AI Developer Toolkit’s documentation changelog records new guides, lessons, and content refreshes for Cursor, Claude Code, and Codex, month by month from February 2026 onward. Tool release notes live on a separate page; this one covers what changed in the guide itself, including new skill deep dives and changelog sweeps.
This page tracks new and updated documentation on developertoolkit.ai — new guides, lessons and content refreshes. For release notes about the tools we cover (Cursor, Claude Code, Codex), see Latest Updates.
New: Compound engineering — the loop, the plugin, and what the guide skipsEvery's method for making each unit of work leave the system smarter, plus an audit nobody else ran. The published guide lists 26 agents and 23 commands; both directories now 404, and the repository holds 33 skills instead. Covers the four-step loop and the three time splits Every publishes for it, the five-stage adoption ladder, the ETH Zurich measurement that argues the whole premise backwards, and the retrieved-versus-always-loaded distinction that decides whether it applies to you. Four copy-paste prompts run the loop with nothing installed. EN+PL.
New: The AI Developer Toolkit MCP serverThis site now answers MCP calls at https://developertoolkit.ai/mcp — a public, read-only endpoint with two tools, `search` across every guide on the site and `fetch` for one article's full markdown. No account, no API key, one line of config in Cursor, Claude Code or Codex. The guide covers the setup for each, the response contract the ChatGPT and Claude connectors expect, and why a Polish query returns Polish articles with no language parameter. EN+PL.
New: Anti-slop writing skills — unslop, miodkuj, and what the evals sayThe category in one pass. What unslop's catalogue actually bans, why its three Python scanners run in CI with no agent in the loop, and the fact nobody else in the genre will tell you: its own core benchmark records a no-ship, while the category leader, at more than 36,500 stars, ships no evals at all. Plus miodkuj for the Polish tells no English word list reaches — officialese, nominalisations, genitive chains — and why it refuses to ship a linter. EN+PL.
New: Software factories — how to build one and keep it from rottingSix agents produce forty pull requests and two people who can meaningfully review them. Faros AI's 2026 telemetry puts numbers on that week: median time to first review up 156.6%, merges with no review at all up 31.3%. The article adjudicates the Dex Horthy / PostHog argument — both are right, about different things — and shows how to meter autonomy loop by loop against a verification oracle instead of declaring it factory-wide. EN+PL.
Dr. Skill — audit what your agent actually loadsA deep dive on drskill, the read-only diagnostic CLI that resolves each coding agent's effective skill and MCP loadout, then checks it against 34 issue categories: shadowed names, duplicate skills, colliding descriptions, prompt-injection surfaces, secrets in MCP configs, and unpinned server packages. Covers the tool-loadout research behind it, the audit command that reads your existing session traces, a CI gate, and the limits — including why a clean scan is not a clean loadout. EN+PL.
New: Pruning CLAUDE.md and AGENTS.md — the ablation protocolContext-file advice inverted in July 2026: stop adding a rule per mistake, start deleting. A dated map of why every guide you have read contradicts every other one, the loading mechanics each tool buries — which files load at startup, which lazily, and the truncation Codex performs with no trace in the TUI — a routing table for all five destinations a line can go to, and nine prompts that do the cutting. EN+PL.
New: Alternative IDEs and agent shellsA survey of the shells you can run Claude Code, Codex and Cursor Agent inside — Zed and JetBrains over ACP, herdr, Conductor, Nimbalyst, Emdash, Sculptor, Warp — sorted into four shapes, with one rule for what even qualifies (does it run your agent?), a decision table, and the repo changes that make parallel agents survivable. EN+PL.
New: herdr — the agent multiplexer with a scriptable APIReal PTY panes and a persistent server like tmux, but every pane classified idle, working or blocked and grouped in a sidebar. The deep dive covers how state detection actually works, sessions that outlive your laptop, the herdr agent CLI and the socket API underneath the TUI, and the orchestrator pattern where one agent spawns and supervises the rest of the fleet. EN+PL.
New: tmux for agent fleets — sessions, send-keys, wait-forClaude Code, Cursor and Codex create and clean up their own worktrees now, so every 'loop over git worktree add' guide is a layer too low. What tmux is still for: seeing which agent is blocked, surviving a closed laptop, and passing work between agents — through sessions, send-keys, capture-pane, wait-for, and hooks that report exact state instead of screen-scraping. EN+PL.
