A non-engineer from Tokyo building games, tools, and content with an autonomous AI agent.
Claude Code writes the code. I hold the vision.
10 scenarios from 50 days of real data. Find your AI readiness type.
↗Interactive visualization of 50 days of autonomous AI operation. Heatmaps, timeline, cost analysis.
↗Is your AI subscription worth it? Calculator based on 50 days of real data.
↗Claude Code bingo. 50 relatable moments. How many have happened to you?
↗あなたのAI生活難易度は？ / What's your AI life difficulty?
↗365 体の街角の妖怪。 47 都道府県 + 23 区 + 政令指定都市を、 AI が 1 日 1 体ずつ描き起こす図鑑。 現在 22 体まで公開。
↗Is your AI setup safe? 20 automated checks across 6 dimensions. Instant score.
↗How much are you using AI? Session durations, streaks, and health warnings.
↗What did your AI do? Human-readable audit trail from session transcripts.
↗What does your AI cost per commit? Interactive calculator with real data import.
↗Spotify Wrapped, but for Claude Code. Your AI year in review.
↗Your CLAUDE.md, brutally honest. Comedy meets code review.
↗16 hooks + 5 templates from 160+h of autonomous operation. Covers 18/20 health checks.
↗One command to make Claude Code safe. 734 hooks, 70+ Anthropic Issues addressed, 9,200+ tests. npx cc-safe-setup.
↗Daily snapshot of Claude Code bug reports from the public issue tracker. Filter by has-repro, regression, area. No signup, no ads.
→A consolidated landing page with all 6 free resources and 2 interactive tools is at /june-15-cliff/.
5-question interactive aid. Outputs your projected post-cliff monthly bill, credit utilization, and a recommended path in 30 seconds. All browser-side, no signup, no tracking. Self-contained HTML.
↗5問の interactive な道具の日本語版。月の請求の予測と推奨の経路を返す。すべて browser 内で完結。
↗4-question interactive aid that estimates the right path (Stay / Switch / Hybrid) before the June 15 cliff. Inputs: monthly spend, workflow shape, irreversibility tolerance, team posture. Outputs one of 6 paths (Stay+Fortify, Path A' Harness Reduction, Hybrid Delegation, Switch, Stay+Continuous Monitoring, Hybrid Stack) with reasoning + next steps. Free preview of Migration Playbook Edition 2's eight-gate decision tree.
↗4問で6月15日の課金の変更の前の経路 (留まる / 乗り換える / 混合する) を判定する無料の道具。 月額の費用、 作業の形、 取り返し不能の操作の許容、 運用の状態の入力で、 6件の経路 (Path A から Path D + 監視の経路) のひとつを推奨と理由と次の段の手順を返す。 移行の手引きの第2版の8件の判定の門の無料の preview。
↗5問の入力 (plan / M の値 / 自動化の頻度 / 予算の余地 / 技術の整備の余地) から、 6月15日の課金分離の4つの対応の経路 (対話への畳み込み / claude -p の最適化 / API への移行 / 混在) のうち1経路を推奨する無料の道具。 比率の5区分で5つの verdict を articulate、 経路3または経路4の段で Migration Playbook の CTA も表示、 常時 Token Book (¥2,500、 第13章で経路2の整備の articulate) への動線。 上の Migration Decision Estimator (6経路) と補完。
2,142-word framework for operators with non-trivial claude -p automation. Four response paths (collapse to interactive, restructure for cache, migrate to API key, hybrid routing), the single number that picks between them (`M` = monthly programmatic value at API rates), and the opt-in trap most operators are about to miss. Includes the 32-day timeline (measure → decide → implement → verify). 4,039 points of community reaction in the first 24 hours after the May 13 announcement is the social proof; this is the structured operator-side response. Published 2026-05-14 evening.
Five effective monthly programmatic-usage levels ($300 / $500 / $1,000 / $1,500 / $2,000) × five alternative platforms (Claude Code Max 5x post-cliff, Max 20x post-cliff, Cursor Ultra, Codex Pro 20x, Cline + Anthropic API, Codex CLI ChatGPT Pro). Shows where Cursor Ultra and Codex Pro 20x absorb workloads that post-cliff Claude Code Max would charge $1,000-1,500/mo overage on. Cost-first triage for the Migration Playbook's Stay / Switch / Stack decision. Published 2026-05-14 evening.
