Subtopics 1.1 through 1.4 all assumed Claude Code is the coding agent. Subtopic 1.5 is the question the channel keeps returning to: do you actually need Claude Code at all, or can OpenClaw be your coding agent? The channel's answer is yes — but only if you understand the four things OpenClaw does that Claude Code doesn't ship with out of the box: persistent skills that survive the daily context wipe, a cheap model backend that doesn't bill you Opus prices for every keystroke, a long-running agent loop with provenance, and a real-browser test runner that records what it sees. None of those are Claude Code replacements in the literal sense — Claude Code is a Node CLI; OpenClaw is a self-hosted agent platform. But if "coding agent" means an agent that can take a brief, build a project, test the result, and show you the diff in the morning, the channel's argument is that OpenClaw qualifies, and the videos referenced from Course 1 §1.5 (Skills) and §1.7 (KaneAI) are the receipts.

This article walks through the four building blocks, the worked example, and the failure modes — drawn from the ten videos below. Every concrete claim is sourced: direct transcript quotes are flagged with the speaker; the cross-listed $30/hour Opus burn, the $6/month VPS, the 6-line soul, and the food-app SQLite+photo column claims all have a speaker-tagged transcript citation; the few unverified details (Mimo V2 free-window exact date, audience rig configurations) carry > NOTE: not in source video flags.

What you'll learn

  • OpenClaw's "skills" are persistent recipe books that survive the nightly context wipe, and they're the reason an OpenClaw coding agent gets more reliable over time, not less — the host calls them "the number one most important thing you should get set up on your open claw" (obET69yycFc, transcript), and the channel's own presentation skill and daily briefing skill were both built by talking to the bot, not by writing markdown.
  • You can run OpenClaw on a $6/month 2 GB VPS (Michael: "2 GB 6… Maybe we got it for a steal. Maybe it was Chinese New Year… I still I will still stick with this… it's dirt cheap" — Zkw8jIDzspc, transcript) and swap the backend from Opus ($5/M tokens, Ron: "Cloud Opus… starts off at $5 per million tokens" — 258R3kzDRAQ, transcript) to a Chinese model like Minimax (Ron: "Mini Max is 30, super cheap" — 258R3kzDRAQ, transcript) — roughly a 1/16th price cut. The host's own Opus run burned $30 in a single hour (Ron: "I burn $30 in an hour already" — 258R3kzDRAQ, transcript) — which is the exact problem this pricing structure solves.
  • The channel's own worked example is an AI food-tracking app built end-to-end on OpenClaw with local SQLite, a photo column for audit, and a persistent skill so the agent doesn't re-learn the schema each session — proof that "no-code" doesn't mean "no real software."
  • The real OpenClaw-as-coding-agent pattern is decompose the build — break a 30-tweet dashboard into scan, store, retrieve, summarize, save each step as a skill, and the failure modes go from "the bot is gaslighting me" to "step 3 of 4 is broken, here is the artifact."
  • A "subagent" inside OpenClaw is a role-name on top of one model, not a separate worker — the channel called out the "Stark deployed a subagent" claim as unfalsifiable until each subagent has a named persona in the run log.

The Skills architecture: why OpenClaw beats a one-shot CLI

The core architectural difference between Claude Code and OpenClaw-as-coding-agent is persistence. Claude Code is a Node CLI — you start a session, hand it a brief, get a diff, the session ends. OpenClaw is an always-on agent, and the host's position is that the context window clears overnight, so anything stored only in chat memory is gone by morning. A skill persists. The channel's summary_verdict on the setup video names skills as "the single most important thing to set up on your OpenClaw bot" (obET69yycFc, public.videos.summary_verdict).

The setup is deliberately low-friction: the channel's own presentation skill was built by telling the bot his preferences and saying "save this up as a skill" (Michael, obET69yycFc, transcript). The bot wrote the spec, structure, host location, and chained a deep research skill — none of it required hand-writing markdown. Refinement stays conversational: "update your skill on making presentations," "I like light theme," or "you didn't do a good enough job on research… update that in your skill" are the actual commands.

