いっちょー＠非エンジニアのアプリ開発記録

Claude Codeのセットアップガイドを新UIに全面対応しました✨ 非エンジニアでも迷わないMac/Win両対応の全7ステップを、動画でツアーしてみました🙌 仲間が増えてほしいので、ぜひ参考にしてみてください☺️ ▼ icchooo.github.io/claude-setup-g… #ClaudeCode #非エンジニア #AIアプリ

X投稿5週間目のご報告です📊 フォロワー68→119人(+75%)🎉 100人突破！😆🎊 無反応だった濃霧の中から少しだけ抜け出せて、健全に伸びてい…る…？？という感覚です🫣笑 開発の中身、進捗よりも自分の心境をあけすけに吐露した投稿が跳ねました⏫️ やはり魂の叫びは伝わるものですね😆 #振り返り

@shakebata しゃけばたーさん、０→１達成おめでとうございます！！🎊🎉 大きな一歩ですねっ！🙌✨️

77,526分。 4月からnoteを始めて、54日と20時間6分。 何もなかった初心者でしたが、 ついに記事を購入していただけました！ ずっと「本当に意味あるのかな」って不安だったので、0→1を達成できたのがすごく嬉しいです。 今日はちょっとだけ、自分を褒めたい。 そしてこの嬉しさを、どうしてもここで報告したかった！ 副業初心者の皆さん。 派手じゃなくても、積み上げはちゃんと力になると思います。

@FlyingDevLab フォローありがとうございます☺️✨️ 今後ともよろしくお願いします！🙇‍♂️

@icchooo_AI いっちょーさん、初めまして！ 返信ありがとうございます。 お笑いもいいですよね。 面白いところは繰り返し見てしまいます。

仕事中に音楽とか聞きますか？ 私は「勇者ヨシヒコ」などのドラマを聞きながら作業してます。 音楽を聞いている人は、もしよかったら何を聞いているか教えてください！ #個人開発 #indeidev

@Mindsync_log これ見た時一瞬バグかと思うほどMASHROさんバズってると思いました🤣笑 信頼は一瞬で積み上がるものではなかなかないですからね😃自分の中の精一杯をこれからも届けていきましょう😊

これ、自分に刺さった。 正直Xを初めてこんなに反応をもらえたのは初めてで驚いてる。 昨日の98万インプは、規格外さんが長年積み上げてきた信頼のおかげで、自分の歯科の知識を乗せてもらった形で見てもらえた。 だからこそ改めて思う。 自分も何かしらの信頼を積み上げていかないと、いざ波が来ても乗りきれない。 まずは自分の持ってる知識を言語化して積み重ねること。

@kbt_sjf @techcrunch 小鳩さん初めまして！フォローありがとうございます😭✨️ AIを使って個人開発に取り組んでいるいっちょーと申します🙏 見つけてもらう▶選んでもらう▶継続利用されるのそれぞれがかなり高い壁ですよね😅 活動を続け、地道に自分の認知を広げて行こうと思います🫡今後ともよろしくお願いします🙇‍♂️

AppStoreはAI製アプリの急増でかなりカオス化してるらしい。AIで誰でも作れる時代になったけど、これからは「作れる」より「継続利用される」が超重要。 The App Store is booming again, and AI may be why techcrunch.com/2026/04/18/the… @techcrunchより

@2Zyuragi 揺らぎ〜さん初めまして！リプありがとうございます🙏 お前のことが心配で言ってるんだぞ。系の説教ってAIも言うんだと思いました🤣 僕はずっとClaudeを使っています😃このやり取りはチャットの方で起こりました！

@icchooo_AI 「心配してるんです」が面白いですね。ちなみにモデルは何を使われてますか？

超小心者の自分は作成した投稿文をよくAIに 「これでどう？」「炎上しない？」「言い回し変かな？」 などと聞いたりするのですが、昨晩ついに見放されました😇笑 そんなイライラしなくたっていいじゃんかぁ🥹

@Mindsync_log 早く寝ろ！は僕もめっちゃ言われます🤣 時間はしっかり空けてるんですが、連続で壁打ちしてると思ってるんですかね？？🙄

@icchooo_AI これ面白いですねw 僕は早く寝てください！早く言われますw

@yuto_kkk5 KOUさん初めまして！フォローありがとうございます😊 AIを使って個人開発をしているいっちょーと申します🙏 前提が揃ってなかったら論争の落としどころなんて見つかるはず無いですもんね😇そりゃあ白熱するかという🤣 けどこれがネットの面白いところだなぁと思ったりもします😁