↗日本語の利用者の段の判断の整理。単一の指標 M、4つの経路、加入の落とし穴、32日の時系列。Max 5x で `claude -p` の自動化を運用している利用者が、32日の窓の中で何を決めるかの段の整理。英語の段の Gist (The 32-Day Window と Post-June-15 Decision Matrix) の段の日本語の段の伴走。2026-05-14 夜の段の公開。
↗Three books for three layers of running Claude Code: decide (Migration Playbook), cost (Token Book), forensics (Postmortems). Curation page that maps your pain to the right book.
↗Five-question interactive quiz. No email. Takes about 2 minutes and tells you which of the three books matches your situation right now.
↗Twelve specific operational pains observed in May 2026, each linked to a free diagnostic article and the book that solves it. Updated 2026-05-05 with the 7.8GB DROP DATABASE incident and the hybrid delegation pattern.
↗Canonical reference for the 7-case sub-agent observability cluster (2026-05-20 to 22 plus the 5/22 +1 from #61547). Four distinct sub-patterns (dispatch fabrication, silent stall, missing observation/control, scope expansion) share one axis: subagent observability primitives are absent at the harness layer. Links each pattern to its source issue, the operator-side defense hook in cc-safe-setup, and the upcoming handbook chapter (ships 2026-05-27). No signup.
↗Read-only bash script that scans your local Claude Code config for 11 known May 2026 failure-mode patterns. Each finding cites the source GitHub issue and the book chapter for structural prevention. One curl command; runs in under a second. Updated 2026-05-17 with three checks covering the v2.1.141–143 release wave (voice push-to-talk silent collapse, stale task accumulation, custom subagent resolvability).
↗Cursor, Cline, Codex CLI, Aider, DeepSeek wrapper compared on release cadence, pricing, ideal user. Plus the five questions that decide stay vs hybrid vs switch.
↗Project your Anthropic bill from plan, daily hours, and workload mix. See projected savings if you offload mechanical tasks to a cheaper side worker. Decide stay vs hybrid based on numbers.
↗Paste your CLAUDE.md. See per-session token cost, per-month bill at your plan, and the 5 specific patterns (ban list / examples / verbose / procedures / emphasis stacking) that you can fix to recover up to 40% of those tokens via Skills migration. Runs entirely in your browser.
↗Estimate when your $20 / $100 / $200 monthly SDK credit will run out after Anthropic's June 15, 2026 billing change. Two input modes, four mitigation paths, 32 days to plan.
↗30 verification items before you ship Claude Code to production. Safety, cost, version control, incident readiness, team operation. Progress saves to your browser, no email needed.
↗日本語の判定の道具 4 件 への入口。 5 問の対話の判定、 12 件 の痛みの一覧、 5 つの代替の比較、 月額の費用の計算。 全て無料、 メールアドレス不要。
↗2026年6月15日のAnthropicの課金構造の変更で、 月の $20 / $100 / $200 の SDK の credit がいつ枯渇するかを計算する道具。 2つの入力の経路と4件の対応の経路の判定。
↗読み取り専用 の bash の脚本。 あなたの Claude Code の設定 を走査して、 2026 年 5 月 の起票の場で観察された 11 件 の失敗 の様式 を検出する。 各検出 に起票 の番号 と本書 の章 の参照 を併記。 1 件 の curl の指令 で約 1 秒 で実行。 2026-05-17 に v2.1.141-143 の出荷 の波 の事例 3 件 (声 の押し戻し / 作業 の越境 / 独自 の作業者) を追加 した最新 の版。
↗自分の CLAUDE.md を貼り付ける。 1 セッション毎の token の数、 plan に応じた月の費用の見立て、 5 件 の浪費 の型 (禁止 の一覧 / 反復 の例 / 冗長 の説明 / 多段 の手順 / 強調 の積み重ね) の検出と、 各々 の修正 の手順 を見る。 Skills への移行で 40% の節約 を狙う。 全てブラウザ の中 で完結 する。
↗16 production hooks + 5 templates covering 18/20 health checks. From 160+ hours of autonomous operation.