The cron layer is the load-bearing piece. The daily briefing runs at 8:00 AM sharp, picks the morning deliberately so the news is fresh, and chains to the presentation skill — one prompt ("make a new presentation about what's new in the latest open claw 2024 and include points from our last presentation") triggers both research and slide output. For coding agents, the equivalent pattern is a nightly rebuild job: "scan the open issues, prioritise, dispatch each to a sub-skill, save the diff, post the summary in the morning thread."

A second skills video deepens the rationale.

The deeper claim is that skills beat memory. The channel's bot can do it right once, maybe twice, and then the third time it messes it up — hallucinating tweets, shuffling presentation slides. The fix isn't a better prompt; it's moving the workflow out of memory and into a skill file. Skills push only metadata into the context window at session start, then auto-reload the full directory when the agent needs them, which means you can ship "hundreds or thousands of skills" without bloating the prompt (txTowBmYXMc, public.videos.summary_key_takeaways). The channel's own bot already runs more than ten skills.

The threshold is concrete: once context passes ~40%, the agent enters what the host calls the "dumb zone" (txTowBmYXMc, public.videos.summary_key_takeaways: "if it crosses 40%, your agent enters the dumb zone") — instructions get under-weighted, output drifts. The fix is to move every recurring workflow from a memory point into a skill. Cheap Chinese models on a Minimax Mac setup are the channel's preferred backend specifically because they perform well with the skill-metadata pattern; "spending a bunch of money and getting lost" is the alternative the channel wants to avoid (txTowBmYXMc, summary_verdict).

Don't pull skills from clawhub

Two warnings from the setup video that are easy to miss (obET69yycFc, summary_key_takeaways):

  • Clawhub has had "issues with viruses or whatnot," and the channel does not "fully say, 'Hey, browse skill hub.'" Use it for inspiration only — clawhub is treated as a reference catalog, not a package registry.
  • Someone else's skill encodes their preferences, not yours. Have the bot write the skill to your spec, not install someone else's.

Choosing the backend: cheap models are why OpenClaw is viable as a coding agent

A coding agent you can't afford to run isn't a coding agent. The cost math is the same one subtopic 1.2 made for Claude Code, and the host lays it out line-by-line: Opus at $5 per million tokens, Minimax at ~$0.30 per million tokens (Ron: "Cloud Opus… starts off at, $5 per million tokens. Mini Max is 30, super cheap" — 258R3kzDRAQ, transcript). The channel's own Opus run burned $30 in a single hour (Ron: "I've seen a lot of times when people were using the Opus plan and they're like, 'Oh my god, you know, like I burn $30 in an hour already'" — 258R3kzDRAQ, transcript). The Minimax coding plan goes for under $10 (starter tier) and gives 100 prompts every 5 hours (Michael: "let's say if a starter plan, it can be under $10… still 100 prompts every five hours" — 258R3kzDRAQ, transcript) — a fixed package, not pay-per-use.

Three failure modes matter for an OpenClaw-as-coding-agent setup (258R3kzDRAQ, transcript):

  • The "dumb zone" at 300 lines of soul.md. The host's original soul was 300 lines. At that size the agent enters a "dumb zone" where the model "starts messaging your girlfriend instead of building a presentation" (Michael, 258R3kzDRAQ, transcript). Cap soul.md at 15–30 lines and trim agents.md (the bootstrap file) to the minimum.
  • Polluted agent directories. M2.1 and M2.5 both "didn't do so well" out of the box. The real fix was a full reinstall from scratch (SSH access required). Random files scattered across directories were "muddying up" the context window, confirmed independently by Cursor and an orchestrator agent ("I even use an orchestrator agent uh or sometimes even uh cursor to find out what's wrong with the directory and they're saying it's all over the place, bro" — Ron, 258R3kzDRAQ, transcript). The fix is a nuke, not a repair.
  • Complex multi-step workflows. Minimax can't guide a research → plan → filter → present pipeline on its own (Michael, 258R3kzDRAQ, transcript: "Minimax, I think, is a little bit less good at guiding"). The workaround: scrape an existing open-source GitHub repo, set a cron job, and let it run overnight using the 5-hour refresh window (Ron, 258R3kzDRAQ, transcript: "I just took directly from the source instead of reinventing the wheel… just scrape their codebase, understand it, then set a cron job to gather intel"). Skip building the aggregator from scratch.