Claude Opus 5 across the model, pricing, and routing guidesOpus 5 shipped July 24 as the default Opus at Opus 4.8's price and half of Fable 5. Swept 187 files: new Opus 5 sections in the model-comparison appendix and Cursor's model-selection guide, the two API breaking changes (thinking on by default; disabling it capped at high effort), Claude Code v2.1.219 and category-based fallback routing, and a refreshed Artificial Analysis snapshot. Also corrected three claims that had gone stale independently: Fable 5's plan access, Gemini's above-200K price tier, and the Opus 4.8 effort default. EN+PL.
Impeccable guide rebuilt for v4 — direction by diceThe Impeccable deep-dive was expanded ~2.5× and re-verified against v4.0.2: direction dealing from 188 curated worlds, four visitor modes, a command-by-command reference for all 23 commands, Live Mode's pins-and-strokes workflow, the critique overlay, design hooks with two-speed reporting, doctor, native iOS/Android alpha, and the detector's growth from 44 to 59 enforced rules. EN+PL.
New: Mobbin MCP — real design examples for AI agentsAsk an agent for a paywall and you get the statistical average of every paywall in its training data. Mobbin MCP connects it instead to 620,000+ screens and 130,000 user flows captured from shipped iOS and web products: the three search tools, setup across Cursor, Claude Code and Codex, query patterns that actually return something, context-cost management, and the other MCP servers solving the same problem. EN+PL.
New: Taste Skill — the anti-slop frontend skill libraryDeep dive on Leonxlnx/taste-skill (~65K stars): how it constrains AI frontend generation up front — brief inference, three dials, a design-system map, and a hard anti-slop ban list — plus all 13 aesthetic variants, per-agent install, and how it composes with Impeccable and the design pipeline. Verified against upstream v2. EN+PL.
New: Superpowers — a disciplined design-to-PR agent workflowDeep dive on obra/superpowers: how a 14-skill plugin turns brainstorming, worktrees, executable plans, TDD, subagent review and fresh verification into one enforced workflow — with all ten installation targets, evidence, security and honest limits. Verified against upstream v6.1.1. EN+PL.
New: What Claude Code Can Do That You Haven't Tried~40 underused Claude Code features from Gui Ferreira's NDC AI 2026 talk — grouped into one-time setup, your daily loop, extensibility, and work beyond code. Every command verified against the current docs. EN+PL.
Tool changelogs refreshed through July 19Claude Code 2.1.208–215 (explicit /verify + /code-review, the EndConversation tool, /fork to background sessions, screen-reader mode, permission hardening), Codex CLI 0.144.6 (stronger dangerous-command detection; GPT-5.6 272K working context), and Cursor's July 17 Slack multi-repo update. EN+PL.
Full July 2026 AI tooling refreshSite-wide EN+PL update for GPT-5.6 Sol/Terra/Luna, Codex in ChatGPT desktop and CLI 0.144.1, Cursor 3.11 + Grok 4.5, and Claude Code 2.1.207 + Sonnet 5. Current recommendations, pricing, configuration and changelogs were re-verified against official sources.
Model comparison and routing rebuiltGPT-5.5 was removed from current defaults and replaced by role-aware GPT-5.6 Sol/Terra/Luna guidance; Sonnet 5 and Grok 4.5 were added with current pricing, availability and benchmark caveats.
Cursor ownership and benchmarks clarifiedThe docs now distinguish SpaceX's pending Cursor acquisition from the inaccurate claim that xAI already bought it, explain why Grok 4.5 does not replace Composer 2.5, and separate provider claims from current Artificial Analysis v4.1/v1.1 results.
New: /grill-me & /grill-with-docs — align before you buildTwo of Matt Pocock's most-installed skills: a relentless one-question-at-a-time interview that resolves every branch of your plan before you build — plus, with docs, a ubiquitous language (CONTEXT.md) and ADRs written as you go. Cross-linked from Plan mode. EN+PL.