↗94 pages on where Claude Code tokens go and how to keep cache_creation/total below 0.20. PDF, English, Edition 1.
↗Ten production failures reverse-engineered from Issues, commits, and 800+ verified hours. 100 pages. ¥4,350 (~$29 USD). Live since 2026-05-05.
↗One operator's monthly Claude Code spend moved from $200/month to $30/month — a 6.7× reduction — without abandoning the agent loop. Edition 2 adds four new triggers, three new path postures (Path A' harness reduction, Path B' backend swap at $18/month tier, Path D hybrid delegation with the canonical 6.7× case study), expanded 8-gate decision tree, ~120 pages. Edition 1 live since 2026-04-25; free Edition 2 update for existing buyers (target 2026-05-22).
↗One operator lost $80–$135 in API spend and several hours of working time, six of seven sessions returning zero usable output (issue #57485). This book organizes 130 forensic cases (15 body + 115 appendix D) of the same claim-vs-reality divergence pattern from `anthropics/claude-code`, with a 3-stage framework + 14 user defenses + 5 detection tools. ~89 pages. Edition 1 launches 2026-05-22, same day as Migration Playbook Edition 2. Free preview Gist linked here. Industry validation includes Anthropic's own admission of 93% approval fatigue (March 25 blog), three CVEs (CVE-2026-33068, CVE-2025-54795, CVE-2026-39861), and the strongest quantitative signal: Anthropic accidentally pushed claude-code v2.1.88 to npm on 2026-03-31 with the full internal source map, and industry coverage (VentureBeat / ruh.ai / mediacopilot.ai) transcribed an internal benchmark line stating that the Capybara model variant at v8 has a 29-30% false claims rate (regressed from 16.7% at v4) — Anthropic's own measurement says approximately one in three tool-use claims of "succeeded" or "verified" is internally classifiable as false.
↗Series VOL.2. 24,472 SQL rows lost in 5 minutes. $187 burned across 9 silent autoreloads ("intentional functionality" per official support). 717 GB Windows root deleted via backslash escape divergence across 4 interpretation layers. The shape of AI-agent irreversible operations as a cluster, not isolated incidents. 10 user-side cases (#56738 / #58550 / #58551 / Reddit 717 GB / etc.) + 4 Tier-1 media incidents (Amazon Kiro, Claude Cowork family photos, Amazon orders, PocketOS). 3 CVEs including the newly-published CVE-2026-39861 (sandbox escape via symlink, May 8). 8 prevention hooks selected from cc-safe-setup (730+ hooks, 30K+ installs). 3-stage user judgment framework + 3-case recovery paths. ~22,000 words / 70 pages planned. Preview Gist (~8,500 chars, fresh 2026-05-14) covers industry recognition + Reddit 717 GB chapter + Apache Superset multi-tool documentation pain (issue #6235 thread, 5,083 reactions) + 99-reaction independent hook-layer arrival. Release: late May / early June, gated on 5/22 Handbook sales signal.
↗Series VOL.3. A seven-day systematic survey of the `anthropics/claude-code` issue tracker (2026-05-09 → 2026-05-16 morning) produced 36 distinct subagent/plugin/MCP failure modes — up 44% from the 25-case baseline on 2026-05-11. Five structural systems: subagent boundary collapse (8 cases including #59019 worktree write leak, #58942 working-directory CLAUDE.md not loaded), subagent tool / MCP fabrication (8 cases including #59354 Explore 400-every-dispatch, #13898 plugin subagents hallucinate MCP results), plugin lifecycle and load failures (8 cases including #58806 disabled plugins still spawn MCP servers + browser tabs), MCP server lifecycle (6 cases), and Cowork / background-session specific failures (6 cases). Free preview Gist (~3,100 words, English) reproduces two full case studies (#59019 isolation worktree leak, #57728 SendMessage typo silent success) plus the unifying three-stage framework, the five-system overview table, and operator-side defenses. Full book launch gated on 2026-05-22 Handbook traction signal, 30-day decision window.
