Development Agent

Ship fast, hand off clean, no lock-in.

AI services · internal AX tools · Web3 dApps handed off as 4 deliverables — code plus AI-readable docs, E2E guardrails, and operational runbooks.

Deliverables

Self

Maintained

Weekly

Progress

01 · WORKING CODE

// auth.ts
async function signIn(token) {
  - if (result === null) return
  + if (!result?.ok) throw err
  return result.user
}

02 · TWO-LAYER DOCS

# CLAUDE.md

## Architecture

- Next.js 14 · Postgres

## Skills

- migrate · deploy · e2e

## Invariants

- 0 LLM calls

03 · E2E SUITE

42 / 42 PASSALL GREEN

auth · sign-in

1.2s

billing · checkout

3.8s

dashboard · usage

0.9s

04 · RUNBOOK

$ pnpm run deploy

✓ build (12.4s)

✓ migrate (2.1s)

✓ e2e smoke (8.7s)

→ rollback: ./rb.sh v0.4.2

→ on-call: pagerduty/dev

What we build

3 areas. All ones we use ourselves every day.

AI Services

Production-grade AI products for internal or external users — not demos.

Internal AX Tools

Accelerate team workflows. AI Usage Tracker itself is one.

Web3 dApps · Tokenomics

From contracts to token sales. 40-50K DAU in production.

What you get

Not just working code.

Handed off as a 4-piece set — code · AI-readable docs · E2E guardrails · operational runbook.

Working Code

Ships to your GitHub

// auth.ts
async function signIn(token) {
  - if (result === null) return
  + if (!result?.ok) throw err
  return result.user
}

Two-Layer Docs

AI docs · human docs split

# CLAUDE.md

## Architecture

- Next.js 14 · Postgres

## Skills

- migrate · deploy · e2e

## Invariants

- 0 LLM calls

E2E Suite

Playwright + CI integrated

42 / 42 PASSALL GREEN

auth · sign-in

1.2s

billing · checkout

3.8s

dashboard · usage

0.9s

Operational Runbook

Deploy · incident · monitoring

$ pnpm run deploy

✓ build (12.4s)

✓ migrate (2.1s)

✓ e2e smoke (8.7s)

→ rollback: ./rb.sh v0.4.2

→ on-call: pagerduty/dev

02 · Two-Layer Docs · in depth

Two doc layers — so both humans and AI understand the system.

Human docs stay as-is. AI docs split into 4 startup + 4 in-progress — so agents like Claude Code absorb system context and decision history immediately.

A · Human

Standard artifacts

Docs developers and managers read directly

README.md
Project overview · setup
ADR (decisions)
Key architectural decisions
API docs
Endpoints · schemas
Ops guide
Deploy · incident · monitoring

B-1 · AI · Startup

Startup context

4 files an agent reads on first encounter with the project

CLAUDE.md
AI agent operating rules
CONTEXT.md
Project context · tech stack
SPEC.md
Feature specs · invariants
DESIGN.md
Design system · UI decisions

B-2 · AI · In-progress

In-progress logs

4 files that accumulate living history every session

state.md
Current focus · next actions
worklog.md
Per-session work log
decision.md
Alternatives compared, then decided
backlog.md
Backlog · priorities

See sample — actual state.md / decision.md excerpts (anonymized)

state.mdAI · in progress

### Current focus
ACME payment pipeline — Stripe webhook
retry queue + idempotency validation

### Next up
1. failed_webhook table + retry policy
2. /api/billing/webhook idempotency-key check
3. E2E: 3 scenarios via stripe-mock

### Blockers
None

### Needs human call
- Refund policy — prorated vs full
- Membership downgrade timing

decision.mdAI · in progress

## 2026-XX-XX: Payment retry — queue vs cron

- **Choice**: failed_webhook table + cron retry
- **Alternatives considered**:
  - (A) Bull/SQS queue — fast but adds infra
  - (B) Table + cron (picked) — simple, easy to monitor
- **Reasoning**: traffic is small (<10/s).
  Ops overhead lower than introducing a queue
- **Impact**: schema · webhook route · cron
- **Rollback**: can migrate to a queue later

03 · E2E Suite · in depth

3-tier test scenarios — from PR guards to full coverage.

We don't run every test the same way. Split into 3 tiers by when and what to verify — fast PR guards, daily core flows, periodic full coverage.

Tier 1

CI/CD smoke

Every PR · blocks merge · ~2 min

5-8 critical scenarios

Tier 2

Daily key flows

Daily cron · ~10 min

30-50 key user flows

Flows users actually use often. Runs nightly → results ping Slack. Broader than Tier 1.

Tier 3

Full Suite

Weekly · pre-release · ~30 min+

600+ full cases

Edge + exception scenarios — full coverage. Run before releases or after major changes. Produces automation / manual / blocked / skip classified report.

Sample report

Tier 3 full report — per-module automation rate · execution results · ambiguous / blocked / manual / skip classification.

See sample QA report

Zero lock-in

You run it yourself when we're done.

Claude Code alone is enough — AI-readable docs + E2E guardrails as the safety net.

Problem

Locked in after handoff

Most agencies keep you on retainer even after handoff.

Solution

AI-readable Docs + E2E

CLAUDE.md / Skills let AI understand the system; E2E validates the golden path every time.

Proof

AI Usage Tracker

Our own tool deployed on client infrastructure. Full source public, zero external dependencies.

Our AX benchmark

We measure and improve our AX every day.

We're the team that uses AI Usage Tracker to measure and improve our own team's AX daily. The system we built through AX — we ship it to you as-is.

Our internal AX measurement tool

AI Usage Tracker

The tool we use to measure and improve our own AX (AI Transformation) level every day. Team AI usage · efficiency · industry benchmarks — we look at these daily and decide what to do next. Built by us, used by us daily — and the AX experience we accumulated building it flows directly to your organization.

Live · team demoOpen →

Telegram · LINE Mini dApp

40-50K DAU

4+ years running Web3 dApps in production. Smart contracts · DEX · Bridge · Governance · Launchpad — full stack. IDO/INO token sale execution experience.

Battle-tested stack

Next.js 14PostgreSQLDrizzle ORMAuth.jsPlaywrightTailwind

How we work

Progress you see week over week.

Week 0-1

Scope + first working code

Requirements aligned · CLAUDE.md skeleton · one golden path working end-to-end.

Week 2-N

Weekly demos + GitHub merges

Working demo every week. Priority retro · agree on next week's scope.

Week N

4 deliverables handed off

Code · Docs · E2E · Runbook fully handed over. No-lock-in wrap-up.

30-min free scoping

Requirements aligned · fit assessed.

Week-1 preview

See first-week results, then commit.

No-lock-in exit

After 4-deliverable handoff, no forced retainer.