コード読む読まない論争、結局責任が取れるかどうかだと思った。趣味や個人利用ならAI onlyで全く問題ないしそれがbestまであるが、受託や業務開発では被害が出たら責任を問われる。本来は後者でどこまでAIを使って攻めるかが面白い議論なのに、前者と混ざってることが多い。

@kuragechan_1002 クラゲちゃんさん初めまして！フォローといいねありがとうございます🙇‍♂️ AIを使って個人開発に取り組んでいるいっちょーと申します😊 学習記録、軽く見させて頂きました👀✨️ 僕は非エンジニアで、こういうの全く分からないので教えてくださいっ！！🙏笑 今後とも仲良くして頂けたら嬉しいです！🙌

JavaはVScodeで書いてたのでeclipseがあまりつかえませぬ eclipseなくても大丈夫(乂'ω')ﾉｰ

@Aki_Works729 伝言ゲームの情報量が膨大なので多少ずれるのは仕方ないかもですが、あまりにもの時ありますもんね😂 神アプデしてくれるのを待ちましょう🙏😇

@icchooo_AI AI達の次の自分達への引継ぎもうちょっと改善できないんですかねwww 説明し直しがホントに時間の無駄すぎて🤣

コンテキストウインドウが圧迫されて話が通じなくなる前に次のルームへここまでやったから次のルームではここからやろう！みたいなまとめ文章作らせてるのに次のルームで、ではここのためにこれをやりましょう！みたいな方針の転換ばかりなのめんどくないです？

@Den_scoth_Mn いえいえ！全く礼には及ばないです！👋 相談とかも乗ってあげたいのですが、feinさんの投稿見ても高度すぎて何やっているか正直分かりません🤣笑 でもビールとおつまみは頂きまーす😍わーい🍺✨️

@icchooo_AI AIエージェント体験はいっちょーさん達、フォロワーさんの背中押しがあってこそ☺️ その節はありがとうございました✨ ものすごい議論になっているようで、想定を遥かに超えてますね😮 いっちょーさんもいっしょに観戦しましょ！ 場所取りもしてあるし、ビールとおつまみ買っておきましたよ🍻

バイブコーディング🆚人力コーディング流行ってるね😆ww まぁええんでないの❓️ 本人が困らないなら、どっちやってようが 私は両方やるけど、どっちみち趣味だからね🎵 でも観戦してるのはおもしろい🤣ww #個人開発

@murakumo_dev 世はAI乱世を極める混迷の時代よ。仕方あるまい。 ちなみにこのやり取りもちょっとそうですからね🤣笑

@icchooo_AI いい歳して黒歴史が生み出されそうになるとは思わなんだ

ツイート文書いてるとついカッコつけたこと言いたくなってしまう 抑えろ...自分は何者でもないだろう...

@yoppy0123 吉岡 裕貴さんおはようございます☀️フォローありがとうございます🥹 フォロー返させていただきます♫今後ともよろしくお願いします！！

@icchooo_AI こちらこそよろしくお願いします！ Claude/Codexどちらも便利なので両方使っています！

Claude CodeでUsage Limitsにすぐ達してしまう方向けに、原因になりがちな使い方と対策をまとめました。 トークン消費の仕組み理解して効率的に使いたい方もぜひ読んでください！ zenn.dev/aun_phonogram/…

@d_create0804 D-CREATEさん初めまして！フォローありがとうございます🙏 AIを使って個人開発をしているいっちょーと申します😊 自分の制作物の進捗を知ってもらうのは大事だと思ってましたが、それをアプリにしようというのは盲点でした😳面白いアイデアですね✨️ 開発仲間として今後ともよろしくお願いします🙇‍♂️

新しい個人開発者向けWebアプリ 「Nokomise」を作ります！ AIで開発してると、 「今日なに直したっけ？」 「残りタスクなんだっけ？」 「LP作るの面倒…」 「スクショもログも散らばる…」 ってなりがちなので、 Markdownログとスクショを投げるだけで、 進捗・仕様変更・次タスク・残タスク・LP下書きを整理するアプリです。 開発者のAIトークンを無駄に使わせない設計にします。 βテスター募集します！ 個人開発してる方、試してくれる方いたらリプかDMください🙏 6月の最初の週には見せれる形になると思います #個人開発 #テスター

