Sources
AI Build Kit grew from the private claude-kernel-builder and its output, claude-kernel, then from the private agentic-dev-skills, and then into this kit.
This list identifies ideas used in the kit and the files that implement them.
Records and planning
Section titled “Records and planning”| Source | Contribution | Where it is in the kit |
|---|---|---|
| 28 Days of Lovable, by Lazar Jovanovic | A single authoritative project document, with a next step and a named check. | The masterplan and the process for recording work, .agents/skills/setup-ai-build-kit/templates/masterplan.md, .agents/skills/records/SKILL.md. |
| Alex Lavaee, the new software lifecycle | Record decisions in documents, with code implementing the current decisions. | Project records and changes to them, docs/PHILOSOPHY.md, .agents/skills/records/references/masterplan-changes.md. |
| Alex Lavaee, from RPI to QRSPI | Build complete features in small pieces rather than implementing one technical layer at a time. | Pieces small enough for one session and the shaping process, .agents/skills/records/references/pieces.md, .agents/skills/shape/SKILL.md. |
| Dex Horthy, Advanced Context Engineering for Agents | Build each piece across the parts needed for a working feature, rather than completing one technical layer at a time. | Pieces small enough for one session, .agents/skills/records/references/pieces.md, .agents/skills/shape/SKILL.md. |
| Matt Pocock’s skills | Offer a provisional answer with each question; use a small prototype to resolve a decision. | Interview and prototype steps, .agents/skills/clarify/SKILL.md, .agents/skills/clarify/references/decision-prototype.md. |
| Matt Pocock’s skills | Reproduce a fault before attempting a repair and choose the workflow according to the type of question. | .agents/skills/repair/references/reproduce.md, .agents/skills/triage/SKILL.md. |
| Matt Pocock’s skills | Keep each piece within one session and place implementation detail beneath a plain-language description. | .agents/skills/build-piece/references/build-limits.md, .agents/skills/records/references/pieces.md. |
| GitHub Spec Kit | Identify requirements that no piece covers, without changing the records. | The requirements coverage review, .agents/skills/records/references/coverage-read.md. |
| OpenSpec concepts | Include proposed changes to project records with each piece and apply them when the work is saved. | .agents/skills/records/references/masterplan-changes.md. |
| Kiro’s proposal to record what a spec was checked against | Record the saved version of the code against which a document was checked. | .agents/skills/records/scripts/trued-mark.sh. |
| Beads | Link new work to the build that found it. | The Found while building link, .agents/skills/records/references/pieces.md. |
| GitHub Copilot’s agentic memory | Use concise, sourced facts only after checking them against the current project. | .agents/agents/researcher.md, docs/MAINTAINING.md. |
| Damien C. Tanner, how Toyo works | Judge a change by what it could damage; use a prototype as a brief. | The risk record and fit check, .agents/skills/fit-check/, .agents/skills/setup-ai-build-kit/templates/masterplan.md. |
| “Don’t read the code” | The person need not read code; check the result against the brief in a fresh context. | .agents/skills/second-opinion/SKILL.md, docs/PHILOSOPHY.md. |
| ReqToCode and StrictDoc | Treat an outdated reference between a written requirement and the code as a failure. | .agents/skills/quality-checks/references/document-read.md, .agents/skills/quality-checks/scripts/document-claims.py. |
| doc-drift and the trust-allocation study | Documentation may omit facts but must not invent them; use automated checks to verify names. | .agents/skills/quality-checks/references/document-read.md. |
| GitHub CODEOWNERS | List the relevant file paths beside each rule. | The risk-area map, .agents/skills/setup-ai-build-kit/templates/foundation/check-sensitive-areas.sh. |
| CCPM | Use GitHub issues as the task record. | Pieces are GitHub issues, .agents/skills/records/references/pieces.md. |
| Architecture decision records, as described by Martin Fowler | Record each settled choice with its reason, so a later session does not reopen it. | The ## Decided section on a piece, .agents/skills/records/references/pieces.md. |
| arc42 and C4 model | Diagram the tool and the external systems it connects to. | The system connections diagram in the masterplan, .agents/skills/setup-ai-build-kit/templates/masterplan.md, .agents/skills/records/SKILL.md. |
Review and teamwork
Section titled “Review and teamwork”| Source | Contribution | Where it is in the kit |
|---|---|---|
| obra/superpowers | Ask one question at a time; pause after three failed repair attempts to reassess the cause and explain the risk; use a reviewer who did not build the change. | .agents/skills/clarify/SKILL.md, .agents/skills/repair/references/fixing.md, .agents/agents/reviewer.md. |