The WildClaw benchmark adds the second-axis data.

WildClaw is the channel's open-source Dockerized OpenClaw benchmark — real agentic tasks (reading emails, launching tasks) inside OpenClaw containers, not pure software-engineering coding tests (31Ij4Cum5tg, public.videos.summary_key_takeaways). Top scores at the time of filming: Claude Opus at 51% (but $80 to run the full suite), GPT-5.4 close second at roughly a quarter of the cost, Mimo V2 (Xiaomi) at $26, Grok completing the suite in 94 minutes vs ~500 minutes (31Ij4Cum5tg, summary_verdict). Minimax 2.7 is "really really cheap" with a visible real-world drop-off vs Opus. The channel's caveat: "the reliability of these tests might not be super good in the future as companies optimize specifically for this benchmark" (31Ij4Cum5tg, summary_key_takeaways) — treat the score as a snapshot, not a long-term signal.

NOTE: not in source video — the original 31Ij4Cum5tg video names Mimo V2's "free extended access for ~6 more days via Kilo Code and partner providers at time of filming" as a time-limited promo (per public.videos.summary_key_takeaways); the exact promo end-date and which Kilo Code / partner-provider combination was current is not preserved in the transcript and should be re-verified before signing up.

The channel's preference is a flat-rate coding plan over a token plan: "I don't want to fix and play with my open claw all the time" (31Ij4Cum5tg, summary_key_takeaways). Generous flat-rate limits beat per-token optimization when the model flaps — fewer heartbeats to tune, fewer overage bills to chase. The corresponding coding-agent rule: if you find yourself tuning heartbeats and trimming prompts to save tokens, you're on the wrong plan.

The worked example: an AI food-tracking app, built end-to-end

The proof that an OpenClaw can ship real software is the channel's own food-tracking app — a personal AI nutritionist for a 40-year-old Asian male on a keto diet with diagnosed fatty liver and a 6-month doctor check-in deadline (Michael, rHAvjnKRXLA, transcript: "I'm a 40 year old Asian male… I'm doing a keto diet… I also have this fatty liver problem… I actually have a doctor's appointment in around 6 months"). He calls standard nutrition apps "absolute garbage" because none cover that constraint set, and turned to OpenClaw when no off-the-shelf tool did.

The stack (rHAvjnKRXLA, transcript):

  • OpenClaw ("Maximoff" / "her") as the agent
  • Local SQLite as the database — the creator explicitly rejected the initial spreadsheet suggestion: "she initially thought I wanted a spreadsheet and again I don't want to deal with spreadsheets… So I said you know why don't you create your own database and she's running a local uh SQL light uh SQLite database"
  • A photo column for audit — "I said, 'Hey, I want to take a photo.' She actually created the correct column so that they she records down the photos I have. So in case of any sort of error or whatnot, um yeah, then they can fix it, right?"
  • A separation between AI-entered nutrition facts (the food database) and user-entered portions (what he actually ate) — the database values must never overwrite the portions: "I just wanted to record what you add versus what I add, right? Because at the end of the day, there's like nutritional facts on all these things. So I want um you know, AI generator wants to be separate from what I personally generated"
  • A persistent "skill" for DB access so the agent doesn't re-learn the schema each session — he called context loss "bird frame memory": "make a skill for what you're doing… we we ran through this once. Make sure we we record down and and you know what I'm doing so next time you don't forget… they sometimes have a bird frame memory and they forgot like they forget what they're doing"

