Tomorrow auto mode becomes Claude Code's default. Plan mode changes too: exploration commands get classifier-reviewed, not prompted. Full semantics: https://dev.to/rulestack/claude-code-plan-mode-what-it-actually-blocks-what-still-runs-and-what-approving-switches-you-22m3 #AgenticAI #devtools

The `never` assignment is still the cheapest exhaustiveness check in TypeScript: widen the union and every switch that forgot the new member stops compiling. No lint rule, no library — the type checker does the sweep for you. What's your go-to guard for drifting unions? #TypeScript #devtools

Reviewing agent-written TypeScript, the fastest signal is the `any` count in the diff, not the logic. Each `any` marks a spot where the model gave up and kept going. Grep the diff first, read the code second — the shaky part surfaces in ten seconds. #TypeScript #CodeReview

An MCP server inherits every secret in the environment you launch it from. 'It only reads files' describes the tool surface, not the process. Give it a scoped token and a working directory you'd be comfortable posting publicly. #MCP #devtools

Custom subagents re-read your whole CLAUDE.md at startup — only Explore/Plan skip it. A 10-agent fan-out ships your rules ×10. What shrinks the multiplier: https://dev.to/rulestack/your-claudemd-loads-into-every-subagent-the-context-multiplier-nobody-budgets-for-440g #AgenticAI #devtools

Your agent's context window is a shared office: every always-on rule, every tool schema, every MCP server is a coworker talking at once. Capacity isn't the spec-sheet number — it's how much attention survives the noise. What's the loudest coworker in your setup? #LLM #devtools

A claim about someone else's tool needs a URL and a quote, not a memory. I write the source line next to the sentence now, because the version of me that edits it in three months has no idea which parts were checked. #BuildInPublic #devtools

Paste the compiler's error, not your summary of it. The agent uses exactly the parts you'd trim — file paths, the second error under the first, the note about where the type was declared. Paraphrase deletes its best clues. #TypeScript #PromptEngineering

Mine writes more disciplined test names than I ever did, and I've stopped pretending otherwise. Honest one: what does your coding agent genuinely do better than you? #AgenticAI #BuildInPublic

Once two agents work the same repo, the rules file stops being config and becomes an API contract between tools. Version it, review changes to it, and treat a silent edit like an unannounced schema change. #AgenticAI #devtools

Prose is the weakest form a rule can take. 'Run X after edits' can be a hook. 'Never touch Z' can be a deny rule. 'Match this style' can be linter config. Write enforcement first; keep sentences for what only judgment can do. #PromptEngineering #devtools

Fastest codebase onboarding in 2026: read the repo's AGENTS.md before its README. Build quirks, forbidden dirs, the real test command — teams finally wrote tribal knowledge down because the agent wouldn't work without it. Humans get it free. #AgenticAI #devtools

we measured what a Claude Code subagent actually costs: ~436k tokens of fixed overhead per agent (spawn + final cache write) before it does any work. embedding an entire product's text into the prompt? ~46k. headcount is the cost driver, not content. #AgenticAI #devtools

Cursor reads CLAUDE.md like AGENTS.md — unconditionally, ignoring alwaysApply. Claude-specific instructions run in Cursor too. The 4 rules files, with doc quotes: https://dev.to/rulestack/agentsmd-vs-claudemd-vs-cursorrules-what-cursor-actually-reads-now-and-in-what-order-17bo #CursorAI #devtools

Agents weigh examples over instructions. One outdated example in your docs outruns three correct rules, because the model copies shapes before it reads prose. Curating examples is rule-writing by other means. #PromptEngineering #LLM

Run /context and read the Memory files list — did everything you expected actually load? Nested files stay out of context until Claude touches their directory, so early in a session they're simply absent. What's missing from yours? #ClaudeCode #devtools

ran the numbers on subagent overhead. splitting one review into 3 perspective agents nearly tripled the token bill — and the merged single agent that actually fetched the docs caught the worst issue the trio missed. agent count is the cost lever, not payload size. #AgenticAI

a Claude Code subagent costs ~436k tokens in fixed overhead before any useful work — ~9.5x the content you hand it. measured across a real review pipeline: https://dev.to/rulestack/what-a-claude-code-subagent-actually-costs-measuring-the-436k-token-fixed-overhead-46g6 #AgenticAI

Flaky tests hurt agents more than humans. A human shrugs and reruns; an agent takes the failure at face value and 'fixes' working code. Determinism in your suite isn't hygiene anymore — it's the agent's ground truth. #AgenticAI #devtools

Daily practical notes on Claude Code, Cursor & Codex — rules that hold up in real projects. 18 packs live: 7 Claude Code skill packs · 4 Codex packs · 4 prompt collections · 3 Cursor rules packs. Start with the $1 AI Coding Starter. https://rulestack.gumroad.com?ref=bsky #ClaudeCode #CursorAI

Context compaction keeps your rules and drops your corrections. The fix you explained 40 turns ago is gone; the line in CLAUDE.md survives and reloads. Anything you've said twice in chat is asking to become a rule. #AgenticAI #LLM

A skill's body is lazy-loaded — write it as long as it needs to be. The description rides in every session's listing, budget-capped and always competing. Body length is free; description length is the scarce resource. #ClaudeCode #PromptEngineering

A metric that falls back to 0 when the lookup fails will one day report success during an outage. Zero is a measurement; null is an admission. Systems that can't say 'I don't know' lie precisely when it matters. #SaaS #BuildInPublic

For me, test assertions moved up the review list and formatting fell off it entirely. Since agents write more of your code: what do you review harder now, and what did you consciously let go? #AgenticAI #devtools

A Cursor rule scoped to **/* isn't scoped — it's a permanent header on every request. Globs are how a rule earns its context cost: attach it to the files where it changes behavior, and nowhere else. #CursorAI #PromptEngineering

Custom slash commands in Claude Code are skills now — same discovery, same listing budget, same lazy-loaded body. Two systems collapsed into one mental model: a command is a skill you address by name. #devtools #AgenticAI

Have the agent review its own diff before you look. It won't catch its blind spots, but it reliably catches the cheap stuff — dead imports, stray logs, the test it forgot to run. Spend your attention on intent, not lint. #AgenticAI #IndieHackers

A reviewer agent 'confirmed' a quote that doesn't exist in the cited source. Our fix wasn't a sterner prompt — it was structural: an external claim now needs URL, fetch date, and a verbatim quote, or it doesn't count as verified. #LLM #BuildInPublic

Read your agent's last ten commit messages in one sitting. If they tell a story of flailing — fix, actually fix, fix again — the problem usually isn't the model; it's that it can't run your tests. What do yours say? #AgenticAI #BuildInPublic

In a monorepo, other teams' CLAUDE.md files ride into your context too — contradictions and irrelevant conventions included. claudeMdExcludes exists for exactly this. Scope your context the way you scope your builds. #devtools #LLM