Tue 09:12
You approve a 600-line diff you only skimmed, because reading it properly would take longer than writing it.
For engineers and teams working with Claude Code, Codex or Cursor
AI Developer Toolkit is the open manual for agentic engineering with Claude Code, Codex and Cursor. Read all open guides free, audit your repository locally with @wondelai/harness-audit, or pick up the eight offline books and team services.
npx @wondelai/harness-audit 100% open web docs · MIT CLI runs locally in <2s · Zero code uploaded.
One task, start to finish, on a Level 4 setup.
Your attention goes to the spec before the build and the evidence after it. The diff goes to checks you set up once.
TicketIntent
PAY-412
Prorate plan changes made mid-cycle.
You write
SpecIntent
spec.md
Acceptance criteria, edge cases, files the agent must not touch.
38 lines
You write and read
PlanIntent
plan.md
Seven steps. The agent proposes them, you approve before it builds.
Agent writes
BuildLoop
+612 −148
14 files changed in an isolated worktree, until the tests pass.
Agent writes
VerifyVerification
760 changed lines read
Checks read
Pull requestRelease
#1184
Opened with the test report attached. You read it against the spec.
24 lines
You read
Engineering teams at these companies use it
Creator of wondel.ai Skills — 2.7k+ stars on GitHub (opens in a new tab)
Once the agent writes most of the code, you become its full-time reviewer. Every task ends in a diff you either read line by line or approve on trust. Approving on trust is vibe coding. Agentic engineering replaces that trust with checks that can say no.
“You are still responsible for your software, just as before.”
Andrej Karpathy, published transcript of his Sequoia Ascent talk (opens in a new tab), 2026-04-30
“Your life is diffs.”
Tue 09:12
You approve a 600-line diff you only skimmed, because reading it properly would take longer than writing it.
Tue 11:30
Your CLAUDE.md started as a paste from someone’s thread. Nobody on the team knows which lines still do anything.
Tue 14:05
Three engineers run three different coding agents on the same repo, with no shared hooks, test gates or review bar.
From our reader survey
“I switch between Claude Code, Cursor, VS Code, GitHub, and other agents, and often have to explain the same project or problem again.”
A reader, July 2026
What two years of telemetry from 22,000 developers shows
Faros AI, “AI Engineering Report 2026”, April 2026 (opens in a new tab)
+66%
epics completed per developer
+242%
incidents per pull request
+441%
median time a pull request spends in review
Output went up, and so did incidents and review time. The way out is checks that read the code for you, set up once and run on every task. That is what this canon teaches.
Dan Shapiro’s five levels describe how much of the work you hand to an agent, from spicy autocomplete to the dark factory. Every open guide is filed by level, so the next step is always named.
Want a real measurement instead of a guess?
Run npx @wondelai/harness-audit locally in under 2 seconds. It inspects your repository across 6 Factory Stations, calculates your Level 0–5 cap, and links directly to the free chapter for each gap.
The dark software factory
“a black box that turns specs into software”
WritesAI
ReadsNobody
The engineering team
“write a spec” … “leave for 12 hours, and check to see if the tests pass”
WritesAI
ReadsTests, mostly
The developer
“Your life is diffs.” … “almost everyone tops out here”
WritesAI, most of it
ReadsYou, full-time
The junior developer
“where 90% of ‘AI-native’ developers are living right now”
WritesYou and the AI, paired
ReadsYou, every line
The coding intern
“you offload specific, discrete tasks to your AI intern”
WritesYou, mostly
ReadsYou, all of it
Spicy autocomplete
“not a character hits the disk without your approval”
WritesYou
ReadsYou
Pick the situation you are in today. Every docs track and the CLI are open and free.
WhenYour agent config grew organically and you don’t know what’s missing.
You get Run npx @wondelai/harness-audit locally (<2s, MIT, zero code uploaded). Get a 0–100 score across 6 Factory Stations, Level 0–5 gating, and --prompt to emit HARNESS-FIX-PROMPT.md.
WhenThe agent’s diffs are bigger than your attention span.
You get The Level 3 to 4 patterns: test oracle protection, Stop hooks, sandboxes and architectural fitness functions so every task ends with evidence instead of a wall of code.
