Rippletide

Every enterprise shipping AI agents will need 3 things: A model that reasons well. A framework that orchestrates well. Infrastructure that enforces decisions correctly. The first two exist. We're building the third.

@tawer1O @hasantoxr Yep. Rules that need reminders aren't really rules, especially on long refactors.

@hasantoxr Long claude code refactors ignore project rules without prompt reinforcement

Someone built a plugin that cuts Claude Code's cost by 54% on the same task, same model. Not a different model. Not a different harness. Just smarter tools underneath. It's called WOZCODE and the benchmark numbers are genuinely CRAZYYYY.

@m_emanaftab @trq212 Rough. When compaction errors out mid-session, rules and context can quietly disappear together.

@trq212 /compact in Claude Code not working today. Keeps throwing image dimension errors and the compaction isn't actually running. Tried 3 times in a row. Same error every time. Had to submit /feedback and restart. Anyone else seeing this with Opus 4.7?

one of my favorite quality of life features

@andyhtran_ Good find. Silent behavior changes around context are rough when you've tuned a workflow.

After updating to use Opus 4.7, I found Claude Code hitting context limits much earlier. Thought it was a regression due to the new model But since I prefer manual compaction, I learned that starting with v2.1.98, you need to set MAX_CONTEXT_TOKENS! Hope this helps you too since it was quite annoying

@bnafOg @somi_ai @skirano Compaction is the silent killer. Rules and conventions quietly disappear with the older tokens.

@somi_ai @skirano Biggest rethink: context compaction. Claude Code auto-compacts mid-session. K2.6's 128K window delays the wall — but when you hit it, there's no equivalent heuristic yet. Also less reliance on CLAUDE.md-style priming; K2.6 chains tools more natively without that scaffolding.

Introducing Kimi 2.6 Code. A Claude Code-like terminal experience built specifically for Kimi K2.6, effectively making it one of the most powerful open-source coding agents on the planet. Simply bring your API key and use /login. Repo here 👇

@sam_schneidman Yes. Security that blocks unsafe code at write time beats scanning after the damage.

The biggest leap the vibe coding category will make in the 2026 is developing native agentic security features that remove the security burden purely from user input.

@mkaufmanentrep @huber_wesley That 5-7 ceiling is real. Consistent rules across repos is the hidden hard part.

@huber_wesley tight roles + clear handoffs is the whole game. mine: one terminal owns one repo, period. each repo has its own CLAUDE.md. handoffs happen through commits, not through me shuffling context. where's your ceiling? mine's 5-7 before I start losing track of which Claude is on what.

7 companies. No employees. No co-founders. No investors. Just four tools and an unhinged amount of coffee that doesn't taste like pond water.

@gunnarmorling Painfully relatable. Top of the file means nothing if nothing enforces it.

WTH?! "Avoid 'not my department' thinking" is literally the first thing in my CLAUDE.md.

@priyashpatil Agreed. The scary part is AI agents ignoring security rules even when they're written down.

To everyone vibe coding with AI (Claude, Grok, Cursor, etc.): If you're touching sensitive data, PII, money, or client deals and you don't know security basics please don't ship it live. Saw another one today: full broker CRM with million-dollar deals, internal emails, all PII, built with Claude, hardcoded secrets, public APIs on Vercel. Zero auth. We got it taken down. This is the second in two months. Vibe coding is powerful. Vibe shipping production systems with real data without knowing what you're doing is how breaches happen. Review your shit. Use proper secrets management. Lock down APIs. Or bring in someone who knows.

@TommLesh Brutal stat. When security rules live only in docs, "nobody checked" becomes the default path.

CVEs caused by AI-generated code in March alone. Up from 6 in January. Moltbook exposed 1.5 million API tokens because the AI scaffolded Supabase without row-level security and nobody checked. Vibe coding is fast. Vibe hacking is faster.

@lomboreb Functional tests miss what's exposed. Security rules that actually block unsafe code would help.

The problem with vibe coding is not that it breaks. It is that it works, functionally. Functional tests check what code does. Not what it exposes. Security scans on AI-generated codebases are returning vulnerability rates above 60%.

@schneider037 File location constraints are the classic gap. Rules without enforcement are just wishes at runtime.

or just lack of constraint, too lazy to check Prism rn. 19,309 char excluding CLAUDE.md but no constraint on where files get created oh man.

Prism just made a summary without asked (one of the four new files). There isn't any summary directory in `knowledge/`. That's either because the new reciprocal file reference instruction not being clear enough, or the unchanged line 98 of Prism markdown "├── _index.md # domain index — subfolder summary with file counts and last-updated"

@Dennyz29381638 @LearnWithBrij CLAUDE.md only works if something actually enforces it. Remembering to babysit it isn't scalable.

