Ironclad — Senior Software Developer

The problem

AI coding assistants write code that looks right. It puts currency in a float, buries a tax rate in an if-statement, writes one "date" column for five different dates, deletes rows with no trace of who did it — and then tells you the result is production-ready.

Here is one that reproduces today: 24.67x4 + 12.09x7 + 23.73x3 at 23% comes to 313.03 in floating point. The customer was charged 313.04. One cent short, silently, on every order, until a bookkeeper cannot close a reconciliation.

What this skill enforces

• Volatile rules stay in config — rates, thresholds, limits and deadlines live in a table keyed by jurisdiction and effective date, never in code.
• Money is exact — decimal or integer minor units, currency code stored beside every amount, FX rates persisted with their date.
• Dates are explicit and typed — event, effective, posting, due, expiry; stored UTC, rendered local, DST and month-end handled deliberately.
• Audit trail on every change — who, when, before and after, and a reason. Records are voided or superseded, never silently deleted.
• Privacy and security by design — least privilege, parameterised queries, secrets outside source control, personal data masked in logs.
• AI assists, humans decide — nothing filed, paid, published or deleted without a person saying yes.
• Honest about readiness — it states what it tested, what it assumed, and what still needs real data. It will not tell you the code is perfect.

What is inside

SKILL.md — the seven non-negotiables, an eight-step method for working any request, a severity-ordered review protocol, and separate output shapes for builds and reviews.

references/domains.md — nine domain patterns with data models, controls and the defects reviewers actually find, plus seven app templates.

references/stacks.md — twelve stacks with concrete rules, and a selection guide for when no stack has been chosen.

references/worked-example.md — one end-to-end example of the expected answer.

Reference files load only when the request touches them, so the skill stays cheap in context until it needs the depth.

Who it is for

Developers, technical founders, consultants and agencies shipping software other people depend on — and anyone directing an AI agent to write code they will have to maintain afterwards.

What it will not do

It supplies no tax rates, statutory deadlines or regulatory figures. It insists they stay configurable and verified against the official source for your jurisdiction. It is not legal, tax or regulatory advice.

Tested before release

Run against live build and review requests before publishing, with every finding folded back into the skill. Version 1.0.0.