New: 4 guides — /goal, /loop, shadcn/improve & ImpeccableCross-tool autonomous-run commands — /goal (native in Claude Code + Codex) and /loop (Claude Code + Cursor) — plus deep-dives on two standout skills: shadcn/improve (audit with your best model, execute with a cheap one) and Impeccable (a design vocabulary that fixes AI slop).
New: MCP token-cost & skill-security guidesTwo new guides — reducing MCP token cost with code execution, and skill supply-chain security (SkillSpector) — plus Chrome DevTools MCP coverage and research-backed prompting updates.
Tool changelogs refreshed through June 18Claude Code v2.1.183 (Artifacts, nested subagents, implicit agent teams), Cursor 3.8 (/automate, cloud subagents), and Codex v0.141 + Record & Replay.
Claude Fable 5 covered site-wideModel comparison, routing strategies, pricing and the Claude Code changelog (v2.1.169–170) updated for Anthropic's new top-tier model, released June 9.
Cursor 3.6–3.7 documentedFull coverage of Canvas Design Mode, Design Mode in the browser, the Cursor SDK (custom tools + auto-review), and the Auto-review Run Mode.
Claude Code v2.1.156–168 documented`ultracode` is now the dynamic-workflow trigger keyword, plus fallback models, `.claude/skills` plugin auto-load, and auto mode on Bedrock/Vertex/Foundry.
11 new official skills publishersOpenAI, Google Gemini, LangChain, Vercel, Prisma, Convex, Deepgram, Box, Contentful, Contentstack and Base added to the official skills guide.
Latest Updates refreshedThe cross-tool release-notes hub now tracks June across Cursor, Claude Code and Codex in one place.
Compound engineering guide — Every’s loop (plan, work, review, compound) as a cross-tool guide, plus an audit of the plugin behind it: the published guide at every.to lists 26 agents, 23 commands and 13 skills — its own directory tree says 14 in the same document — while agents/ and commands/ both 404 and the repository ships 33 skills, 27 reviewer personas owned by ce-code-review, and 14 agent hosts; the marketplace URL it gives you redirects after a rename. Also the five-stage adoption ladder, the three different time splits Every publishes for one loop, why /lfg does not merge, and the argument almost nobody raises — the compound step as the guide describes it writes into CLAUDE.md, the always-loaded configuration Gloaguen et al. measure as net negative, which the shipped plugin avoids by making docs/solutions/ retrieved instead. Four copy-paste prompts run the loop with nothing installed; verified against plugin v3.23.3, EN+PL (#337, Aug 24)
The AI Developer Toolkit MCP server — the site answers MCP calls at https://developertoolkit.ai/mcp: a public, unauthenticated, read-only endpoint carrying the two tools the ChatGPT and Claude connectors look for — search across every guide on the site and fetch for one article’s full markdown. The guide covers the one-line setup for Cursor, Claude Code and Codex, the response contract that makes a connector’s citations resolve, why a Polish query returns Polish articles with no language parameter, and what the endpoint does not do (no writes, no account, no rate limit worth documenting yet), EN+PL (#319, Aug 20)
Anti-slop writing skills — a deep dive on theclaymethod/unslop and bartekpucek/miodkuj: the three deterministic Python scanners that run in CI with no agent in the loop, the progressive-disclosure mechanics that let a skill ship a 50 KB rule catalogue for about 100 tokens of standing cost, the eval-first maintenance loop, and the finding that separates unslop from the rest of the category — its own core benchmark records a no-ship, while the most-starred skill in the genre ships no eval suite at all. miodkuj covers the Polish tells no English word list reaches (officialese, nominalisations, impersonal fog, genitive chains) and makes the opposite architectural bet, declining a deterministic linter on the grounds that a match list cannot resolve context, EN+PL (#323, Aug 19)
/goal and /loop tool support re-verified and corrected — the two autonomous-run guides asserted that /loop was native only to Claude Code and that Codex’s /goal was experimental and had to be enabled. Both were out of date: Cursor 3.5 shipped a /loop skill on May 20, 2026 (a month before the article was written), and Codex CLI 0.133.0 made goals stable and on by default on May 21, 2026. Re-checked against Claude Code 2.1.231 and Codex CLI 0.145 rather than the cached doc snapshots, which stop at 2.1.37 and February 2026. Also folded in the /loop behaviour that had accumulated since — the v2.1.196 rule that a scheduled fire only runs self-invocable skills, ScheduleWakeup stop: true in v2.1.202, the cloud-provider fixed-interval fallback, jitter and the 50-task cap — plus Codex’s /goal edit, and new parity rows in the feature matrix and command reference, EN+PL (Aug 13)
