, and the fieldwork of using /explain, /first-principles, Grill-me, and checkpoint to rein their workflows back in.
- Updated:
People Who Scored High on the GMAT Work Noticeably Better With AI
An unwritten observation: people with high GMAT scores collaborate noticeably better with AI. So I wrote two GMAT-style questions about AI ability. How many can you get right?
- Updated:
More Subagents Won't Make You Faster
From wanting 13 subagents to chat for me, to one person running 1000, to agents arguing across a table — the real bottleneck is the human main agent, and the fix is deduping before you dispatch.
-
PII Guard TW: A De-identification Tool Built for Taiwan
There's no off-the-shelf de-identification tool for Taiwan's PII formats. So I built one that keeps sensitive data on your machine while you send the rest to AI.
- Updated:
No Error Doesn't Mean Success: Silent Failures, Opus Fabricating Tool Output, and the Hook That Finally Stops It
Exit 0, HTTP 200, and a model saying "done" are all unreliable success signals. Cases from my own code, CLIs, and APIs through to a model inventing commit hashes out of thin air, ending with the one thing that actually holds.
- Updated:
Evaluating AI Models: Why Temperament Matters More Than Benchmarks for Daily Work
With benchmark gaps narrowing, daily experience with AI models comes down to temperament. A breakdown of flagship models across transparency, dialogue stability, and conversational logic to help you find a partner you can actually work with long term.
- Updated:
AI Tool Philosophy — Pick a Mindset, Not a Side
After making nearly thirty tutorial videos: tools change, thinking doesn't. Don't dismiss others' tool choices. What matters is the orchestration logic — task decomposition, validation planning, responsibility assignment.
- Updated:
7 Practical Tips for Talking to Claude 5: Less Drift, Less Chatter, Less Slacking, Less Nagging
Some practical tips I have been sharing with friends lately on talking to the Claude 5 generation of models: a home-made explain skill for alien-speak, answering everything in one pass instead of letting the conversation drift, voice-input setups, first-principles pruning, a hook that treats slacking, a minimal CLAUDE.md, and a fix for the phantom pushback habit.
- Updated:
Where the Quota Actually Goes: Claude Code Cache Bugs, Opus 4.7's 2x Burn, and Runaway Subagents
Four quota incidents in full: two official cache bugs, an 80x jump in API requests pulled from JSONL, Opus 4.7 measured at 2x consumption, and one session of mine that burned 90% in half an hour.
- Updated:
Two Worlds of Web-Based Claude Code Setup
Web Claude Code has two completely different runtimes — cloud VM and remote-control. Here is what I learned setting up both, the small bugs I hit, and why some user-level configs simply cannot be lifted to the cloud as-is.
- Updated:
A Year In, My Answer to "Terminal or Desktop App?" Has Changed Three Times
From March's Cowork token burn, to April's "still the terminal," to May's "beginners should just start with the desktop app," to August's "the CLI is the firstborn." One question, three different answers across five posts, laid out in the order I gave them.
-
Three design mistakes I keep seeing in enterprise AI training
After running AI adoption training for several companies — manufacturing to services, 20 to 100+ people — three mistakes come up almost every time.
- Updated:
Fable 5's One-Week Return: Turning the Most Expensive Model into a Skill Distillation Engine
Fable 5 came back for one week only. I did not spend it on daily busywork — I spent it on high-leverage judgment, distilled into skills that cheap models can follow.
-
Free Gets Fifteen Minutes: Four Rules for Making Money From What You Know
Four rules from fifteen years of teaching: free consultations get fifteen minutes, be a community contributor before you think about money, don't look down on things you assume anyone could google, and how to spot the frauds in your own field.
- Updated:
The More Rules You Add, the Less Claude Listens: Trimming 36% of Baseline Context with an AI Fleet
Piling on rules makes LLMs miss what actually matters. When a 1M context window gets clogged with noise, you need an overhaul. Here is how an automated team of subagents trimmed my baseline harness by 36% by stripping out narrative fluff and conflicting constraints.
- Updated:
Leveling Up the iPad Workflow — Surviving Internet and Power Outages, and Why I Ditched the Magic Keyboard
A follow-up to the full iPad-runs-Claude-Code guide. This one is not about setup, it is about robustness: home internet in Bangkok drops, the rainy season kills the power, so how does the host machine survive? Plus why I gave up the Magic Keyboard for my mouth and some gestures.
-
Your AI Tools Are Now the Attack Vector: npm and Python Supply-Chain Backdoors, a $1000 Stolen-Key Bill, and a Scan-Before-You-Install SOP
Attackers are planting instructions in Claude Code, Cursor, and Gemini CLI configs so your own assistant runs the exfiltration script. Two supply-chain waves, one $1000 stolen-key bill, and the five minutes to spend before installing anything.
-
Your Automated Video Pipeline Is Silently Dropping Work: Gotchas From Remotion, ffmpeg, the YouTube API, and HyperFrames
Four months of video automation gotchas, from Remotion through HyperFrames: truncated downloads that exit 0, thumbnail titles cut mid-word, transcript chunks returning zero lines, large uploads that hang. None of them raise an error.
-
ABC Legal's Agent Fleet: Turning AI Experiments Into a Governed Production System
Six cards summarizing Anthropic's ABC Legal case study: 50+ production agents, agents managed as software in git, a steering committee with no software developers on it, and the principle that trust comes before automation.