@kludgelog kludgelogさんリプありがとうございます！✨️ 自分の中だけで良くしようとこねくり回していても答えが出ないことも多いですしね😌 即決、即断、即行動！これからも意識したいところですね😊

@icchooo_AI 間違い無いですね！ 早く行動すればするほど、はやく次に繋がりますよね

今週の学び💡 「失敗するなら早い方がいい」 サイトを出してみて、自分の凡ミスにより大幅な改修を迫られましたが 最低限の実装のみを行ってすぐ公開したことで、結果的に傷は浅かったです😃 まずは出す。擦り傷、転倒上等で致命傷だけ防げればいい。実体験として良い血肉になりました😁 #個人開発

@Joelc_eth @SentientAGI Nice to meet you, Mr. Joel! Thank you for the like! ✨️ My name is Iccho, and I’m an independent developer working with AI. 🙏 I have a strong interest in AI and plan to continue using it in my work, so thank you for compiling this information. 🙇‍♂️

SPIN-Bench: In-Depth Analysis of Large Language Models' Strategic Planning and Social Reasoning Capabilities Sentient @SentientAGI The recent, swift progression in artificial intelligence research has ushered in remarkable advancements, especially in the competence of Large Language Models (LLMs) across text generation, context understanding, and various agent-based tasks. This leap forward has effectively paved the way for deploying LLM-based agents in sophisticated real-world scenarios, including advanced digital assistance, complex decision support systems, and collaborative human-AI partnerships. Nevertheless, numerous tasks within these crucial domains demand competencies that significantly exceed simple question-answering or short-term, superficial inferences. They require deep strategic planning and nuanced social intelligence. Strategic Planning fundamentally involves the capacity for long-term deliberation, which mandates the formulation of multi-step goals within vast action and state spaces. Social Intelligence, conversely, encompasses vital abilities such as discerning the hidden intentions and beliefs of other agents, effective negotiation, complex cooperation, and strategic competition. These two capabilities are profoundly interconnected and form the bedrock of truly intelligent behavior, particularly in dynamic, multi-agent environments. A critical limitation of current evaluation sets is their narrow focus, primarily concentrating on either isolated planning tasks or single-agent reasoning problems. This leaves a significant gap, as they lack a comprehensive, integrated framework to measure these two pivotal abilities simultaneously and holistically. To comprehensively address this deficiency, the Strategic Planning, Interaction, and Negotiation (SPIN-Bench) framework was meticulously developed. Its central aim is to systematically measure the LLMs' combined strategic planning and social reasoning capabilities across multiple distinct domains, with complexity levels that systematically and progressively increase. SPIN-Bench achieves this integration by unifying several critical task types under one robust umbrella: 1.Planning Domain Definition Language (PDDL) tasks. 2.Competitive board games. 3.Cooperative card games. 4.Complex multi-agent negotiation scenarios (like Diplomacy). This unified structure allows for an unprecedented, in-depth analysis of how LLMs integrate their aptitude for methodical, step-by-step planning with their capacity for conceptual social inference. Crucially, it enables researchers to observe precisely how this integration performs and where it breaks down as the inherent complexity of the task intensifies. This comprehensive work details the SPIN-Bench framework, presents the empirical results derived from its use, and outlines the profound insights gained regarding the current capabilities and critical limitations of LLMs. Definition and Philosophy of the SPIN-Bench Framework SPIN-Bench is conceptualized as a multi-domain, highly comprehensive evaluation and simulation environment specifically engineered to rigorously measure the sophisticated reasoning and strategic behavior of LLMs within social contexts. Its core philosophical tenet is the recognition that true intelligence—as demonstrated by strategic and social reasoning—is fundamentally more complex and sophisticated than mere isolated planning or reasoning tasks performed in static environments. Distinct from most existing benchmarks, SPIN-Bench avoids a focus on narrow planning or single-agent reasoning. Instead, it consolidates PDDL tasks, competitive games, cooperative games, and complex negotiation scenarios into a unified framework, functioning simultaneously as a benchmark and a realistic arena for stress-testing AI agents' reasoning and strategic acumen. The foundational design principle involves systematically manipulating the scale of action spaces, the complexity of the state representation, and the total number of interacting agents. This deliberate variation is used to construct diverse social environments where ultimate success hinges not solely on a methodical, step-by-step decision process but equally on accurate conceptual inference about the other, potentially adversarial or cooperative, participants. Through this, the framework seeks to meticulously identify the specific conditions under which LLMs excel, the precise capabilities where they exhibit deficiencies, and the fundamental reasons