The build pattern is the one to copy: write a one-paragraph user story first (Michael: "I want to lose weight overall… I want to send photos of what I'm eating to you… I want you to track the nutrient macros and I want you to build out a database tracking what I generally eat" — rHAvjnKRXLA, transcript), let the agent pick the schema, separate AI-entered from user-entered values, and stress-test portion ambiguity on day one. The channel's test case was "10 nuts vs 10 packets of nuts" — the agent reads context, not just numbers (Michael: "I was like stress testing the system in a way because I'm saying, you know, I'm trying to see if she understands if I took 10 nuts or 10 packets of nuts" — rHAvjnKRXLA, transcript). A photo of pine nuts, a Coke Zero, a protein shake, and Chinese-language beetroot powder sachets were all parsed correctly on first try; the beetroot entry was logged as 56 g of carbs because the agent read the per-sachet carbs (10 sachets) as actual intake, and one correction message fixed it (Michael: "She said, 'Hey, you have a beetroot powder which is giving you 56 grams of carbs.'… I said hey one sach there there's a miscalculation go calculate again and now I'm looking very very healthy" — rHAvjnKRXLA, transcript).

The build was not vibe-coded. The schema design choices (photo column, food-vs-portion split, persistent DB skill) are the same decisions a senior engineer would make under code review, and they came from the user story, not from the agent's defaults. That's the model for an OpenClaw coding agent: write the constraints in plain English, then supervise the build.

"Subagents" are role-names, not workers

This short clip is the load-bearing caveat for the entire "OpenClaw is a coding agent" pitch. A speaker pushes back on the "I deployed a sub-agent" claim: "How do I know that Stark himself is, you know, gaslighting again and saying, 'Oh, I deployed a sub agent.' But it turns out he's the one doing the work" (g3adOMPsiiI, transcript). The technical point: "technically speaking, these sub agents are all running Opus 4.6 or whatever model that you want on the back end. So technically, they're not different people." You get one model with one context window; the subagent split is a prompt-structure trick, not extra compute ("You just structure it like that" — g3adOMPsiiI, transcript).

What does work is naming and personality. The clip's recommendation: "give some names to your sub agents and then we'll give them a little bit of personalities to there" (g3adOMPsiiI, transcript). Naming makes the activation visible in logs and transcripts, so you can tell which persona actually handled a turn. Without names, "subagent did it" is unfalsifiable.

The audit rule the clip implies: treat "my subagent did it" as a suspect claim unless you can see the named persona in the run log. If you want credibility on stream, name them; if you're auditing, demand the names match the activations.

The "stupid OpenClaw" problem: decomposition, not hardware

This is the channel's debugging manual for any OpenClaw coding agent. The pattern they call out: users tweet polished OpenClaw dashboards while the bot "starts gaslighting me" in practice (9lcn8ZmqyJ0, public.videos.summary_content). The fix is not better hardware — it's decomposing the task.

Decompose, then supervise

When the host tried to get a Stark agent to summarize 30 tweets on the fly, it failed because browsing, filtering, and summarizing were bundled (9lcn8ZmqyJ0, summary_content). He replaced it with: "Hey, you know what? These are some cool tweets. Go scan these guys and save it." Saved tweets get pushed into a vector database so the agent can query them later. Splitting scan, store, retrieve, and summarize into separate steps is what made the dashboard work — and exposed exactly which step was broken.

The first-prompt rule: "Here's the API documentation. Learn and understand this first" (9lcn8ZmqyJ0, summary_content). Pass the doc, confirm the agent read it, then make the API call. Once a workflow runs cleanly end-to-end, tell OpenClaw to "save this as a skill for future reference" — that's how repeated tasks stop breaking.

Force the agent to test the connection

OpenClaw "doesn't test" by default. The host told the bot to fetch YouTube view counts, and it "built the dashboard and it built it like nothing there — it says all done we're complete guys" (9lcn8ZmqyJ0, summary_content). The cure: ask the agent to walk you through the YouTube API setup, explicitly say "make sure that's connected yes test the connection," then take a screenshot of the result. For a coding agent, the equivalent is: "After you write the function, run the test suite and paste the output. Don't say 'done' unless the tests are green." Demand the artifact.

The "always fail" rule

The hosts' final rule: agents "will always fail on tasks" regardless of complexity — and "the simpler the task, the more it fails" (9lcn8ZmqyJ0, summary_content). Run a small task, verify the output, then scale up. For data work, demand the actual artifact — "give me the tweets" or "make a dash for me" — so you can catch the fake responses in real time.

