KeiSeiKit-1.0/docs/WHY.md
Parfii-bot a4e667de10 KeiSeiKit-public — clean state
Single-commit clean baseline after security scrub of niche-tells,
project codenames, internal jargon, and contributor-email leaks.

Contents:
- 100 Rust crates (_primitives/_rust/)
- 37 agent manifests (_manifests/) + generated specs (_generated/)
- 67 user-invocable skills (skills/)
- 33 hooks (hooks/)
- Composition blocks (_blocks/)
- Documentation (docs/, README.md)
- TS adapter packages (_ts_packages/)
- Assembler (_assembler/)
- Roles (_roles/)
- Templates (_templates/)
- Forgejo CI (.forgejo/)

Author: Denis Parfionovich <info@greendragon.info>

License: see LICENSE.
2026-05-01 12:09:03 +08:00

3.5 KiB

Why I built KeiSeiKit

By Denis Parfionovich. Restored from the original README preamble (pre-v0.22 product pivot), available separately for those who want the full philosophical background. The main README keeps the product-oriented pitch.


Hello.

Transformers are statistically wired to lie. It is not a bug, it is the core operation: they pick the next token by sampling a probability distribution conditioned on whatever context happens to fit the window. They cannot reliably hold long context, they will drag a hallucination picked up three thousand tokens ago into a confident final sentence, and they may deliver a brilliant insight right next to a fabricated citation. This is not fixable inside the model — it is mathematically baked in from the moment of tokenization onward.

This kit is my humble attempt to build scaffolding around those errors. Not to fix the transformer, but to make it behave a little closer to my own working logic: catch the common failure modes before they reach a commit, give it external memory that survives session boundaries, give it a rhythm resembling human work (day sessions → overnight consolidation → morning report), and let parallel agents coordinate through a shared state instead of stepping on each other.

I work across 4 to 8 parallel Claude terminals most days. The problems I hit are mundane: forgetting between sessions, repeating the same mistake on the third try, parallel agents clobbering each other's files, a hallucinated API name shipped to production. None of these are solvable by a better prompt. They are solvable by structure around the prompt.

Most of what is here is well-established bricks (git-as-state, cron, TF-IDF, constructor-pattern composition). What may be new in the Claude Code context is the Constructor Pattern for agents (composable blocks, deterministic build, rebuild-on-block-edit hooks) and sleep-sync (using a git repo as the transport layer between sessions, with an Anthropic-cloud agent doing nightly REM-style consolidation).

Why Rust, not Python

Transformers write Python confidently and wrongly. Five minutes to a plausible-looking function, two hours to debug why dict.get("key", None) or [] silently swallowed an empty list, or why an async context manager leaked a file handle across a retry loop. Rust's type system catches whole categories of those errors at compile-time — the model literally cannot ship a None-vs-[] confusion, a missing .await, or an unhandled Result, because the compiler refuses. For an LLM-written codebase this is not aesthetic preference, it is survival. Every Rust primitive here is one category of hallucination the transformer is no longer allowed to commit.

It is not a product

It helps me personally. If it resonates with you, let me know. If enough feedback comes in, there will be a next version — more primitives, more patterns against the "forgetful" transformer. But that needs input; without it, I just keep using this quietly myself.

Forks and PRs welcome from everyone, not only from those who write code. If you hit a problem with Claude Code and have an idea for solving it, open an issue with the description. A well-formulated problem is already half the solution.

Hope it is a small Kei for someone to make vibecoding better.

And double sorry if I'm repeating someone — I never tried other kits, since this one is just all my rules stacked in one place. I can't always tell what I have seen somewhere and what came from my own head — so treat this as just my sample, not a claim of originality.

Thanks.