Joshe

Member
  • Content count

    3,294
  • Joined

  • Last visited

Everything posted by Joshe

  1. I like the Kia Forte I bought a few years ago.
  2. Check out this AI banger:
  3. What do transitions between stages actually look like? We talk about the stages constantly here but almost never about the transitions. So I'm curious about people's actual experience. Not the theory, the lived version. Pick whichever transition you consider your most significant and describe it. If you've got more than one, pick the biggest. I've put some questions under each one to give it some shape, but answer however you want. I'll share mine below. Red to Blue If you had a phase of impulsiveness, defiance, or chaos and then took on structure, faith, discipline, or a code: what was it? What was life like before, and what made the structure appealing? Was there a person or institution involved? How long did it take to feel settled? Blue to Orange For those who grew up inside religion, tradition, strict family rules, or a tight community and left it, or moved toward ambition, achievement, and self-direction. What were you like as a kid inside it, the one who fit or the one who asked questions? Did your siblings, raised in the same house, go the same way, or are some still in? Was there a trigger, or did it erode? How long did it take? What can you do now that you couldn't then? Orange to Green For those who were driven, career-focused, competitive, and then shifted toward relationships, feeling, community, or social concern. What did the Orange period feel like from inside, and what stopped working? Was there an event, a burnout, a relationship, a book, a retreat, or did it just fade? Did people around you go through the same shift at the same time? What did you gain, and what did you give up? Green to Yellow For those who feel they've made this one. What did Green feel like from inside while you were in it, and what was the first crack? Was there a moment, or was it gradual? What can you do now that you couldn't do at Green, concretely? How would someone else be able to tell the difference from the outside? And how do you experience "seeing all the stages" now: is it something you do, or something you're aware of doing? Anything else If your biggest shift doesn't fit these, describe it anyway. And if you feel like you've never really shifted, that you've been more or less the same person with the same outlook since you were young, I'd like to hear that too. A few things I'm especially interested in across all of these: how long transitions take, what triggers them, whether other people in your life moved at the same time, and what specifically you could do afterward that you couldn't before. Are you able to see clear transitions in your own history, and if so, what did it look like?
  4. Well, I've only got a sample size of 1, so IDK how rare it is, but I'm guessing at least 5% of people can handle this decoupling without a problem.
  5. Of course if judgement shuts down inquiry and understanding, it's a problem. But it's possible to moralize and to then suspend it in favor of understanding. So if you catch someone moralizing and then chalk them up as Green or someone with shallow understanding, that can be a mistake. It's not the best heuristic, because some people have no problem turning off the morality when it's time to think, yet then turn it back on when they're done, lol.
  6. You may be mistaking your preference for emotional detachment as a universal prerequisite for deep understanding. Some cognitive architectures can run moral judgement and deep understanding in parallel without the understanding getting contaminated. There are people who study atrocities and monsters for a living and those people are moralizing the subjects of their study. It's possible to judge Hitler as an all out piece of shit, and to understand his mind through and through, even on a deeper spiritual level where one can empathize with the evil-doer in their exploration/mapping out of the subject, and then when the exploration is done, still hold "Hitler was a piece of shit" in a top slot of their Hitler model.
  7. You can have high moral reactivity and still see clearly. You don't necessarily need to drop it. You just need to expand your thinking and to care about getting an accurate view of the map. Most people develop lower or more calibrated reactivity as a consequence of living several decades. Some people are born with low reactivity. 10 years from now you'll very likely be less reactive without any intentional effort on your part. Low moral reactivity is largely a byproduct of temperament + age/maturity, not some sophisticated next level unlock. If emotionality prevents clear seeing, that's a problem, but you won't find the solution by studying SD/Green/Yellow. Seems to me the solution would be to practice regulating while exposing yourself to people, situations, or ideas that you're inclined to judge, and reflect upon those experiences. It seems like a very basic skill to acquire for those who actually want it. Seeing clearly is largely a matter of acquiring, observing, connecting, and judging data objectively. You can do all that and be high in moral reactivity at the same time. This nuance seems lost on many self-appointed Yellow folk. Careful not to let people gaslight you with something like "your reasoning is tainted because you're green". It often goes like this: someone states a position, then a self-appointed yellow person sees the position is flawed and if they detect emotional charge, they lazily chalk up your bad reasoning to being Green. When in reality, your reasoning may have been faulty because you were simply missing some key data or failed to consider a variable. An error can have many causes. So keep an eye out for people quick to use that cheap heuristic. It's more often than not them pulling rank on you(childish status games), not them trying to improve your thinking. If they wanted to improve your thinking, they'd actually highlight and explain your error rather than pull rank on you. This is one of the more toxic things that occurs here.