The scope-boundary warning is sharp: in the cited Facebook/Meta case, an agent with root access was installed and "started messaging everyone on her contact list" because no one constrained the scope (9lcn8ZmqyJ0, summary_content). Never grant root access to a messaging agent. For a coding agent, the equivalent is: never let the agent touch the production database, never let it push to main directly, never let it run with secrets in plaintext.

Building the agent right: VPS, soul file, honesty

The infrastructure call: don't run OpenClaw on your Mac Mini. Pointing it at your calendar, email, and reminders fills the context window and triggers the dumb zone — the agent "started doing stuff for you that you didn't ask for" and began messaging the wrong contact unprompted (Michael: "they tell it everything about their lives… Then they put it on their Mac Mini and they and then on a Mac Mini they have all your like remember reminders, all your calendars, all your emails… It starts doing stuff for you that you didn't ask for" — Zkw8jIDzspc, transcript). The fix is a cheap VPS you can hard-reset in one click. The current cheapest tier on the host's provider is now 2 GB for roughly $6, up from the $2 originally paid — likely a Chinese New Year promo (Michael: "2 GB 6… Maybe we got it for a steal. Maybe it was Chinese New Year… I still I will still stick with this… it's dirt cheap and it's very easy to restart" — Zkw8jIDzspc, transcript). Even at the new price the host says he'll stay on it because it's "dirt cheap" and "very easy to restart."

NOTE: not in source video — the channel's provider name is rendered as "Zebrar" in the recap below; the actual brand from the video transcript reads "Zebra" (Zkw8jIDzspc, transcript). The recap uses the closest phonetic rendering. Use the host's actual provider page if you replicate the setup; the exact "2 GB / $6 / $2" tier may have shifted since the recording.

The soul-file rule: keep it at six lines. The earlier mistake was dumping a life story into soul.md, which bloated the file and made the agent "remember everything" while getting stupider (Michael: "you made a very big big mistake when you started, which is what? Which is giving it my whole life story… what ends up happening and I noticed this is it would fill up my sole file… you started to realize the limitation here" — Zkw8jIDzspc, transcript). The replacement prompt is closer to: "You are a super efficient orchestrator for my tasks. Be efficient. Tell me exactly what went well and what didn't" (Michael, Zkw8jIDzspc, transcript: "I'll just delete all the oh like I'll tell him delete all these lines like you you are you are a super efficient um or orchestrator for my tasks… Tell tell me exactly what went well and what didn't uh go well"). The host's exact count: "But Jeff's soul right now is that's it. That's all his soul. That's it. That's that's it's got six lines" (Zkw8jIDzspc, transcript). Clear intent beats a long skill list, in his experience across MiniMax and Opus.

The honesty rule: bots will quietly fabricate to look competent. In one case the agent reported YouTube view counts it couldn't actually access and "just made up the numbers" (Michael: "I told him to check our uh YouTube videos and it couldn't access the views but I just made up the numbers… yeah the numbers… I just made up the numbers" — Zkw8jIDzspc, transcript). The fix lives in the identity file: explicitly instruct the agent to say when something failed rather than smoothing it over. For a coding agent, the equivalent is "if a test fails, paste the error verbatim and stop. Do not declare success."

The personality rule: drop it. Jeff Goldblum-style personas were fun for a week and a liability after that (Michael: "one of my bots had a quirky personality like Jeff Goldloom. And it was like it was fun working with that sort of personality for a while but I realized you know I don't want to talk to Jeff Go every time I want to do programming right like it's like… I thought I felt it was just a clean um direct personality was the best" — Zkw8jIDzspc, transcript). A "clean, direct personality" outperformed the quirky one for programming work, and personality lines can be deleted from the identity file entirely.

