The AI development publication
In-depth guides, honest reviews, and battle-tested workflows for building software with Cursor, Claude Code, Copilot and the modern AI stack.
AI Coding AssistantsOne 53-line diff, six known bugs, five runs across two models and three effort levels — plus the routing table that decides how deep the review actually goes.
AI Coding AssistantsThe prompt behind /security-review, its 17-item exclusion list and confidence-8 cutoff, measured against a diff with six known vulnerabilities in it.
AI Coding AssistantsWe pulled the default keymap out of the shipped binary: 180 bindings, 82 keys, 22 contexts, 30 actions you can only reach by binding a key yourself.
AI Coding AssistantsA measured precedence ladder for Claude Code settings, plus the discovery rule that quietly disables a monorepo config and the malformed file nobody warns you about.
AI Coding AssistantsNot a copy of the docs table: a census of what one build actually answers to, split by the three different replies the CLI gives.
AI Coding AssistantsRules backed by a failing gate: 0 violations in 93 articles. Rules that exist only as sentences: 92. We re-measured a year of the same corpus to find out what changed.
AI Coding AssistantsWe measured --safe-mode, --restricted and --bare on one machine. Safe mode removes your config and nothing else — the flag that removes capability is a different one.
ComparisonsWe priced the gap between an agent and a chat client on one machine: 17,709 tokens of tools and project context, 63.4% of a default Claude Code prompt.
AI Coding AssistantsA technical breakdown of Claude Code Auto Mode, autonomous tool execution loops, permission hierarchies (default, acceptEdits, auto, bypassPermissions), and safety controls.
AI Coding AssistantsA comprehensive guide to Claude Code permissions, tool rule schemas (allow, deny, ask), settings hierarchy precedence, glob matchers, and /permissions commands.
AI Coding AssistantsHow Claude Code manages session state, command line flags (-c, --continue, --resume), session forking, multi-terminal concurrency, and prompt caching economics on resume.
AI Coding AssistantsWhat a checkpoint is on disk, how much of a session it covers, and the measured gap between what rewind restores and what your working tree actually changed.
AI Coding AssistantsWhat /clear removes, what it leaves on disk, and the measured floor a cleared session starts from — across 165 transcripts and 16 releases.
AI Coding AssistantsThe full ~/.claude layout measured on a working machine: what is in the .jsonl transcripts, how project directory names are built, and what a purge leaves behind.
AI Coding AssistantsShift+Tab is the plan mode hotkey, how many presses it takes from each mode, why optional modes never move plan, and the fallbacks when your terminal intercepts the key.
AI Coding AssistantsA section-by-section layout for CLAUDE.md, derived from three real config files rather than a template, plus the length ceilings that keep each rule findable.
AI Coding AssistantsEvery documented way to leave plan mode, whether it keeps or discards the plan, and exactly which permission mode the session lands in after each one.
AI Coding AssistantsHow Claude Code compaction works under the hood, before-and-after token measurements, prompt caching dynamics, and when to use /compact vs /clear.
AI Coding AssistantsLocal Node.js process RSS vs remote LLM context memory, V8 heap benchmarks, child process spikes, and how to optimize Claude Code on RAM-constrained machines.
AI Coding AssistantsDoes CLAUDE.md have a hard character or token limit? We debunk the 150k limit myth, test instruction compliance degradation, and show how to modularize large rule sets.
AI Coding AssistantsThe median turn fills 96% of a 200k window and 46.9% exceed one outright. Context came back down exactly zero times in 89 sessions.
AI Coding AssistantsAcross 20,285 billed turns the operator typed 47,296 tokens and the input side came to 4.25 billion. Shorter prompts move a rounding error.
AI Coding AssistantsA median turn takes 3m 08s and 44 steps to get there. The lever is the number of steps, not the size of the context.
AI Coding AssistantsEvery assistant turn stamps its own version and its own token count, so the question is answerable retroactively. Across 17 versions the cost per turn fell 47% from its peak.
GuidesA local model closes one of seven outbound services. The other six are yours, and the one that stops you is memory rather than model quality.
GuidesZero injection, zero committed secrets, every admin surface guarded — and four unescaped sinks, no CSP, and 548 packages nobody read.
GuidesThe contract governs the vendor. Files on each developer's machine govern the agent — and most of them never reach a diff.
GuidesThroughput is measurable to the byte. We counted what an agent produced in 33 days, who it was produced for, and how much of it ever reached a reader.
ComparisonsChoosing between Anthropic's CLAUDE.md and the open AGENTS.md standard. Scoping, memory injection, and the single-source bridging strategy.
GuidesThe multi-agent instruction format adopted across 60,000+ repositories. How it is structured, discovered, and executed.
GuidesWhy agent sessions exhaust quota in under 20 turns. How rolling windows, reasoning multipliers, and subagents drain usage caps.
GuidesThe published workflows are lists of rules. A workflow is the subset that executes whether or not anyone remembered it.
