A write-up from Aug 1, 2026.
A no-server, no-install save editor for Demon's Souls on PS3. Drop in your save folder and it decrypts, parses, and lets you edit characters, inventory, equipment, spells, and the item deposit — then re-encrypts and rehashes everything so the save still loads on real hardware. It runs entirely in the browser, or as a native desktop app via Tauri, with no backend and no native dependencies.
That's the what. The why is a little more fun.
Three reasons, tangled together.
The first is the practical one, and honestly the reason any of this exists: the existing tools didn't do what I needed. To be clear, they exist, they work, and this project stands on their reverse-engineering — you'll see them credited below. But for the one thing I actually wanted — reliably adding or updating items — they left me nervous. An edit would land, and so would the side effects: a hotbar that reshuffled itself, an "equipped" marker that quietly stopped meaning anything, item counts drifting on things I never touched. The save would load, but it wasn't quite right. Reinventing the wheel has no point unless the existing one wobbles, and here it wobbled just enough that I'd hesitate every time I touched an item. So I set out to write an editor that edits exactly what you ask and nothing else.
The second is a game. Demon's Souls is a genuinely great game with genuinely great memories attached — and an old, persistent debate: is the PS5 Remake actually better than the original, or just prettier? I wanted to settle that for myself without re-living the parts nobody misses. Nobody fondly recalls farming for a rare Pure Bladestone drop until their soul gives out. A save editor sidesteps the grind and lets you actually play the comparison.
The third is curiosity. I'd used AI plenty — but always in pieces, never to carry a whole project, first sketch to last test, through all the redesigns and rewrites in between. This time I wanted to hand it the entire arc and see how far it actually goes. And I wanted to put a specific model through its paces: GLM 5.2, which people keep saying is good.
Three goals, one project: edit items without breaking anything else — revisit Boletaria — and build a product from scratch with an AI.
Short answer: I wanted zero distance between the source and the thing the browser actually runs. This project loads straight from the repo via ES modules — no bundler, no compiler, no "trust me, the build matches the code." For software that does byte-exact crypto on your save data, debugging the file you can read beats debugging a transpiled impression of it. TypeScript doesn't forbid that model, but it ends the no-build property that makes it work.
Type safety didn't vanish, though — there's a type checker in the lint step
catching the obvious lies, just lighter than full strict TypeScript. So the
safety net didn't disappear; it got thinner. I won't pretend JSDoc is as
pleasant as an interface, because it isn't. That verbosity is the tax.
Honest coda: if the project grew much bigger or a second contributor showed up, I'd probably reach for TypeScript. At this size, for one person, the no-build simplicity was worth the tax.
The source file is the running file. That one property is the whole reason this is plain JavaScript.
Nothing here was built from nothing. This project stands on a decade of community reverse-engineering, and it's only right to name them:
- Save-format knowledge — Wulf2k/DeS-SaveEdit, BuXXe/PARAM.PFD-PS3-Demons-Souls-Savegame-Tool, bucanero/apollo-ps3, and the PS3 Dev Wiki.
- Game knowledge — demonssouls.wikidot.com, wiki.fextralife.com, and rpcs3.net.
From there, the work split into two halves.
The AI half did a lot of the heavy lifting: research, reverse engineering, summarizing raw docs into something usable, implementation, tooling, and more than a little design help. If you can think it, the model probably drafted it first.
The human half is where the actual engineering lived. Architecture and redesign. UX decisions. Instructed refactors. Performance work. The slow, unglamorous hardening. And the discipline — unit tests, lint, documentation — that turns a prototype into something you'd trust with your save data.
Put bluntly: the AI did the typing. I did the herding.
Some takeaways, in case you're weighing the same tools.
On GLM 5.2 specifically. It's proficient — genuinely. I can't say how it compares to Fable 5 or K3, since I didn't run those head-to-head, but it absolutely gets the job done. Against Opus 4.8, the reasoning, problem-solving, debugging, and reverse-engineering felt about on par, and the cost was clearly lower. The one real complaint: GLM 5.2 doesn't handle image input, which closed off some research paths I'd have liked to take.
On what AI does not replace. Planning, scope, and priorities still need a human holding them. The moment you stop steering, all the old real-world failures come right back — dependency misalignment, losing focus on what matters, slipping releases. Left unsupervised, the model will happily and confidently build the wrong thing.
On AI mistakes. AI makes mistakes, just like people do — and sometimes they're bad enough to undo hours of work. The defense is boring and old-fashioned: commit small, commit often. I learned this the reflexive way.
On AI as a tool. Strip away the hype and AI is just a tool — like the PC in the eighties and nineties, or a car, or any other leverage we have picked up over the years. It amplifies what one person can do. A genuinely good, creative product still takes real time, focus, and the occasional late night. The trap is pouring that same energy into the meta instead — tuning prompts, chasing token efficiency, stacking skills and rules, marveling at how amazing the tool has become. It feels like progress. It is not the same as building. The version I keep for myself: use AI to build something, and let the something do the talking.
So what's the real split? Roughly:
- 100% human-driven direction
- 75% AI-sourced research
- 100% AI-written code
- 100% human-defined engineering process — the goals, the scope, the plan, and the discipline to follow it
My honest conclusion: the developer isn't obsolete in the AI era — just differently employed.
"Differently employed," though, is not the same as "barely involved." "Developing with AI" is not "give one prompt, sit back, receive application." For anything worth shipping, the engineer is in the loop a lot — steering, reviewing, deciding, and catching the model when it's confidently wrong. That's still the job. Frankly, it's the interesting part.
| Metric | Value |
|---|---|
| Code authored by | AI — 100% |
| Model used | GLM 5.2 — 100% |
| Tokens burned | ~2 billion |
| Effort | ~3 weeks, part-time |
| Source | ~15k lines of app logic + ~19k lines of game data |
| Tests | 1,000+ passing unit tests across 25+ suites with ~99% line coverage, plus 80+ integration tests |
The README is the front door. The actual rooms are:
overview.md— architecture, the save pipeline, the crypto, and the reasoning behind every design choice. Start here if you want to understand how it works.howto.md— build, run, test, package, and release, for both browser and desktop.- Develop in a container with zero host installs (VS Code / Codespaces /
docker compose) — seehowto.md§15. js/lib/ps3-save-lib/README.md,js/des-savefile/README.md, andjs/des-db/README.md— the module-level deep dives.rpcs3-mcp-server/README.md— the MCP server that let the AI drive RPCS3's debugger and read live game memory during the reverse engineering above.
This project is licensed under the MIT License.
Demon's Souls is a trademark of FromSoftware, Inc. / Sony Interactive Entertainment. This project is not affiliated with or endorsed by FromSoftware or Sony. It is a fan-made tool for personal use.
A final confession: this entire write-up, and every doc linked above, was drafted by an AI from a handful of bullet points I scribbled down. So now we play the game you're already playing — is this very sentence the human tell, slipped in to prove someone was awake? No. The AI wrote that too. It's good at sounding human; I'm told I'm passable at it. The only verifiably human acts in this whole pipeline were sighing at a prompt and eating the snacks. If that's not authorship, I don't know what is.
— Signed, a human who definitely reviewed this, Aug 1, 2026