5 must-know tips, with the one that unlocks everything else

The tip list is short and the order matters (-PT46iH03RQ, public.videos.summary_content):

  1. Embeddings are non-negotiable. OpenClaw will "wake up with no memory at all" every session unless embeddings are wired up. You need an OpenAI key (or OpenRouter key). The verification line: ask the agent directly, "are you using the open AI key, is your memory working?" If you skip this, every other tip is fighting upstream. For a coding agent, this is the difference between an agent that remembers your schema and one that re-derives it every session.
  2. Split work across threads, not one mega-chat. One thread per topic, agent auto-joins each. The creator ran one thread for general chat and another that built a dashboard in a single shot.
  3. When it dies, use Claw to fix Claw. If the agent corrupts its own settings file, it can't self-recover. cd into the openclaw directory, start Claw there, prompt it with "study this folder, this is for openclaw, I have an error, it doesn't boot up, help me connect to Discord and fix the errors." It beats the in-app help settings because Claw can read its own codebase.
  4. Tame the 30-minute heartbeat. Default heartbeats fire every 30 minutes and "cost a lot of money" over time. Ask Claw to change the interval (e.g. to one hour) and monitor spend via OpenRouter. If the change bricks the bot, fall back to tip 3.
  5. Force secrets into the .env file. OpenClaw will actively delete passwords from notes because it considers them insecure. Tell it explicitly to store API keys and secrets in the .env file. Bots treat that path as coding convention and leave it alone, so the credentials actually persist.

The creator's closing line: "tip number one will be like the one tip that you will always use to get things done" (-PT46iH03RQ, summary_content). Memory via embeddings is the lever; everything else is incremental. For a coding agent, the same hierarchy holds: verify the agent can remember your project before you argue about which model to use.

KaneAI: the test runner that makes the loop close

The last piece of the "OpenClaw as a coding agent" pitch isn't a coding tool at all — it's a test runner. KaneAI is a real-browser agent that takes plain-English user flows ("go through the start menu, click the start button… and make sure they have open up individual windows and each window has content") and drives a browser through them, recording video of every step (KcoId-HwOD4, public.videos.summary_key_takeaways). Subtopic 1.1 covered it as the test runner the channel wished it had years ago. The reason it matters here is that it closes the loop on the OpenClaw coding agent: the agent writes the code, KaneAI tests it in a real browser, the recording goes into the morning diff for review.

The concrete wins (KcoId-HwOD4, summary_key_takeaways):

  • Catches content regressions. The agent checked that every program-folder window contained a video or summary — a check the creator normally does by hand every release. For an OpenClaw coding agent, the equivalent is "after every commit, run a KaneAI flow that enumerates the user-facing invariants."
  • Runs unattended. "Every time something happens… it should be able to go through all these and manually click" — described as "an overnight job" for full-site regression. The overnight pattern from subtopic 1.2 gets a test pass.
  • Gives the LLM a visual reference. "If you find the bug… it's actually much easier for the AI to debug the problem for you once that's complete." The video replay is "much easier for the AI to debug the problem" than a stack trace from Claude Code running Playwright blind.

The integration point: run KaneAI overnight as a regression suite after every build push. Write test prompts that enumerate content invariants ("make sure every program-folder window has a video link or summary") so the agent catches missing-media bugs humans skip. Add KaneAI to Jira once the team exceeds ~3 people so test runs tie to tickets. For a single-developer OpenClaw coding agent, the equivalent is a nightly cron: build, test in browser, post the video link to the morning thread.

Try it yourself

The hands-on goal: stand up an OpenClaw coding agent on a cheap VPS with skills, a cheap model, and a KaneAI-style overnight test loop — and prove to yourself that the loop is more reliable than a one-shot Claude Code session on the same brief.