  8. Let your freak flag fly. There will be many who don’t like it.p, but it filters for the people who do.
  9. Damn, I loving this dude's content.
  10. Jesus. I didn't know Opus 5.5 was such a big leap. This levels the fuck out of the playing field, which kind of sucks, lol.
  11. If you identify with a description and then try to operate like the description, you got issues.
  12. If you're the type to let a model tell you who/what you are, as opposed to the ground truth. You can see a description and say "yeah, that's what I'm like, that's how I operate". Acknowledging what you're like does not necessarily entail locking yourself into that. It seems that only an ideological mind, or some needy/wanting to be a particular way kind of mind, or something of the sort would be at risk of being hemmed in. As soon as ground truth changes and the descriptions no longer match, that's just that.
  13. This is just part of my current solution for working on large projects over the span of months. It's overkill for projects you can complete in a weekend. It's a "meta-harness" that can run any other harness. The idea is you set agents up inside an organization with you, the CEO at the top, and you "hire" other agents. You can set up a supervisor agent, research agent, or any agent you want, then delegate work to them, and they can all work together and communicate with one another. This dude is magoo, but here ya go: I've never designed a conversational chatbot but what you want sounds fairly simple. Depending on your volume, trusting AI to converse on your behalf is risky. If it were me, I would want to approve all AI responses. Are you trying to make users think they're talking to a real human? Why not just let them know they're being vetted by a chatbot and you'll get back to them? The only way I think this makes sense is if volume is too much to handle on your own.
  14. @LordFall Have you heard of Paperclip AI? Seems like something you'd resonate with.
  15. @LordFall What are you currently doing with coding agents? Claude or ChatGPT? App or command line? VS Code? What's your CRM idea? How would it be different from any other CRM?
  16. Well said! I wish I was as articulate as you, haha. Sure thing. I'll share more later once I get my current idea worked out. Up till now, I've just been winging it from the start of each project and implementing solutions as I go, and the other day I realized I had 3 large projects and each had different setups, so I audited them and am consolidating all the best stuff. The main thing I'm working on right now is an app to assist me with my freelance work. It's huge. A ton of file management, new project setup and management, time tracking, file modification, and just a ton of stuff. I wish I could share screenshots but the app is populated with all my client data and I can't share due to NDAs. It's amazing, really. It's saving me so much time and energy. Outside of work, I have about 6 projects going. Things like a command center app with tons of features, including a custom launcher like Raycast. A custom file explorer. A global notification system guided by AI that is aware of all my emails, my schedule, calendar, to-do list, etc. And making it so the notifications can reach me wherever I'm at. Like if I'm not at my desk, then notifications will come through to my android phone. And I want to set it up with a text to speech model like Kokoro-82M so it can talk to me in real time. I've been working on a design system and component library so I can build UI quick. I feel like if you're going to be building a lot of apps, you need UI solved so the agent doesn't get bogged down with it. Here's a sample of my component library I put all those components in an npm package and let every app I have consume it. So when I need any UI element, it's already built and styled the way I like. Back to the workflow. My current setup is too much to explain but maybe this will suffice: The main thing you want is to avoid filling up the context window. If you're at 50%, that's too much. You need to run /clear often. But when you do that, you need a way to get back to where you were without losing important facts and context. I solved that with a /handoff skill. I have the agent do one thing that might take up 15-25% of the context window, then I