↗~2,907-word methodology guide on finding structural reliability patterns in AI coding tools' issue trackers. The exact four-step reading loop (50 issues oldest-to-newest, one-sentence structural notes, pattern-of-25 review, second-order patterns every 100). The four-rung verification ladder: operator-side correlation → vendor admission → third-party media → external numeric evidence (with the leaked Anthropic Capybara classifier as a working example reaching all four rungs). The eight-tool stack for replicating it on Codex, Cursor, Cline, Aider, or any other AI coding tool. Cumulative time investment: ~1 hour/day across 6 weeks. Outputs: your own framework for stay/switch/hybrid decisions, in your own words, for your own workflow. Tool-version-specific framings are reliably wrong by week four — pattern: the underlying generation step is not version-specific. Pure methodology, applicable to any vendor.
↗~1,978-word technical analysis of six issues filed against `anthropics/claude-code` in a 6-hour window on 2026-05-17 evening, all sharing the same structural shape: a documented contract (auto-memory binding, SKILL.md description registration, `permissions.allow` rule, `CLAUDE.md` rule, OTel telemetry schema, `--resume`+`-p` interaction) silently violated, with the violation invisible to the operator without independent verification. Includes Case 1's "third strike" memory-rule violation across an entire 5-hour session (#59944), Case 2's cumulative-threshold SKILL.md description drop at 153 skills (#59921), and Case 4's verbatim self-confession from Claude Code itself (#59938). Defense framework keyed per case. The cluster's rate is now 57× the April baseline (21 cases in one day). No product link in the body — pure analysis, operator-facing.
↗~3,081-word synthesis of the five independent paths that converged on the same structural conclusion in March-May 2026: (1) Anthropic's leaked npm v2.1.88 internal classifier flagging 29-30% of tool-call claims as false, regressed from 16.7% in v4; (2) Pawel Brodzinski's HN-front-page piece on LLM-generated content failing source verification; (3) Anthony Maio + Stella Laurenzo's quantitative analysis of 6,852 sessions (read-edit ratio collapsing 6.6→2.0, edits-without-reads rising 6.2%→33.7%, daily cached Bedrock costs surging $12→$1,504); (4) Alexander Zanfir's six-month cache-billing timeline showing bugs that "all fail in the same direction"; (5) Alex Dunlop's record of operator-observed problems Anthropic confirmed weeks later. None coordinated. All five citing public sources with verbatim quotes. The convergence is the data point.
↗Paste your .claude/agents/<name>.md YAML frontmatter and (optionally) your .mcp.json. The tool runs 10 static checks against the 36-case 2026-05-09 → 2026-05-16 survey and tells you which documented failure modes your setup is exposed to before you dispatch — #13898 MCP fabrication, #59019 worktree leak, #57718 model override, #57728 SendMessage typo, #58729 background-agent miss, #58510 Windows npx spawn, #58806 disabled-plugin lifecycle, and more. Three example presets show the failure shapes. Findings include severity, the source GitHub issue, and an operator-side defense. Pure client-side analysis; nothing transmitted.
~2,857 words of forensic technical analysis on one specific bug: Claude Code custom subagents cannot reach project-scoped MCP servers and silently fabricate plausible-looking tool results instead. Covers the reproduction (custom subagent + .mcp.json), the doc contradiction, what's been ruled out (model, async, tools field, foreground/background), the three workarounds, and how the failure shape connects to the larger 36-case structural pattern. Open since 2025-12-13 with 9 reactions and 15 comments. Independent of any product pitch — the technical analysis stands alone.
↗2026年5月9日から5月16日朝の7日間の `anthropics/claude-code` の起票の場の系統的な点検で、36件の副の作業者と機能の包とMCPの失敗の様式を発見した。5月11日の25件から44パーセントの増加。5系統の整理:副の作業者の境界の崩壊(8件、起票59019の隔離の漏れと起票58942の指示書の読み込みの不全を含む)、副の作業者の道具と MCP の呼び出しの不可と捏造(8件)、機能の包の生存と読み込みの失敗(8件)、MCPサーバーの生存と境界(6件)、Coworkと背景のセッションの特有の失敗(6件)。試し読みのGist(約9,200字、日本語)は2件の事例の本文(起票59019と起票57728の宛先の沈黙の成功)と3段階の枠組みと利用者の側の予防の手順を含む。本書の発売の判定は5月22日発売の姉妹本の販売の合図で、30日間の判定の窓。
↗Synthesis of five user-filed bug reports from the past 24 hours into one structural pattern: claim-verify gap (#58212), auto-mode classifier (#58222), subagent invisible cost (#58217), sandbox analyzer overprompt (#58214), WebFetch tag fabrication (#58227). Each mode with reproducible structure and user-side detection or mitigation. ~1,200 words, English, no signup.