Software factories guide — what a software factory is, why Dex Horthy (“Why Software Factories Fail”) and PostHog (“Can software factories actually work?”) are both right and are not arguing about the same thing, the ceiling SlopCodeBench measures (structural erosion rising in 77% of trajectories, agent code 2.3x more verbose and 2.0x more eroded than human code), the self-inflicted wound below it — pipelines that split “decide” from “build” and hand the agent a ticket stripped of production context — and how to meter autonomy loop by loop against a verification oracle, EN+PL (#257, Aug 12)
Dr. Skill deep dive — a full guide to dbreunig/drskill, the loadout diagnostic CLI: the tool-loadout research behind it (routing accuracy collapsing past 30 tools, a 600-skill enterprise agent), the 34 checks grouped into skill hygiene, description quality, injection surfaces and MCP configuration, drskill audit reading the session traces your agents already write, a CI gate with a fingerprinted ack ledger, the invocation-time shell-command exposure, and the limits that make a clean scan less than a clean loadout; verified against version 0.6.7, cross-linked from the skills-ecosystem index, skill security, best practices and the skills CLI reference, EN+PL (Aug 12)
Answer-first rewrite of the corpus — every non-index article now opens with a self-contained paragraph that answers its own title before any preamble; template headings that said nothing were removed, faq and contentType frontmatter was added where it applies, and the rule was promoted into the build gate so a regression fails the build rather than shipping quietly (#214–#225, Aug 6)
The ablation protocol for context files — pruning CLAUDE.md, .cursor/rules and AGENTS.md as a budget to cut rather than a log to grow: a dated map of why the guides contradict each other, what each tool actually loads and when (including Codex’s silent 32 KiB truncation), a routing table for all five destinations a rule can live in, nine prompts that do the cutting, and observed stumbles; three accuracy corrections followed post-merge, EN+PL (#212, #213, Aug 4)
herdr deep dive — the agent multiplexer that classifies every pane idle, working or blocked: how state detection actually works, sessions that outlive a closed laptop, the herdr agent CLI and the socket API beneath the TUI, the orchestrator pattern where one agent spawns and supervises the rest of the fleet, and where herdr fleets break down; the Moshi iPhone/iPad terminal was added to the mobile-access section shortly after, with its “no relay” claim scoped to hosting rather than data locality, EN+PL (#196, #205, #206, Jul 30)
tmux for agent fleets — what tmux is still for now that Claude Code, Cursor and Codex create and clean up their own worktrees, which puts every “loop over git worktree add” guide a layer too low: session per task, send-keys as the agent-to-agent channel, capture-pane polling, wait-for barriers, and hooks that report exact state instead of screen-scraping, EN+PL (#196, Jul 30)
Alternative IDEs and agent shells survey — the tools you can run Claude Code, Codex and Cursor Agent inside, organised into four shapes (editor + swappable agent over ACP, terminal-native, desktop orchestrators, boards), with a decision table, a terminal baseline to try before buying anything, and the repo changes that make parallel agents survivable, EN+PL (#195, Jul 30)
Mobbin MCP guide — connecting an agent to 620,000+ screens and 130,000 user flows from shipped iOS and web products: the three search tools, setup for Cursor, Claude Code and Codex, query patterns that return usable results, context-cost management, the MCP-versus-REST-API split, and the other design-reference servers including Refero, EN+PL (#183, Jul 25)
Impeccable v4 refresh — the guide now covers v4’s direction dealing (188 hand-approved worlds, five direction tests, system board + first-surface mock before code), the six-intent command regrouping with a full command-by-command reference (craft deprecated in favor of conversational job classification), Live Mode’s pins/strokes/three-variant workflow, the critique overlay, design hooks with two-speed reporting, doctor staleness checks, the native iOS/Android alpha, and the detector catalog at 64 patterns / 59 enforced rules; verified against v4.0.2, EN+PL (Jul 25)