| Builder.io on subagents, EveryInc compound engineering, and Maxi Contieri’s notes | A builder is a poor judge of their own work; independent review needs a separate run. | .agents/agents/reviewer.md, .agents/skills/second-opinion/SKILL.md. |
| Cloudflare, AI code review at scale | Keep the number of review findings small and check each finding again. | .agents/agents/refuter.md. |
| Claude Code subagents | Define each agent in a short file that specifies its tools and skills. | .agents/agents/*.md. |
| “How we vibe code at a FAANG” | Read a plan closely before building. | Shaping and the requirements coverage review, .agents/skills/shape/SKILL.md, .agents/skills/records/references/coverage-read.md. |
Checks and evidence
Section titled “Checks and evidence”| Source | Contribution | Where it is in the kit |
|---|---|---|
| Simon Willison, agentic engineering | A passing test suite is not proof by itself; report what ran and what happened. | .agents/skills/evidence-run/SKILL.md. |
| trailofbits/skills | A check that inspects nothing must fail. | The starter placeholder check, .agents/skills/setup-ai-build-kit/templates/foundation/checks.yml. |
| TDAD and TestPrune | Identify code affected by a change and run the tests that cover it. | .agents/skills/quality-checks/references/reach-check.md, quality.py reach. |
| dependency-cruiser | Use the import graph, a map of dependencies between code files, to identify code affected by a change. | depcruise, .agents/skills/quality-checks/scripts/quality.py. |
| StrykerJS incremental testing, mutmut, and Awesome Testing’s field note | Deliberately change code to test whether the tests detect the fault. | .agents/skills/quality-checks/references/test-strength.md, quality.py strength. |
| Vulture and lizard | Report only what a tool is certain of and use its stated complexity limit. | .agents/skills/quality-checks/scripts/quality.py, .agents/skills/quality-checks/references/trim.md. |
| sentrux, ArchUnit, and import-linter | Enforce rules that restrict which parts of the code may access other parts. | .agents/skills/quality-checks/references/boundary-rules.md. |
| Semantic code clone detection study | Limit the quarterly duplication check to copied code. | .agents/skills/quality-checks/references/waste-read.md. |
| ESLint ignore files | Exclude installed kit files from project linting, which checks code against style and error rules. | .agents/skills/quality-checks/references/check-floor.md. |
| Agentless, LLM-guided bisection, Cursor Debug Mode, and SWE-Doctor | Reproduce a fault, identify the saved change that caused it, and remove temporary diagnostic code. | .agents/skills/repair/references/reproduce.md, .agents/skills/repair/references/fixing.md. |
| OneUptime’s structured logging guide | Record each request’s actions in a consistent format, without sensitive content. | The request logging check before launch, .agents/skills/setup-hosting/scripts/run-recipe.sh; .agents/skills/repair/references/reproduce.md uses those logs. |
| ImpossibleBench and tests written after code inherit its flaws | Keep existing tests fixed; explain why a test changes. | .agents/skills/setup-ai-build-kit/templates/foundation/check-tests.sh. |
| Martin Fowler and Birgitta Boeckeler, Harness Engineering for Coding Agent Users | Use automated checks for testable rules and written instructions for decisions requiring judgement. | docs/PHILOSOPHY.md, .agents/skills/quality-checks/SKILL.md. |
Skills and standing instructions
Section titled “Skills and standing instructions”| Source | Contribution | Where it is in the kit |
|---|---|---|
| HumanLayer, Skill Issue: Harness Engineering | Start with a small set of instructions and review work in separate agent sessions. | docs/PHILOSOPHY.md, .agents/agents/. |
| jeffallan/claude-skills | Keep each skill short and load reference files when needed. | .agents/skills/*/SKILL.md and their references/ folders. |
| Agent Skills specification and Agent Skills best practices | Give each skill one job and load its references only when needed. | The layout of .agents/skills/*/. |
| IFScale | Evidence that models can miss instructions when too many apply at once. | Short command files and on-demand background skills, .agents/skills/, .agents/tools/check-structure.py. |
| ETH Zurich, Evaluating AGENTS.md, Configuration Smells in AGENTS.md, and Anthropic’s memory guidance | Keep standing instructions focused on practices that the code does not reveal, and remove unnecessary context. | .agents/skills/maintain/scripts/standing-lines.sh, .agents/skills/setup-ai-build-kit/templates/foundation/AGENTS.md. |
Running and release work
Section titled “Running and release work”| Source | Contribution | Where it is in the kit |
|---|---|---|
| Geoffrey Huntley’s Ralph technique and Clayton Farr’s Ralph Playbook | When a long run fails, change the instructions and run it again. | .agents/skills/implement/references/running-longer.md. |