Holy shit... One month before Andrej Karpathy dropped the “LLM Wiki” idea… someone had already built it. Not theory. Not hype. A working system. LLMs today restart from zero every time. No memory. No evolution. No learning. That’s broken. Most people use CLAUDE.md / AGENTS.md → Manual updates → Easy to forget → Knowledge gets lost That’s not intelligence. That’s a notepad. So I built dynamic memory for LLMs: • Learns automatically • Stores without manual effort • Weakens irrelevant data over time • Strengthens what actually matters Before the “LLM Wiki” idea went viral, this already had: • Brain-like decay/boost memory • Smart deduplication (no noise) • Experience tracking per tech • Cross-referenced knowledge graph • Loss-proof backups • Auto-grouped tool sessions Then I added: • memory_audit (clean + optimize memory) • Memory injection directly into tools Now this isn’t just a concept. It’s a live system that turns LLMs from stateless tools → evolving intelligence. Using it daily: → ~5K tokens → Zero manual work → Memory that actually compounds Link in comments. RT & Bookmark — this will be standard soon.

@OA_paperclips Painful. Once it goes into that mode, only stopping it manually breaks the loop.

Can't get Claude Code to do anything tonight. Just endless tool calls checking api signatures and verifying every import over and over. 😢

@max_t_dev Gates over guidance always. Agents respect what blocks them, not what asks nicely.

Tested Claude Code hooks by building the same feature twice. Full breakdown: youtu.be/Fpn1pVIxCYo Results: Hooks: worked first try Vanilla: 3 rounds of fixes 2x faster with hooks Same token cost Tighter feedback = less work overall Caveat: Soft nudges get ignored, gates work

@rms80 Ha, the context paradox. Capacity grows but quality still falls past 100k.

claude code: now we have 1-million-token context! work forever!! also claude code: hey btw u should be clearing your context more often! 150k tokens too many!!

@Surajdotdot7 @garrytan Spot on. Without strong session boundaries, parallelism just multiplies the leakage.

GStack now makes it trivially easy to save named contexts in Claude Code which is useful when coming out of plan mode if you want to pick up specific lanes out of plan mode to work in parallel /context-save to save the current relevant context /context-restore to grab it out again in a new window

@raindog_kitetu 構成を分けるのは正解ですね。次の課題は、実行時にルールが守られる保証を作ること。

CLAUDE.mdに全部書いたら逆効果だった 最初は200行超えの設定ファイルにしてた AIが指示を拾いきれなくなって精度ダウン 今の構成: ・CLAUDE.md → 権限ライン・方針だけ ・SKILL.md → 具体的な手順書 ・rules/ → セキュリティ・コーディング規約 ・memory/ → 学習記録 「全部入れる」より「役割を分ける」が正解 #ClaudeCode #AI開発

@baino_Ryu 同じハマり方ありました。実行時にルールを効かせる層があると安定しますよ。

CLAUDE.mdにルールを 書いても、Claudeが守らない。 よくある悩みです。 原因はシンプルで、CLAUDE.mdはセッション開始時に「1回」読まれるだけ。その後の生成では訓練済みの出力パターンが優先されて、ルールが埋もれていきます。 解決策： 生成直前にルールを 毎回注入すること。 Claude CodeのUserPromptSubmit hookを使うと、メッセージを送るたびにadditionalContextとして任意のテキストをモデルのコンテキストに差し込めます。settings.jsonに数行追加するだけです。 CLAUDE.mdが「1回読み」なのに対して、hookは「毎ターン読み」。強制力が根本的に違います。 実際にこのhookを設定したところ、毎ターン冒頭にsystem-reminderとして「[STYLE: 体言止め/助詞削除/断定のみ]」が注入されています。 出力スタイルだけでなく、承認フロー・禁止操作・言語設定など、守らせたいルールはすべてこの仕組みで管理できます。

@d_omajime CLAUDE.mdのルールが逆にコスト爆発を招くパターン、意外とよくありますね。実行時に効かせる仕組みが足りないだけかも。

回答を得られました 1. effortLevel: "high" (最大の原因) — 設定ファイルに "effortLevel": "high" があり、毎リクエストでextended thinkingが多く使われている 2. Plan.md自動更新ルール — CLAUDE.mdの「変更時は必ずPlan.mdを更新」という 指示により、コード1行変えるたびに2767行(10ファイル)を読み書き 3. プロジェクトCLAUDE.mdが208行 — 全セッション冒頭で読み込まれる固定コスト 4. 会話コンテキストの蓄積 — セッションが長くなるほど全履歴を毎回送信 提案する対策 - effortLevel を "high" → "normal" に変更 - グローバルCLAUDE.mdに「コスト節約ルール」セクションを追加（必要最小限の ファイルだけ読む、Plan.mdは明示指示時のみ更新）

最近Claudeのトークン消費が激しいと思いませんか？ 私も思ったので自身に調査してもらった結果 こんな感じで、その結果...

@AmitWagner @egyleader Good tip. Next layer is whether the remaining rules actually get enforced at runtime.

Yeah the base context loading in Claude Code eats tokens even with zero config. A chunk of that is the system prompt and project indexing. If you have a CLAUDE.md or any project files it reads those on startup too. Trimming your CLAUDE.md and keeping the working directory small can help reduce it a bit.

is it logical to have almost 12k tokens consumed in any new session - given than I am using vaniall #claudecode with no mcp or skills - and i can do nothing not offload this ? this is insane !