clear. I'm not sure if it would be helpful to anyone but it looks like this: --- name: handoff description: Wrap up the current session so Josh can safely /clear — assess whether now is a good stopping point (and say so if not), commit green work, write the worklog handoff entry, push, and hand Josh his next-session opener. Use when Josh types /handoff or says anything like "wrap up for handoff" / "let's stop here". --- # /handoff — end-of-session wrap-up Josh is about to clear context. Your job: make the next session resumable for the cost of reading one worklog entry. Follow the phases in order. ## Phase 0 — GATE: is now actually a good stopping point? Assess the working state first. **Bad** stopping points: - Code is mid-edit or broken (build/typecheck/tests failing because of half-applied changes — not pre-existing failures). - Mid-debugging with an untested hypothesis (the understanding in your head is the most expensive thing to lose). - A multi-step operation is half done (schema half-migrated, rename half applied, dependency half swapped). If the state is bad, DO NOT wrap up silently. Tell Josh: 1. Why this is a bad freeze point (one sentence). 2. The nearest good one and what it takes to reach it ("finish wiring X and run its test — I can do that now, ~a few minutes"). 3. Offer both paths: (a) push to that point, then hand off (recommended); (b) emergency handoff anyway — then capture in-flight state in the worklog's **In-flight** section (see Phase 3) at forensic detail: every touched file, the intent, the exact next edit, the current hypothesis if debugging, AND what's already been ruled out (with the evidence — dead ends are as expensive to reconstruct as the live hypothesis). **Exception:** if this close is driven by context pressure, skip (a) — do the forensic capture immediately, while there are still tokens to write it well. A capture is cheap; a fix that dies mid-flight leaves the state bad AND undocumented. Remember: /clear does not touch disk. Uncommitted changes survive; only _understanding_ is lost. Write down the understanding. ## Phase 1 — land the work - Run the relevant checks (tests/typecheck for touched workspaces). - **Pre-existing failures don't block landing.** If a check fails, verify against the base commit (`git stash && <check> && git stash pop`, or check whether the failure touches files you changed). Failures that predate this session get noted in the worklog State section and do NOT block committing this session's work. Do not spend context fixing breakage unrelated to the session's task unless Josh asks. - Commit completed, green work with a proper message. - Do NOT commit broken code to main. Unfinished-but-safe changes stay uncommitted on disk and get described precisely in the worklog instead. - **Cross-repo writes:** if this session filed intake specs into josh-ds, check ITS tree (`git -C ...\josh-ds status --porcelain`) and report what's uncommitted there in the worklog State section. NEVER commit or push in josh-ds from an app-side session — that repo's git belongs to its own sessions. - **If nothing shipped** (pure design/discussion session), Phase 1 is a no-op — skip it and say so in the worklog. Do not invent work to satisfy this phase. ## Phase 2 — persist durable knowledge (the repo is the brain) The bar for persisting: **a future agent would get something wrong without it.** Do not persist anything derivable from reading the code, and do not write "no new facts this session" filler — if a doc has nothing to gain, don't touch it. - New domain facts Josh revealed this session → `docs/DOMAIN.md`. (Domain facts = truths about the business/problem space, not about the code.) - Decisions that crystallized → `docs/adr/` (status: proposed unless Josh explicitly accepted in-session — his positive reaction to a suggestion is NOT acceptance; update the ADR index). - Changed operational facts (commands, repo map, doc pointers) → `CLAUDE.md`. A new trap that cost this session → `docs/GOTCHAS.md` (append with the next number; never renumber). Tool behavior/rules → `docs/APP-TOOLS.md` or `docs/WORKSPACE.md`. - Scope/backlog changes → `docs/ROADMAP.md`. - **Triage `docs/PARKING-LOT.md`**: move every item to its real home (ROADMAP / DOMAIN / adr / a future slice's notes in BUILD-ORDER) and delete it from the lot. The worklog