GuidesThe generator is not the bottleneck. On a real Next.js codebase, 87% of modules need a fixture before the first assertion runs.
ReviewsGenerated code does not rot. It gets rewritten in a very small number of places, and this measures which ones.
Guides76 modules, 169 import edges. A third of them can be reviewed from the diff alone. Nine of them cannot be reviewed from a diff at all.
GuidesTen seeded faults, every one a failure this project actually hit. Three produce no message for any debugger to read.
GuidesWe counted every turn in the sessions that built this site. The median run between two human sentences is 30 tool calls; the longest is 236.
ComparisonsTen tools, thirteen attributes, and a provenance class on every cell. Forty per cent of the table is a gap, and no tool has a complete row.
Comparisons99% of this site's source is read by an automated check. A misspelled Tailwind class passes all of them and ships to the browser doing nothing.
ReviewsOne of ten obvious install commands gets you the tool. Three get you unrelated software that ships a command with exactly the name you meant to run.
ComparisonsTwo stores, two completely different rankings, and both of them put unmaintained software near the top. Here is the measured shelf and the three signals worth ranking on.
PricingThe same eight files are 162,409 tokens on one tokeniser and 134,707 on another. A $/MTok column compares prices quoted in different units — here is the unit that converts.
ReviewsEight of eleven are genuinely open source, two are source-available, one is freeware. None of that changes the token bill, which is where free stops being free.
ComparisonsThree vendors, one 354,835-token session, and three different pricing shapes. Gemini overtakes Claude Opus after 80 minutes, and the winning row rests on a rate nobody publishes.
GuidesSWE-bench Verified counted from the dataset rather than the paper. Every gold patch but one edits Python, which makes the leaderboard a Python leaderboard and almost nothing else.
GuidesOne question, one model, two configurations. Removing the harness cost the right answer and 72% of the loaded context, and the loop sent the transcript three times over.
ComparisonsSeven Q4_K_M candidates, real byte counts from Hugging Face, one ordinary Windows laptop. Zero fit the GPU, four fit RAM, and three separate runtimes refused to start.
ComparisonsOne vendor charges a fee for the cache, the other charges rent. On a measured session the metered gap looks like 110x and reverses after 53 minutes.
GuidesNine files from one repository, measured by differencing two headless turns. The folklore constant undercounts every one of them, and the session was 31,979 tokens deep before anyone typed.
ComparisonsNine published cards, one measured request, a 10.5x spread — and two vendors who publish no context window on the page everyone quotes prices from.
ComparisonsTwo turns of the same prompt, minutes apart. The model term is exactly 2.5x for every session you will ever run. The other two terms are the ones costing you money.
GuidesThe filter that builds SWE-bench keeps roughly 2.5% of the pull requests it sees. Point it at your own repository and you learn more than the leaderboard will tell you.
AI Coding AssistantsOne five-word prompt billed two different models. The resolution chain that picks them ends somewhere your repository cannot see, and the script that proves it.
ComparisonsSix configuration concepts audited across one real repository. Copilot CLI is the first target that scores 6 of 6 — and the first that reads a competitor's directory by name.
ComparisonsTwo tokens of a 32,005-token request were typed by a person. Before you can compare two coding models, you have to control the thing that wrote the other 32,003.
ComparisonsTen agents surveyed on the two signals that survive a rewrite: whether the project is still moving, and whether your configuration comes with you.
ComparisonsCline reads AGENTS.md but not CLAUDE.md, and its MCP file is a byte-identical copy of ours that belongs in your home directory. We measured both against the same repository.
ComparisonsRoo Code was archived on May 15, 2026 and roocode.com now redirects to an unrelated product with no migration notice. What to carry out of a Roo repository, and where to.
ComparisonsAider has no subagents, no skills, no hooks, and discovers no instruction file. We measured what that costs and what it buys against Claude Code on the same repository.
ComparisonsGoogle replaced Gemini CLI on June 18. We measured what a Claude Code repository actually carries into Antigravity CLI — and which single line has to change first.
ComparisonsFeature tables have converged, so we measured lock-in instead — which configuration surfaces survive a move between Claude Code, Codex CLI and Antigravity CLI.
ComparisonsBoth are terminal agents and the feature lists have converged. The decision that costs you a day is configuration portability — so we measured it on a real repo.
DeploymentD1 without Workers: one wrapper module, the measured transport cost, the four error codes worth handling, and why batching and bound parameters cannot both be used.
DeploymentWhich of your variables are frozen into the bundle, which are read per request, and why a value that works locally can be absent in production.
Next.jsThe attribute is not undefined — it is deleted from the tree before compilation. What changed in v6, why the repo is archived, and the three ways out.
AI Coding AssistantsClaude Code's own toolset costs 122,137 bytes as an MCP server. That number is the budget every server you install is competing against, and nothing surfaces it.
AI Coding AssistantsFive servers start concurrently in under three seconds. The bill that actually matters is 30,218 bytes of tool definitions sitting in context on every single message.