underpinning these observed shortcomings. SPIN-Bench is envisioned as a vital catalyst for propelling future research into robust multi-agent planning, sophisticated social reasoning, and effective human-AI team formation. The Four Main Components of SPIN-Bench SPIN-Bench's evaluation of LLMs' strategic planning and social reasoning capabilities is structured around four distinct, yet integrated, main components, each escalating the complexity and the demands on the LLM: 1. PDDL Tasks (Classical Planning):This component is the entry point, establishing a baseline by testing core reasoning skills in a fully controlled, deterministic environment. It includes: -- Factual Retrieval: Tests the basic recall of facts across domains like elevator and floor tile, often with expanding state spaces. -- Spatial Reasoning: A scenario to test the ability to track an agent’s position based on sequential, relative movement instructions. -- Full Classical Planning Tasks: A large set of 1,280 tasks across 21 domains demanding sequential decisions, resource management, and constraint handling. The goal is to isolate failures, identifying causes like poor state tracking or chain-of-thought breakdown. Competitive Games: Designed to assess adversarial reasoning, this component uses three popular turn-based games of increasing strategic depth: -- Tic-Tac-Toe: Focuses on short-term foresight. -- Connect Four: Requires deeper, multi-step tactical analysis. -- Chess: The ultimate test of deep tactical calculation and complex, long-term strategy against established AI solvers (e.g., Minimax, Stockfish) and human baselines. 2. Cooperative Game (Hanabi): This challenging card game evaluates multi-agent collaboration under conditions of incomplete information. Players cannot see their own cards but must communicate indirectly (via limited clues) to coordinate actions and achieve a shared, perfect score. It is a rigorous test of advanced Theory of Mind reasoning, requiring LLMs to infer the beliefs and intentions of teammates to execute coordinated actions. 3.Strategic Game (Diplomacy): As the most demanding component, Diplomacy is a multi-player game focused on negotiation, dynamic alliances, and strategic deception. Players must formulate long-term military plans while simultaneously engaging in complex, open-ended communication to secure or betray alliances. This environment captures a broad spectrum of realistic social behavior, testing the LLM's capacity for strategic adaptation, persuasion, and managing trust in a partially cooperative and partially competitive setting. Evaluation Metrics and Experimental Setup SPIN-Bench employs a dual system of evaluation: objective rule-based metrics and qualitative LLM-assisted social metrics, designed to provide a comprehensive performance analysis across a wide range of state-of-the-art LLMs. Rule-based Metrics --Classical Planning Tasks: Accuracy (is the plan correct?) and N-Step Look Ahead (how far ahead does the model plan?). --Competitive Games: Top-k Action Percentage vs Solver (how close is the move to the optimal move?) and Internal Elo Rating (relative skill comparison among LLMs). --Cooperative Game (Hanabi): Final Score (direct measure of cooperation success). Strategic Game (Diplomacy): Factual Consistency (do orders follow game rules?), Order-level Success Rate, and Final Game Outcome (supply centers controlled). LLM-assisted Negotiation Metrics For negotiation-heavy environments like Diplomacy, the quality of the communication is assessed using a powerful LLM (Claude 3.7 Sonnet) as an evaluator across six dimensions on the chat logs: 1.Strategy Alignment Ratio: Consistency of messages with the stated strategy. 2.Acceptance Rate: Rate at which the LLM's proposals are accepted by others. 3. Mutual vs. One-sided Benefit: Tendency to propose fair or unilaterally advantageous deals. 4.Peace vs. Conflict Inclination: The general tone of the communication. 5.Perspective-taking: Evidence of understanding other agents' hidden beliefs or intentions. 6.Conditional Negotiation: Frequency of "if... then..." tactics in proposals. Experimental Setup A broad range of commercial (closed-source) and open-source LLMs were rigorously evaluated, including GPT-4o, o1, Claude 3.5 Sonnet, Llama3.1-70b, DeepSeek-R1, among others. A unified interface provided models with the current state, history, legal actions, and game-specific Chain-of-Thought prompts. The Diplomacy engine was fully integrated to simulate complex multi-turn interactions with messaging. Analysis of LLMs' Fundamental Planning Competencies The analysis focused on isolating fundamental planning deficits: Factual Retrieval: While LLMs handled simple One-hop Queries (e.g., "Which locations does Russia control?") well, performance dropped notably on Multi-hop Queries (e.g., "Attack Analysis"), which require combining multiple facts and deeper inference. This indicates that complex information synthesis remains a significant challenge. Spatial Reasoning: Tasks requiring models to track their position through long sequences of relative movements showed that performance degrades as the sequence length increases. Advanced models like o1 proved more robust, maintaining high accuracy up to 29 steps, but many LLMs struggle to sustain complex spatial relationships over time. Error Analysis (Classical Planning): Errors were classified: 1.Breaking Constraints (BC): Execution errors. 