Taste Skill deep dive — a full guide to Leonxlnx/taste-skill (~65K stars), the anti-slop frontend skill library: how it constrains AI generation up front (brief inference, three dials, design-system map, hard ban list, pre-flight check), all 13 aesthetic variants with their install names, per-agent install, honest scope limits, and how it composes with Impeccable and the design pipeline; verified against upstream v2, cross-linked from the frontend-skills, Impeccable and design-pipeline pages, EN+PL (#170, Jul 21)
Superpowers workflow guide — a deep dive on obra/superpowers: the 14-skill design-to-PR workflow (brainstorm → worktree → executable plan → TDD → subagent review → fresh verification), with installation targets, evidence, security and known limits; verified against upstream v6.1.1, cross-linked from the skills ecosystem, core methodology and the scorecard guide, EN+PL (#159, Jul 11)
“What Claude Code Can Do That You Haven’t Tried” — ~40 underused Claude Code features from Gui Ferreira’s NDC AI 2026 talk, grouped into one-time setup, daily loop, extensibility and work beyond code; every command verified against the current docs, EN+PL (#146, Jul 3)
Tool changelogs + model refresh — Claude Code 2.1.208–215 (explicit /verify + /code-review, EndConversation, /fork to background sessions), Codex CLI 0.144.1–0.144.6 (GPT-5.6 272K working context), Cursor 3.11 + Grok 4.5; model comparison and routing rebuilt around GPT-5.6 Sol/Terra/Luna and Sonnet 5 (Jul 20)
/grill-me & /grill-with-docs alignment skills — a relentless one-question-at-a-time agent interview that resolves your plan before you build, plus a ubiquitous language (CONTEXT.md) and ADRs with the docs variant; cross-linked from Plan mode and the skills ecosystem, EN+PL (Jun 24)
/goal & /loop autonomous-run guides — cross-tool coverage of goal-directed (/goal: Claude Code + Codex) and looping (/loop: Claude Code + Cursor) agent runs, with the gaps in each tool mapped honestly, EN+PL (#129, Jun 24; tool-support corrected Aug 13)
shadcn/improve & Impeccable skill deep-dives — an architecture audit that writes plans for cheaper executors, and a design vocabulary that catches AI slop; both verified across Claude Code, Cursor and Codex, EN+PL (#129, Jun 24)
Claude Fable 5 site-wide refresh — model guides, pricing, routing strategy and Claude Code changelog v2.1.169–170 (Jun 10)
Tool changelog refresh — Claude Code v2.1.156–168 (ultracode trigger keyword, fallback models, .claude/skills plugin auto-load, auto mode on Bedrock/Vertex/Foundry) and Cursor 3.6–3.7 (Canvas Design Mode, Cursor SDK, Auto-review Run Mode, Enterprise Organizations); Codex confirmed current at v0.137 (Jun 8)
11 new official skills publishers documented — OpenAI, Google Gemini, LangChain, Vercel, Prisma, Convex, Deepgram, Box, Contentful, Contentstack, Base (Jun 8)
Latest Updates + update tracker refreshed with June model/version coverage (Jun 8)
This page tracks new and updated documentation on developertoolkit.ai -- new guides, lessons, and content refreshes. Latest Updates carries release notes for the tools the guide covers: Cursor, Claude Code, and Codex.
What documentation shipped in August 2026?
Six new guides -- the site's own MCP server, software factories, the Dr. Skill loadout audit, the ablation protocol for pruning CLAUDE.md and AGENTS.md, the anti-slop writing skills, and compound engineering -- plus an answer-first rewrite of every article in the corpus, now enforced by the build. All of it shipped EN+PL.
What documentation shipped in July 2026?
The Claude Opus 5 sweep across the model, pricing, and routing guides (187 files); the Impeccable guide rebuilt for v4; new deep dives on Taste Skill, Superpowers, Mobbin MCP, herdr, and tmux for agent fleets; a survey of alternative IDEs and agent shells; "What Claude Code Can Do That You Haven't Tried"; and tool changelogs refreshed through July 19.
How often is the documentation updated?
Continuously. New or refreshed documentation lands here most weeks, and the changelog is grouped by month, currently running from February 2026 through August 2026.
Where do I request new documentation?
Through the feedback page -- subscriber requests shape the next sprint. The Product Roadmap shows what is already planned.