Immediately usable artifacts from jyoung.dev essays on engineering practices for AI coding agents. Each tool operationalizes one post's framework — the essay is the argument, the artifact is the move.
Copy a skill directory into your repo's .claude/skills/ (project-wide) or ~/.claude/skills/ (all your projects):
cp -r skills/audit-claude-md /path/to/your-repo/.claude/skills/Each skill's script also runs standalone without Claude Code.
| Artifact | What it does | Canonical post |
|---|---|---|
agent-task-go-no-go |
Runs a candidate task through the five-gate go/no-go pass — runnable check, human-time estimate, verification cost, verifiable done, blast radius — and returns the delegate/decompose/keep verdict at the first failing gate | What AI Coding Agents Are Actually Good For (And When to Skip) |
audit-claude-md |
Inventories your CLAUDE.md loading tiers — always-loaded vs lazy vs skills, @imports expanded, token estimates — then routes every section to the tier that should own it | How to Structure CLAUDE.md: It's a Loading Policy, Not a Document |
audit-jit-retrieval |
Audits an agent's JIT retrieval surface against the five-row failure ledger — dead references, tool-description routing, empty semantic results, uncapped loops, fallback re-injection — reporting which failures are loud and which are silent; bundles a standalone tool-catalog audit script | Where Just-in-Time Context Retrieval Silently Breaks |
decompose-epic |
Converts one epic into a task graph whose nodes each carry an exclusively owned write glob, a single-owner interface contract, explicit acyclic edges, and a mechanical acceptance predicate — with a standalone checker that tests the globs against the real repo for overlap, contracts for two owners or none, edges for cycles, and acceptance criteria for prose, and calls a one-node graph what it is: run it monolithic | Task Decomposition for AI Coding Agents: Draw the Graph First |
isolation-gate |
Walks a proposed multi-agent split through the two-gate flowchart — isolation need first, coordination readiness second — and returns a verdict at the first failing gate | When One Agent Stops Being Enough: The Isolation Gate |
lint-claude-md |
Runs every line of a CLAUDE.md through the post's per-line decision table — instruction-ceiling budget, self-evident/negative/vague/hook-shaped flags, canary probe check — and emits a keep/rewrite/cut/hook verdict report | CLAUDE.md Instruction Ceiling: Maintained Config, Not a README |
loop-preflight |
Preflights an unattended agent loop before you walk away: checks the wrapper for the max-iteration, no-progress, and spend brakes plus externalized state and a non-self-graded check, and prices the always-on context re-paid every iteration | Loop Engineering Breaks Your Single-Shot Context Playbook |
size-agent-task |
Runs a proposed agent task through the five-gate sizing flowchart (one-sentence diff, independent verifiability, layer crossing, exploration cost, decomposition floor) and bundles a standalone script that measures any git diff against the PR-sizing research thresholds | How to Size Tasks for AI Coding Agents |
tier-agent-authority |
Classifies every action an agent can take — Claude Code permission rules mechanically, ambient credentials by inventory — into four authority tiers by reversibility and blast radius, flags any action holding a control looser than its tier's default, and proposes a measured graduation ledger | Tier Your AI Agent's Production Authority by Task Risk |
verify-agent-pr |
Runs an agent-authored PR through the six-move verification checklist as a fresh grader — evidence over self-report, tests read as claims, spec-forward verification, scope diff, blast-radius depth — with a standalone diff-triage script | How to Verify AI Coding Agent Output: A Reviewer's Framework |
coordination-contract-lint |
Validates a JSON handoff contract against the post's schema (state enum, objective, ownership, observations, mutations, verification, approval_gates) and enforces the done-state promotion guards, exiting non-zero on a malformed or incomplete contract so coordination logic is testable in CI instead of buried in a prompt | Coordination Is an Architecture Layer, Not a Prompt Instruction |
cost-guardrail-checklist |