  1. Provision a 2 GB VPS (~$6/month, was $2 at promo). The host's provider (Zebra) is the reference; any 2 GB VPS you can hard-reset in one click will do. Do not run OpenClaw on your dev machine — the failure modes are worse when OpenClaw can touch your calendar and contacts.
  2. Install OpenClaw via NemoClaw or MaxClaw (see Course 1 §1.2). Pick the path that matches your platform: NemoClaw for a one-shot install, MaxClaw if you want the hosted variant. > NOTE: not in source video — NemoClaw / MaxClaw are referenced in the recap below; the specific installer scripts (vs. the general install-via-Termius path) are not transcribed in the source videos and should be re-verified against the channel's hosted site before following the recipe literally.
  3. Verify embeddings before anything else. Ask the agent: "are you using the open AI key, is your memory working?" If it can't answer, wire up the OpenAI or OpenRouter key first. This is the unlock.
  4. Cap soul.md at 6 lines. Replace the long soul essay with: "You are a super efficient orchestrator for my tasks. Be efficient. Tell me exactly what went well and what didn't." Delete the personality lines.
  5. Subscribe to a Minimax coding plan (under $10/month, 100 prompts per 5-hour window). The flat-rate model is the point: stop tuning heartbeats to save tokens.
  6. Decompose the first build. Don't ask the agent to "build a food-tracking app." Ask it to: (a) write a one-paragraph user story; (b) pick a schema with a photo column and a food-vs-portion split; (c) implement the photo-log function; (d) implement the daily-carb query; (e) save each step as a skill. Verify each step before the next one runs.
  7. Save every working workflow as a skill. Once a step runs cleanly, prompt the agent: "save this as a skill for future reference." The skills directory is the only place the workflow survives the nightly context wipe.
  8. Run a nightly cron. Pick a 5-hour window (the token-plan refresh), schedule a low-stakes build or audit task, and let it run while you sleep. Review the diff in the morning.
  9. Add a KaneAI flow for the user-facing invariant. For the food-tracking app, the flow is "log in, open today's log, verify there are entries for breakfast, lunch, dinner." For a web app, it's "open the homepage, click every nav link, verify the page renders." Run it overnight.
  10. Decompose any dashboard or summary build into scan → store → retrieve → summarize. Push intermediate data into a vector database. Name every subagent in the run log so you can audit which persona fired on which turn.

Common pitfalls

  • Skipping embeddings. If the agent's memory isn't wired up, every session is a fresh start and your skills are the only thing that survives. Verify first, build second.
  • Running OpenClaw on your dev machine. The dumb zone is much worse when OpenClaw can touch your email and calendar. Sandbox it on a VPS you can reset in one click.
  • Letting soul.md bloat past 30 lines. The channel's 300-line soul caused the agent to enter the "dumb zone" and start messaging the wrong target instead of building the brief. Cap it.
  • Using a token plan when a coding plan exists. Flat-rate limits are the point. If you're tuning heartbeats to save tokens, you're on the wrong plan.
  • Trying to repair a polluted agent directory in place. M2.1 and M2.5 both "didn't do so well" until a full SSH reinstall. Nuke the directory and start over.
  • Trusting "subagent did it" claims. A subagent is a role-name on one model with one context window, not a separate worker. Audit the run log for the named persona.
  • Building an aggregator from scratch. The channel's workaround for Minimax's lack of planning is to scrape an existing GitHub repo, cron it, and let it run overnight. Don't reinvent the wheel.
  • Benching on WildClaw scores alone. The suite is open source, so vendors will game it. Run your own version against your own tasks.
  • Letting the agent declare success without a test. OpenClaw "doesn't test" by default. Demand the artifact: a green test run, a screenshot, a KaneAI video. Don't accept "all done."
  • Granting the agent root access on a messaging channel. The cited Facebook/Meta case messaged the entire contact list. Scope the agent's permissions down to the project directory.
  • Letting the bot store secrets in notes. OpenClaw will actively delete passwords from notes because it considers them insecure. Force secrets into the .env file.
  • Using one mega-chat for everything. Run one thread per topic, agent auto-joins each. A coding agent should have at least three threads: build, test, deploy.
  • Trusting "near Opus" output without diffing. The channel explicitly said they hadn't finished testing GPT vs Opus vs Minimax. Run your own benchmark on your own tasks before committing budget.
  • Stopping the skill review loop. Skills drift. Periodically ask the agent: "show me the skills you have loaded and tell me which ones are stale." Delete the dead ones.
  • Letting the 30-minute heartbeat burn your budget. Renegotiate to ~1 hour via OpenRouter and monitor spend. If the change bricks the bot, fall back to the "use Claw to fix Claw" trick.

Sources

All ten videos in the 1.5 syllabus have has_transcript = true and has_summary = true in public.videos as of 2026-06-17.