WhenThe task is too big to hand over in one prompt.
You get The brief chain: intent.md, spec.md, plan.md. The agent works from a spec you wrote and approved (spec-driven development), in Claude Code, Codex or Cursor.
WhenYou want a structured solo path from Level 2 to Level 4.
You get The open Developer Track: write one tight AGENTS.md, isolate loops with worktrees and permissions, and move code review from line-by-line reading to evidence verification.
WhenYour team works at five different levels on the same codebase.
You get The open Tech Lead Track plus harness-audit --interview --role lead: baseline the repo, standardize hooks and CI gates, and run code review as a queue with WIP limits.
WhenYou want an external diagnostic or hands-on rollout with your team.
You get Fixed-price work directly on your repository, delivered by Michał Jaskólski: a written Repo Diagnostic from $499, a Team Working Day, or a measured org pilot.
Everything you need to learn and audit your repo is free. Pay only for the 8-book offline bundle or hands-on team services.
The eight books, at a comparable ebook’s price$383
The Offline Ebook Bundle, paid once$59.99
140+ architecture and mindset guides for Claude Code, Codex and Cursor, filed by level, in English and Polish. Covers context engineering, verification and org rollout — linking to official docs for volatile CLI flags.
Free100% open web
Open-source MIT diagnostic for harness engineering. Deterministic <2s local scan across 6 Factory Stations, optional --deep LLM review, --interview, --prompt and safe init scaffold.
MITopen source
The four-title library (96 chapters, 541 pages) plus four role paths for developers, tech leads, CTOs and executives. DRM-free EPUB & PDF in EN & PL, or single editions on Amazon. See the shelf
$59.99all 8 books
Also free without an account: the open MCP server so your agent can search the canon, the Agentic Engineering Glossary, the /harness-audit skill in wondelai/skills, and the field-notes newsletter.
The four library titles run to 541 pages, and the four role paths reuse those pages in the order your role needs. Get all eight DRM-free offline for $59.99, or individual editions on Amazon.
Also on amazon.com: Kindle on amazon.com, opens in a new tab
Also on amazon.com: Kindle on amazon.com, opens in a new tab Paperback on amazon.com, opens in a new tab
Also on amazon.com: Kindle on amazon.com, opens in a new tab Paperback on amazon.com, opens in a new tab
Also on amazon.com: Kindle on amazon.com, opens in a new tab
Also on amazon.com: Kindle on amazon.com, opens in a new tab
Also on amazon.com: Kindle on amazon.com, opens in a new tab
Also on amazon.com: Kindle on amazon.com, opens in a new tab
Also on amazon.com: Kindle on amazon.com, opens in a new tab
1 / 8
Day 1
Run npx @wondelai/harness-audit --prompt inside your main repo (<2s, 100% local). See your score across 6 Factory Stations and generate HARNESS-FIX-PROMPT.md for the top missing guardrails.
Days 2 to 5
Follow the free chapters linked from your audit to wire a Stop hook and test gate, then write intent.md and spec.md for one real task. Review the test report and evidence bundle.
Week 2+
Put harness-audit --ci in GitHub Actions so hygiene never regresses, share the open role tracks, or book a fixed-price Repo Diagnostic or Team Working Day to accelerate the rollout.
MIT open source · Zero account required · Works on Claude Code, Codex and Cursor repos.
> npx @wondelai/harness-audit --prompt Score: 64/100 · Level 3 (Review Manager) Cap: Verification (missing oracle guard & Stop hook) Wrote HARNESS-FIX-PROMPT.md (4 prioritized fixes)
Every chapter focuses on durable engineering principles and links directly to official Claude Code, Codex and Cursor documentation for fast-changing CLI flags and shortcuts.
Each guide shows when it was last audited and updated, so you always know how fresh what you are reading is.
@wondelai/harness-audit is MIT-licensed with zero runtime dependencies. Every check is readable TypeScript you can run offline or gate in CI.
Every statistic on this page names its publisher and when it was published. No invented productivity multipliers.
Every guide links to its GitHub issue tracker, and wondelai/skills is maintained openly alongside the documentation.
All web documentation and `@wondelai/harness-audit` are free. Buy the DRM-free 8-book bundle for offline reading, or hire Michał directly for fixed-price team delivery. Prices in USD. The total for your country, with any VAT or sales tax, is shown at checkout before you pay.