-
Once It Ships, It's Gone: Three Times an Agent Misfired on an External System
Calendar invites, replies, and the internal notes on a hold event are all one-way doors. I got each of them wrong once this month, so here are the shapes of the accidents and the checks that catch them.
-
Channel Value and Deal Value Belong in Separate Ledgers
Three rules that came out of two weeks of back-to-back BD calls: score a deal's channel value separately from its deal value, use The Pumpkin Plan to decide which slice to take, and write down a hard ceiling for free consulting.
-
Running Local Models Won't Save You Money
If you're buying hardware to run local models just to save money, do the math with me first: capable rigs start at NT$100k, commercial models are priced to the bone, and privacy is the only irreplaceable edge left.
-
Two Skills for Making AI Speak Plain Language: ASD-STE100 for Grammar, ISO 24495 for Structure
A simple asd-ste100 skill I threw together three weeks ago got reshared and picked up three PRs. To make AI speak plain language, pair ASD-STE100 for grammar and vocabulary with ISO 24495 for logic and structure.
-
Information Diet: I Pulled My Own Chrome History and Audited Where My Attention Goes
People obsess over whether every bite of food is clean, then swallow piles of dirty info online without a second thought. So I pulled my own Chrome history db and had it analyze my browsing habits. A few sections came out uglier than I expected.
-
AI Won't Go to Prison for You: What We Actually Need to Learn Is Accountability
In discussions about AI taking over, timelines and taste miss the point—because AI will never go to prison for you. While everyone rushes toward vibe coding, frontline workers carry an asymmetric risk premium. Learning to audit toolchains and catch errors before signing off is the one skill that can't be replaced.
-
NT$18,000 a Month for an AI Course: Price and Value Came Apart a While Ago
An AI-course cash grab overheard at a convenience store, set against the two things I can actually show: a SKILL distilled from fifteen years of material, and a mock exam interface pulling past a thousand USD a month. Price tracks marketing, not delivery.
-
My Context Window Is More Fragile Than Today's Frontier Models: I Open-Sourced a Document Review SKILL
Working with an agent drains my attention, so I worked out a review loop: he writes the full proposal, I review it by voice, he revises, I go eat and exercise, then v2 shows up.
-
Why I Built My Own Newsletter
Once I started using AI, the feeds I follow changed completely. Threads is full of third-, fourth-, fifth-hand reposts with extra seasoning added, so I pulled together 300-plus sources and had AI build my own weekly digest.
-
The Resource Paradox: From Stalling Old Laptops to the Real Baseline for Local AI
Many put up with aging laptops down to 1.8 GB of storage churning through swap, while dreaming of running local LLMs on a NT$5,000 machine. A look at the resource paradox, realistic hardware baselines for modern local models, and how to weigh usable speed against machine upgrades.
-
Installing Gemini CLI With Beginners Cures Most Claude Code Ailments
Five steps I walk Claude Code / Codex beginners through when installing the Gemini Antigravity CLI: multimodal use cases, getting comfortable with a CLI, a quick win from voice transcription, and finishing on data governance and local mode.
-
The Alignment Gap Is Closing. Next Comes Taste and Verification
The gap between AI and human intent is closing fast, so what separates good output moves toward taste and verification — and verification is where human responsibility stays.
-
Other People Fly to Korea for Cosmetic Surgery. I Had Codex Do Mine.
Someone on Twitter chained Codex, Hyperframes, IndexTTS2 and HeyGen into a one-person media pipeline. I pulled the repo, tried it, played with a local TTS model, then rebuilt a video days later on Claude Code.
-
Six New Context Engineering Rules for Claude 5, and the 1,473 Lines I Cut
After reading Anthropic's context engineering guidance for Claude 5, I turned the key points into nine cards, then cut 1,473 lines from a harness I had built up over four model generations.
-
Fewer Prompts Is Not Less Control
Claude 5 and GPT-5.6 official guidance is converging on the same thing — retiring old-style prompt stacking. But trimming is not letting go; control just moves. And the payoff of maintaining that layering is switching tools without losing a step.
-
My Local Model Lineup on a Mac mini, Plus Three Bad Habits I'm Owning Up To
Ten days of offline models on a 48GB Mac mini M4 Pro — the picks, the task split, a terrifying swap write rate, and three bad habits I admit to.
-
Porting Old Prompts to Opus 5: A Nine-Card Migration Guide
Nine cards I made after reading Anthropic's official Opus 5 prompting guide: response length, agent narration, task boundaries, subagent delegation, self-correction, and what happens when you turn thinking off.
-
Two Voice Prompting Modes: Push-to-Talk for Chores, Long Audio for Deep Thinking
When working with Claude Code or Codex, use push-to-talk for quick prompts and full recordings for deep thinking. Here is a zero-setup browser trick and a headless CLI setup for agents to transcribe audio.
-
I'm Stuck at Stage 2.5 of AI Adoption
After watching Boris Cherny break down the stages of AI adoption, it hit home: I'm stuck at stage 2.5, where limited attention plus low trust becomes a vicious cycle. This week I used late-night schedules, a morning dashboard, and herdr auto-spawning sessions to cut a 4-5 hour workflow down to 1-2 hours.
-
Two Traps Running Claude Code on Local Ollama: Truncated Context, and A3B Buckling Under Heavy Verification
Two things I logged: cco (Claude Code driven by local Ollama) had long been giving off-topic answers, and the root cause was not a weak model but a full harness whose system prompt had ballooned to 30-50k tokens and was being silently truncated; then I put A3B on the Mac mini for two days as a night worker.