Pre-ship checklist for agent spend controls: seven layers, each requiring both an attribution dimension (which task spent) and an enforcement ceiling (what refuses the next call), with a verified LiteLLM per-session config | You Can't Cap What You Can't Attribute: Per-Task Cost |
done-gate |
Runs a .done/.json predicate file (decompose-epic's command, test, file_exists, glob_count plus file_absent, grep_count and unchanged_since) against the repo's end state: --hook exits 2 with failing predicates on stderr for a Claude Code Stop or SubagentStop hook, --ci exits 1 with a summary for a required status check, and lint flags prose criteria and exit-code predicates with no unchanged_since guard on the tests | Your Agent's "Done" Is a Self-Report. Make It a Predicate. |
edge-justification-worksheet |
Per-edge worksheet that forces every proposed agent-to-agent edge to name its forcing constraint (context pollution, parallelism, or specialization), its risk tier, its four-field delegation spec, its review node, and both bills — token and reviewer-minute — with a checker that deletes any edge holding an empty or invalid cell | The Topology You Can Review Is the Topology You Can Run |
harness-audit |
Read-only scan that answers the post's seven harness questions against a repo plus your user-scope config — permission rules by scope, the three enforcement buckets, sandbox keys including excludedCommands, registered hooks with their exit-1 fail-open paths, telemetry switches, and runtime version — marking every control DEFAULT, SET, or UNDETERMINED without grading any of them | Audit Your Agent Harness: The Deterministic Layer Nobody Reviews |
harness-portability-audit |
Read-only scan that prints the post's portability ledger for one repo: every harness artifact across Claude Code, Codex CLI, Gemini CLI and Cursor (instruction files, skills, slash commands, MCP client files, hooks, permissions, subagents, worktrees, dev container) sorted into PORTS or VENDOR-BOUND, with the documented bridge named where one is missing (no @AGENTS.md import, context.fileName unset, non-spec SKILL.md keys, .mcp.json url without type, hooks or permissions declared for one vendor only) | AGENTS.md vs CLAUDE.md: Swap the Agent, Keep the Harness |
model-swap-gate |
Decision checklist for coding-agent model swaps: error-analysis entry condition plus five gates (noise floor, instrument validity, harness parity, workload-weighted eval, staged rollout), each with a pass/fail predicate and the published measurement behind it | Stop Picking Your Coding-Agent Model Off a Leaderboard |
ownership-cost-gate |
Walks one build-vs-buy candidate through the five-gate ownership-cost flowchart — prices the 60/60 maintenance tail from your first-build estimate, then gates on differentiation, continuous verification, compliance-as-product, and the internal-build base rate | Build vs. Buy Agentic AI: Ownership Is the New Decision |
size-review-capacity |
Turns your reviewer count, focused review-hours, and PR volume into attention-per-PR and a clears/exceeds verdict for the next agent you want to add, with the post's six-gate scorecard as a copy-paste checklist | Review Capacity Is the Real Ceiling on Your Agents |
task-spec |
Copy-paste task-spec skeleton with the post's nine sections (goal, context, constraints, non-goals, acceptance criteria, verification), plus a lint script that fails on missing sections and flags vague acceptance criteria | The Anatomy of a Perfect AI Agent Task |
trace-field-survivability-audit |
Read-only field-presence audit against exported OpenTelemetry GenAI or OpenInference span JSON (array, JSON-Lines, or full OTLP export): flags every LLM-shaped span missing the post's four unrecoverable fields, decision content, resolved response model versus request-model-only, pinned prompt version or content, and the model's retirement date, and exits non-zero so it can gate CI the way the post's closing instruction asks for | Rank Your Agent Trace Fields by What You Can Never Get Back |
Each artifact directory carries an artifact.json (name, form, canonical post) and a Grounding section citing its sources. Skills with a scripts/ directory also run standalone, for example:
skills/audit-claude-md/scripts/audit-claude-md.sh [target-dir]MIT