| addyosmani/agent-skills | Build several pieces after an explicit decision enabling that workflow, and consult the supplier’s own documentation. | .agents/skills/implement/references/running-longer.md, .agents/skills/research/SKILL.md. |
| Agent Flywheel, by Jeffrey Emanuel | List destructive commands and distinguish written prohibitions from enforced restrictions. | .agents/guard/blocked-commands.md, .agents/skills/setup-ai-build-kit/templates/foundation/guard-secrets.py. |
| Carsten Jørgensen, harness engineering | Disable dangerous commands instead of relying on a warning. | .agents/skills/setup-ai-build-kit/templates/foundation/claude-settings.json, .agents/skills/setup-ai-build-kit/references/blocked-commands.md. |
| tw93/waza | Keep the command set small and fixed, adding capabilities to existing commands. | .agents/tools/kit-inventory.txt, .agents/tools/validate-kit.sh. |
| Husky and Codex non-interactive mode | Integrate push protection with the project’s Git hook configuration and run replay conversations through codex exec. | .agents/skills/setup-ai-build-kit/scripts/use-push-guard.sh, .agents/tests/replay/. |
| GitHub rulesets and branch protection and branch protection API | Require the project check to pass on the main branch, including for administrators. | .agents/skills/merge/scripts/protect-branch.py. |
| GitHub CLI environment and Codex configuration | Give a new Codex process access to the current GitHub sign-in without changing global settings. | .agents/skills/setup-ai-build-kit/scripts/, .agents/skills/setup-ai-build-kit/references/codex-github.md. |
| Codex hooks and Git configuration | Check GitHub access before work begins and give Git access to the session’s sign-in. | .agents/skills/setup-ai-build-kit/scripts/, .agents/skills/setup-ai-build-kit/references/codex-github.md. |
| GitHub CLI authentication refresh, authentication status, and pull request checks | Request workflow permission in the person’s terminal, check the active account, and treat pending checks as a separate result. | .agents/skills/setup-ai-build-kit/scripts/workflow-upload-check.py, .agents/skills/setup-ai-build-kit/scripts/check-tooling.sh, .agents/skills/merge/scripts/merge.py. |
Hosting and recovery
Section titled “Hosting and recovery”| Source | Contribution | Where it is in the kit |
|---|---|---|
| coolify-devops | Define a hosting request that the person passes to the server administrator. | .agents/skills/setup-hosting/recipes/nextjs-supabase-on-coolify/hosting-request.md. |
| Google’s SRE book, DORA delivery metrics, and Continuous Delivery | Treat recovery as a separate step and check the live tool after a release. | .agents/skills/live-copy/SKILL.md, .agents/skills/merge/scripts/merge.py. |
| BoringStack and Boring Stack | Use a short menu of proven defaults and treat an empty project as a normal start. | .agents/skills/setup-hosting/recipes/, .agents/skills/setup-ai-build-kit/references/completion-report.md. |
| Thoughtworks Technology Radar | Review tools on a fixed cycle and date each review. | .agents/maintainer-skills/stack-research/SKILL.md. |
| Intercom, The Safety of Speed | Plan rollback explicitly and keep changes small. | .agents/skills/live-copy/SKILL.md, .agents/skills/build-piece/references/build-limits.md. |
| OpenAI, harness engineering | Review documentation for descriptions that no longer match the code. | .agents/skills/quality-checks/references/document-read.md. |
Screen design
Section titled “Screen design”The starter copies its components and design variables with the
shadcn/ui command, using the
Radix option.
The copied files live in the starter’s src/components/ui/ folder. The
current form guide pairs
Field with react-hook-form and zod.
better-tailwindcss
checks colour classes against the design variables. Its
entryPoint setting
points at the Tailwind 4 CSS file. The starter uses the plugin’s own
no-unknown-classes and no-conflicting-classes rules.
Playwright’s axe guide
supplies the sign-in scan in e2e/accessibility.spec.ts. A separate assertion
checks that the email label is visible, since axe accepts labels hidden from
sight. Passing these checks does not prove a screen is accessible.
| Source | Contribution | Where it is in the kit |
|---|---|---|
| Vercel Web Interface Guidelines, Impeccable craft floor, GOV.UK Design System, Nielsen Norman Group, Carbon, Shopify Polaris, and WCAG 2.2 | Interface rules for visible labels, errors, keyboard focus, spacing and empty states. | .agents/skills/screen-check/SKILL.md. |
Licence notice
Section titled “Licence notice”The screen-check skill adapts some interface rules from W3C’s “How to Meet WCAG” quick reference, and W3C’s licence requires this notice.
Copyright © 2023 W3C®. This software or document includes material copied from or derived from How to Meet WCAG, Quick Reference.
Project checks in CI
Section titled “Project checks in CI”GitHub’s branch protection guide
warns that job names must be unique across workflows. The starter reports
recipe-check beside the foundation’s project-check, and the branch
protection script requires both for a recipe project.