Next.jsYour Next.js project already contains the code the MCP server wants. Importing it fails with an error about Client Components — and the fix is one CLI flag.
AI Coding AssistantsTool counts, cold and warm startup, and the context each server spends before doing anything — plus the two servers every list recommends that npm returns 404 for.
AI Coding AssistantsThe three scopes, every field that does something, and the unset-variable trap that passes the literal string ${VAR} to your server instead of failing.
AI Coding AssistantsWhy -32000 Connection closed means four different things, the corrupt-stdout failure the health check calls ✔ Connected, and a committed script that separates them.
AI Coding AssistantsThree surfaces, one reference: shell subcommands, CLI flags, and in-session slash commands, reconciled against what the binary actually accepts.
AI Coding AssistantsTwo real subagents committed to this repo, a zero-cost discovery probe, and the debug-log evidence for three failure modes the session never shows you.
AI Coding AssistantsWe audited 27 articles against 14 written rules. Rules backed by the build had zero violations. Rules that existed only as sentences had 89.
AI Coding AssistantsThe message names your subscription, but Anthropic lists auto mode as available on all plans. What actually gates it, and why plan mode can trigger it spuriously.
AI Coding AssistantsThe three places plan mode spends tokens, the classifier round-trip nobody counts, and the decision rule for when planning is cheaper than the rework it prevents.
AI Coding AssistantsThe complete set of plan mode controls on every Claude Code surface, which ones persist, and why /visual-plan does not belong on the list.
AI Coding AssistantsThere are six permission modes, not two, and the two people compare most are the two that compose. What each mode actually approves, measured against a real allowlist.
AI Coding AssistantsCLAUDE.md is the only thing Claude Code loads in full regardless of length. Here's how to measure the context floor your project pays on every session.
AI Coding AssistantsA plan is a tool-call argument inside a session transcript, not a document on disk. Verified across 126 transcripts, plus the settings that change what happens on accept.
PricingCopilot and Cursor both dollar-denominate included usage now, which makes them directly comparable for the first time. The answer is not the one the seat prices suggest.
ComparisonsThe three converged at $20/month for individuals and diverged completely at team scale. Windsurf is now Devin Desktop, and the team pricing is where the real difference is.
GuidesZero data retention, content exclusion, SOC 2. Every one of these controls has a documented gap, and the gap is in the vendor's own docs. Here's what to read before you buy seats.
Next.jsNext.js gives you the sitemap, the metadata and the schema. What it does not give you is a warning when your anchors, your lastModified or your internal links quietly break.
AI Coding AssistantsCustom commands merged into skills. What that actually changed, how to build one from a procedure you keep pasting, and the token cost nobody documents.
ComparisonsEvery major tool now has an agent, a plan mode, and a config file. What still differs is where it runs and how it meters you. Pick on those two axes.
ComparisonsCopilot has a terminal agent with plan mode now, so the old framing is dead. The real split is where the agent runs, how it bills, and which models it can reach.
Next.jsA complete account of building this site with Claude Code — the stack, the architecture the platform forced, and every failure that turned into a build-time check.
ReviewsEmployed developers get a seat. You buy one. Here's the break-even between a $20 subscription and bring-your-own-key, using every vendor's live August 2026 pricing.
ReviewsLicence and price are the same across all three. Release cadence, maintenance signals, and ownership are not — and that is the axis that should decide it.
AI Coding AssistantsInstallation, CLAUDE.md, plan mode, hooks, MCP, and subagents — the mechanics of Claude Code, drawn from running it in production on a real repository.
DeploymentA real GitHub Actions setup using this site's own typecheck, lint, and build scripts, plus where an automated AI review step should not run unattended.
PricingThe seat price is a floor, not a ceiling. Here's how Copilot's credits, Cursor's usage billing, and Claude Code's Max tiers actually work once you go past the included amount.
ComparisonsThe workflow difference that actually matters when picking between them — not a feature checklist, since both ship the same feature list within a quarter of each other.
Next.jsThe workflow, the CLAUDE.md rules that prevent silent failures, and a real build error this exact setup hit and fixed.
ReviewsWe checked every major vendor's live pricing and docs pages on the same day. Here's what each free plan actually caps, and which numbers are stale.
AI Coding AssistantsHow to enter plan mode, what it actually blocks, the four approve options, and the decision rule for when planning beats editing straight away.
AI Coding AssistantsTwo 2026 studies found file length and structure don't move adherence. Here are the four CLAUDE.md mistakes that do cost you.
DeploymentThe four-step deploy, the DNS records that actually matter, and the three Vercel runtime limits that break real apps after the first green build.
AI Coding AssistantsWhere the file goes, the structure we run in production, and the 200-line ceiling where a CLAUDE.md starts working against you.
// THE DEVVENTA DIGEST
New tools, deep dives, and the workflows that actually moved the needle. No fluff. Unsubscribe anytime.