entry is the triage receipt — do NOT leave "(earlier triage …)" notes behind in the lot. If an item's home is genuinely unclear, leave it listed with a note — but that should be rare. ## Phase 3 — the worklog entry (the baton) Append to `docs/worklogs/<YYYY-MM>/WORKLOG-<today>.md` (create the month folder and/or file if new): - **Did:** what shipped/changed — **at most three sentences + pointers** to where the detail lives (slice doc / WORKSPACE.md / commit). The entry is a baton, not a record; if Did retells what a slice doc already holds, cut it. Title the section `(no slice)` when CURRENT is none (COLLABORATION.md → "No-slice sessions"). - **State:** tests passing? any pre-existing failures observed? anything uncommitted (list files + why)? - **In-flight:** (emergency handoffs only — omit otherwise) the forensic capture from Phase 0 path (b): touched files, intent, exact next edit, live hypothesis, ruled-out hypotheses with evidence. - **Pickup:** the exact next action, specific enough to start cold ("next: X in file Y; watch out for Z"). - **Waiting on Josh:** open questions, if any. If a question BLOCKS the Pickup action, mark it **[blocking]**. ## Phase 4 — ship and release Josh - Commit the docs, push, confirm the push landed. - Confirm `CLAUDE.md` points the next session at the latest worklog (the newest file in the newest `docs/worklogs/<YYYY-MM>/` folder). If that pointer is missing, add it now — a baton nobody picks up is wasted. (The month folders ARE the archive — no sweep step exists.) - **Size check** (the drift guard, est. 2026-09-20): BUILD-ORDER's banner (the `>` block) must be under ~20 lines, and `CLAUDE.md` under ~20 KB. Over either → move narrative to its home (slice doc, milestone entry, `docs/APP-TOOLS.md`, `docs/RELEASING.md`) before committing. Report the sizes in the worklog State line only when something was moved. - **No-slice session?** Leave the BUILD-ORDER banner alone. If the session's asks were three or more related changes to one area, say in the worklog that the next one should get an S-number. - Final message: confirm it's safe to /clear, then give Josh a copy-paste opener for the next session. Pick by priority (first match wins): 1. **In-flight section exists:** `Resume in-flight work per the worklog: <one-line state summary>. Read the In-flight section before touching anything.` The next session must know it's landing in a hot zone, not starting a fresh slice. 2. **[blocking] questions exist:** `Before continuing: answer [blocking] items in today's worklog (<one-line summary of the questions>), then continue per the worklog.` 3. **Otherwise:** `Continue per the worklog: <one-line slice description>`
  17. When you get into using AI with large systems, it becomes clear you can't rely on it to remember important facts and decisions. It also becomes clear you have your own shortcomings that work against the AI, such as not being disciplined in your communications with it. And it becomes obvious that you need a system to keep yourself and the AI on track if you're to effectively co-create within a complex system together. This is a design problem. And if you can solve them well, it puts you far ahead of any random Joe with a Claude Pro subscription. A lot of people think it's all about the "prompt" and maybe the user memory. But those are just two tiny factors. It's mostly about gates and figuring out what happens at those gates, what is allowed to occur between them, and then designing the system so it operates within those contraints. We have a lot of tools at our disposal for this, such as hooks, scripts, skills, git hooks, agent memory, etc. When combined, these tools allow for massive problem solving creativity and AI system design. The above is a refactor(WIP) of my AI workflow for building apps. I just started refactoring it this week and will share the whole thing when it's done.
  18. TBD. I think we're good for another 2-3 years. Hopefully. Time to sharpen the sword.
  19. Yeah, me too. Maybe they did but he just hasn't been back yet?
  20. I agree. Cold. Transactional. “Bad for the brand”
  21. Maybe time for a new chapter.
  22. 100%. It’s not easy at all to do this effectively when you’re dealing with a complex business/system. I spend a lot of time on this. We should set up a thread for this topic and share ideas.