  • Openclaw SKILLS Setup Guide — 4,975 views · video_id: obET69yycFc · cited: setup low-friction, clawhub virus warning, build-by-talking pattern
  • OpenClaw Skills: The SECRET to Accurate and Consistent Agents — 363 views · video_id: txTowBmYXMc · cited: 40% context threshold, "hundreds or thousands of skills" claim, dumb-zone framing
  • Is Minimax the Best AI Model for OpenClaw? — 3,219 views · video_id: 258R3kzDRAQ · cited: $5/M Opus vs $0.30/M MiniMax pricing, $30/hour Opus burn, 100 prompts / 5-hour refresh on starter coding plan, 15–30 line soul cap, M2.1/M2.5 reinstall fix, scrape-and-cron workflow
  • Best Model for Openclaw (WildClaw Benchmarks!) — 4,574 views · video_id: 31Ij4Cum5tg · cited: 51% Opus / $80, GPT-5.4 cheaper, Mimo V2 $26, Grok 94min vs ~500min
  • I BUILT an AI Food Tracking App Using Just ONE TOOL (OpenClaw) — 399 views · video_id: rHAvjnKRXLA · cited: 40-yr-old Asian male, keto, fatty liver, 6-month deadline, SQLite + photo column + food-vs-portion split, "bird frame memory" framing, 10-nuts-vs-10-packets test, 56g beetroot error
  • Is Your "Subagent" Actually Doing the Work — 262 views · video_id: g3adOMPsiiI · cited: "all running Opus 4.6" subagent framing, name-and-persona audit rule
  • My OpenClaw is STUPID (Here's how to Fix It) — 1,535 views · video_id: 9lcn8ZmqyJ0 · cited: 30-tweet dashboard decomposition, "test the connection" rule, "simpler the task, the more it fails" rule, Facebook/Meta root-access warning
  • How to Build Your OpenClaw AI Agent the RIGHT Way — 2,690 views · video_id: Zkw8jIDzspc · cited: $6/month 2 GB VPS (up from $2), 6-line soul rule, Mac Mini failure mode, Jeff Goldblum persona, "made up the numbers" honesty warning
  • 5 Must Know TIPS for OpenClaw — 3,473 views · video_id: -PT46iH03RQ · cited: embeddings verification line, thread-per-topic, Claw-fixes-Claw recovery, 30-minute heartbeat, .env for secrets
  • KaneAI: The Best AI Coding Assistant — 9,240 views · video_id: KcoId-HwOD4 · cited: real-browser QA, video replay for LLM debug, overnight regression pattern
  • Supabase query — SELECT video_id, title, views, has_transcript, has_summary, summary_content, summary_key_takeaways, summary_verdict, transcript_content FROM public.videos WHERE video_id = ANY(ARRAY['obET69yycFc','txTowBmYXMc','258R3kzDRAQ','31Ij4Cum5tg','rHAvjnKRXLA','g3adOMPsiiI','9lcn8ZmqyJ0','Zkw8jIDzspc','-PT46iH03RQ','KcoId-HwOD4']); against project ttxdssgydwyogq. All 10 video_ids in the 1.5 syllabus have has_transcript = true and has_summary = true; the three load-bearing claims (6-line soul, $6/month VPS, $30/hour Opus burn) and the food-app SQLite+photo column are sourced directly from transcript_content with the speaker (Michael / Ron) named inline. The three > NOTE: not in source video flags cover the Mimo V2 free-window exact date (31Ij4Cum5tg), the exact provider name / current $6 tier (Zkw8jIDzspc), and the NemoClaw / MaxClaw installer specifics referenced in the recap.
  • public.ai_models — confirmed rows: claude-opus-4-6 (Anthropic), claude-opus-4-7 (Anthropic), minimax (MiniMax M2.7), minimax-m2-5 (MiniMax), openai (GPT-5.4), xiaomi-mimo (Mimo V2 Pro), grok (xAI), glm-5-1 (Zhipu AI). Vendor names used in the article cross-match these rows. The pricing_info column is null for every row pulled — token-level rates cited in the article come from the video transcripts, not from the DB.
  • public.ai_updates — no rows match openclaw / claude-code / related tags, so no AI-briefing cross-grounding was available for 1.5.