All 8 books in DRM-free EPUB & PDF (EN & PL) for focused offline reading.
$59.99once
12 months of updated editions. Downloaded files stay yours permanently.
Buy Ebook Bundle Download your booksPaid once. 30-day refund on a first purchase.
All 8 offline books for up to 10 engineers, on one company invoice.
$299.99once
12 months of edition updates for the whole team.
Buy Team BundlePaid once. 30-day refund on a first purchase.
Fixed-price work on your repository, delivered personally by Michał Jaskólski.
From $499
excl. VAT, invoiced by Wondel.ai
See services for teamsA free intro call first
Read every guide and run `@wondelai/harness-audit` free without creating an account.
On a first Ebook Bundle purchase, for any reason. Email support and it is refunded.
Every EPUB and PDF you download stays yours permanently after the 12-month update window.
Polar issues an invoice immediately for bundles; Wondel.ai invoices consulting engagements.
FIELD NOTES · LOW VOLUME
Get new architecture essays, major @wondelai/harness-audit releases, and an immediate starter checklist in your inbox. Zero fluff, unsubscribe in one click.
Each of these names for building software with AI, and the one that means a different job, has its own page, defined and set against vibe coding.
Vibe coding vs agentic engineering, the long guide All glossary terms
Something else? Write to support@developertoolkit.ai and the author answers.
Yes — 100% open with no article meter and no paywall. You can read every guide and run @wondelai/harness-audit (MIT) free. Revenue comes from readers who want the 8 DRM-free offline books (bundle or Amazon) and teams who bring Michał in for fixed-price diagnostics, workshops and rollouts.
@wondelai/harness-audit inspect, and does it upload code? The default scan runs 100% locally in under 2 seconds with zero API keys and zero network calls. It checks 6 Factory Stations (Intent, Harness, Loop, Graph, Verification, Release) and caps your Autonomy Level when critical gates are missing. Optional --deep, --interview, and --prompt modes run only when you ask.
Official docs describe one tool’s features and fast-changing flags — and we link directly to them for those details. This canon teaches the cross-tool mindset, architecture, verification gates and organizational rollout across Claude Code, Codex and Cursor.
Karpathy published a transcript of his Sequoia Ascent talk (opens in a new tab) (30 April 2026), and one line of it sums up the difference: “Vibe coding raises the floor. Agentic engineering is about extrapolating the ceiling.” In vibe coding nobody reads the diff and no check stands in for the reader. This canon teaches agentic engineering: you review the spec and the test report, and checks you set up review the diff.
It starts there and goes further. AI pair programming is the way of working where you read every line the AI writes; AI-assisted coding is the umbrella over all of it, vibe coding included. The guides begin at pairing and move one loop at a time up the autonomy ladder, until checks you set up read the diff for you.
That works. Each tool has its own 10-chapter architectural core (and its own offline book), and the underlying workflow stays identical if you add another tool later.
No. You pay $59.99 once for all 8 books (EPUB + PDF, EN + PL) and 12 months of updated editions. Nothing renews automatically, and every file you download stays yours permanently.
Yes, 30 days from your first purchase of the Offline or Team Ebook Bundle, for any reason. Email support@developertoolkit.ai and it is refunded, as set out in the Terms.
Yes. Polar, the merchant of record, issues an invoice with the VAT or sales tax for your country right after checkout. Team consulting services are invoiced directly by Wondel.ai.
Your work moves to the spec, the architecture and the stop conditions. In the published transcript of his Sequoia Ascent talk, Karpathy says you are “in charge of taste, engineering, design, and whether the system makes sense.” This canon is about doing those well.
One tap, no sign-in.
What are you here for?
What is holding you back?
How does it read?
Thanks. If you left an email, Michał will reply.
Answering your questions?
One tap each. Close whenever you like.
no sign-in
“level 2, and every level after it, feels like you are done. But you are not done”
Dan Shapiro, “The Five Levels” (opens in a new tab), 2026-01-23
MIT open-source CLI · 100% open web docs · Fixed-price B2B diagnostics & workshops.
Michał replies, not a bot. The address is used for this reply only.