2.Goal Not Satisfied (GS): Logically incomplete plans failing to reach the goal. Analysis revealed distinct weaknesses: Claude 3.5 Sonnet had a high GS rate (44.77%), suggesting good constraint adherence but poor goal-oriented plan completion. In contrast, RL-trained DeepSeek R1 minimized GS (3.65%) but had a higher BC rate, suggesting better goal-logic but worse execution. The data showed that the nature of failure is model-dependent. Effect of Action Complexity and Multi-Agent Scale on Planning Performance Complex or Expansive Action Spaces PDDL Domains: Model accuracy showed a stronger negative correlation with the total state-action space than with the average number of legal actions per step. This suggests LLMs are cognitively burdened by the vast potential for future state expansion (the "search burden"), hindering their pathfinding ability. Competitive Games: In Connect Four and Chess, all LLMs were defeated 100% against solver-level opponents, demonstrating that deep, multi-step tactical calculation remains far beyond their current capacity. Internal Elo ratings placed o1-preview and o1-mini highest, indicating relative superiority in strategic management compared to other LLMs, but still significantly below human expert or solver performance. Scaling Number of Agents -- Hanabi (Cooperative): The best model, o1, saw its score drop as the number of players increased from 2 (16.4) to 5 (14.2). Crucially, no LLM approached even the first quartile of human scores. This highlights a profound deficit in coordinating strategies and tracking information states across multiple incomplete channelsa failure of social intelligence under uncertainty. --Diplomacy (Strategic): Increasing the number of agents led to decreased order accuracy and successful attacks. The performance of strong planners like o1 was often negatively affected by the introduction of negotiation, suggesting that the cognitive load of managing social interactions and dynamic alliances can undermine the model's core planning coherence. Social Intelligence and Negotiation Performance Cooperative Scenarios vs. Human Baselines (Hanabi) The vast majority of human-played Hanabi scores cluster between 15 and 25 points. LLMs consistently failed to reach this range, confirming a pervasive social intelligence gap, particularly in perspective-taking and coordinated action based on inferred team knowledge. Negotiation in Diplomacy The inclusion of an explicit negotiation phase in Diplomacy often led to counterintuitive results. For the most capable planners like o1, end-game scores (Supply Centers) dropped significantly with negotiation, indicating that intense social interaction potentially disrupted their extended chain-of-thought planning. (Negotiation Message Analysis): A deep-dive analysis of negotiation messages using the LLM-assisted metrics provided insights into agent behavior:o1 showed the highest Acceptance Rate (%67) and a strong preference for Peaceful messaging (9.26 ratio). DeepSeek-R1 showed a higher inclination towards Conflict (0.95 peace ratio) and the most frequent use of Conditional Tactics (51%). Models generally achieved high Strategy Alignment ($>0.98$) but struggled with consistent, long-term strategic adaptation in the face of betrayals. Conclusion and Future Perspectives SPIN-Bench has served as a rigorous and comprehensive framework, revealing both the strengths and critical limitations of LLMs in strategic planning and social intelligence across a spectrum of tasks. The key takeaway is that while LLMs handle fundamental tasks well, they face significant performance bottlenecks in deep, multi-step reasoning over expansive state spaces and in coordination requiring social skills under uncertainty. No LLM demonstrated human-level competency in high-stakes, incomplete information, or highly social strategic games. The observation that social interaction can actively destabilize the planning coherence of even the strongest LLMs (like o1) is a crucial finding that necessitates further investigation. Future work should focus on developing models that are more capable of adaptive prompting, extending the benchmark to cover even broader aspects of social intelligence, and specifically addressing the cognitive friction between deep planning and managing dynamic social interactions. In conclusion, SPIN-Bench clearly underscores the current insufficiencies of LLMs in managing long-term, multi-agent scenarios. We anticipate that this benchmark will be instrumental in catalyzing the development of LLMs that are not only more strategically adept but also possess the necessary social intelligence for robust and effective human-AI collaboration.