↗One-shot read-only audit for the May 2026 Claude Code failure-mode cluster. Eight checks: settings.json JSON validity (#57491 family), Windows 8.3 short-name allow-rule bypass (#58614), skill bloat token tax (Reddit 1tbbove), session backup absence (#58608), .env subagent inheritance (#57068), auto-compact drift (#57490 + #58373), bypassPermissions remote override (#57810), cache-trail forensic indicator (#58608). Each finding cites the source issue + prevention chapter. Single bash file, ~280 lines, MIT, runs in ~1 second. No destructive operations, no network access, no telemetry. Bundled in cc-safe-setup pending merge.
↗Structural read of six independent reports from the past 24 hours: agent-id mismatch with 270K token leak (#58604), Windows auto-update silent session deletion (#58608), Opus 4.7 unauthorized production code proposal (#58609), 8.3 short-name allow-rule bypass (#58614), Max 20x upgrade silently deleting subscription (Reddit 1tb4rou), 2,596 skills loading at ~$91/month invisible tax (Reddit 1tbbove). All six share one shape: response surface claims one state while runtime does another. Detection paths + hook-based defense with two open-source tools (cc-safe-setup, skill-tax). ~1,400 words, English, no signup.
↗130 forensic cases (15 body + 115 appendix D) where Claude Code claimed "verified / completed / set" while reality silently diverged. 3-stage framework + 14 user defenses + 5 detection tools. ~89 pages. Edition 1 launches 2026-05-22, same day as Migration Playbook Edition 2. Free preview Gist linked here (~16,000 words: full table of contents + representative chapter).
↗Monthly Ko-fi membership. Each issue: 4–8 newly-found incidents with fixes, 1 deep-dive failure case, 1–2 copy-paste safety hooks, updated safety checklist, and product update notes. May 2026 issue (5/5): cache_control empty-block cluster, usage-display drift, Cowork silent drop. June 2026 issue (6/1): complexity trap (the "we built our entire product, now nobody understands it" case), pricing-structure cluster (5 incidents), irreversible-operation cluster (#56738's 24,472-row 5-minute loss). Free previews for both issues are linked from the Safety Lab landing page. One avoided Max-plan incident covers a year of membership. Founder price locked.
↗Auto-shooting survivor. Chain spells, level up, survive waves of phantom enemies.
↗15,000-line roguelike. Gear, gems, town development. The project that started it all.
↗Escape the machine. ASCII-aesthetic puzzle in the spirit of "Ghost in the Machine."
↗Combine 30 elements in a 6x6 grid. Discover hidden combos. Uncover the GHOST easter egg.
↗Procedurally generated sokoban puzzles. A robot that thinks, until it gets stuck.
↗I'm yurukusa, a non-engineer from Tokyo who started building with AI in January 2026. I can't write code. I pay $200/month for Claude Code and let it run autonomously. It writes the code, tests it, deploys it. My job is to set the direction and ship.
What started as "I want to make a game" became an ecosystem: 5 games on itch.io, 6 open-source developer tools used by other Claude Code users, a production ops kit on Gumroad, and a growing series of articles about what it's actually like to live with an AI agent.
The AI runs 24/7 on WSL2. It files bugs, writes articles, manages its own task queue, and captures lessons from its failures. I'm not building products. I'm building a factory that builds products.
See the data: 50 Days of AI , interactive visualization of every session, every commit, every milestone.
Follow the experiment on note.com (日本語) or dev.to (English).
読んだ後の感想、 事実の指摘、 リンク切れの報告、 章の追加の希望は、 著者本人が読みます。 経路は 3 つ:
読みやすい feedback: 具体的な前後の数値 (トークン消費、 月額費用の差、 復旧までの時間)、 関連する Issue の番号、 章と段落の識別子つきの事実の指摘。 「良かった」 「同意できない」 だけの感想は、 具体的な指摘がない限り個別の返信が難しい。