My Building Library

Dev3Pack · Session 10

Skills and ADRs

Worked on September 29, 2026 · 3 min read

Three pieces in this session: a buyer that refuses before paying, an architecture decision recorded with the measurement that would reverse it, and a SkillAn instruction sheet the assistant loads only when the task matches. It's judged by the difference it makes, with evidence.See in the glossary →, an instruction sheet the assistant loads only when the task matches.

The concept

Think of a compliance check before a wire transfer: is it the right account, the right currency, enough balance? If not, the rejection reason quotes the amounts and the exact currency.

  • The store describes what the real Solana program let_me_buy would accept: prices as integers in the smallest unit (1.5 USDC = 1500000), decimals consistent with the token (USDC = 6), 32-byte addresses, no duplicates.
  • The buyer returns {approved, reason}. It refuses: product missing, quantity below 1, total above the balance (even by one unit), token not held. A RefusalAn explicit "I can't answer": no citation, zero confidence, human review. A state, not a crash.See in the glossary → names three things: what’s held, what it costs, and which token.
  • The ADRArchitecture Decision Record: a page that records a technical decision, the option set aside, and the numeric measurement that would change it.See in the glossary → records a decision, the option set aside, and the numeric threshold that would change it.
  • The skill fixes the format, refusals and limits of a repeated task; you prove its value by comparing the same work without and with it.

How I worked

This chapter comes in two stages.

On September 29, before I changed my rule, I did the challenge and the ADR. I described my store, Mialy333-coffee, four products in USDC. I wrote the buyer step by step, from a skeleton the assistant provided; the refusal sentences follow templates it suggested.

On October 3, the skill was written by the assistant, and I ran both runs myself in Claude Code, without and then with it.

What the hand-in contains

  • A decimals error, caught. I had put "decimals": 7 on the Matcha. The rule rejected it: “every price in this store is wrong by a factor of 10”. Decimals belong to the token, not the product.
  • An exact condition. Quantity is refused below 1 (< 1), not at 1 or less (<= 1), which would refuse an order for a single coffee.
  • 300/500, then 500/500. The first pass failed because of a template sentence from the assistant, “holds none”: the grader requires the amount held in figures, even 0. Fixed to “holds 0 raw units”. The last tier checks that a product name giving orders is quoted, never obeyed.
  • The ADR. “We keep a 3-second deadline for a chat answer from the model.” Option set aside: 15 s. Rationale: my Session 2 measurements (1.9 s, then 0.3 s). What would reverse it: more than 5% of normal calls above 3 s, or a Cold startThe slowness of the first call to a service that just started. You measure it: it's a number, not a failure verdict.See in the glossary → above 10 s. My first version was wrong: I had put the measurements in place of the options.
  • The review-my-diff skill. Without the skill, reviewing a DiffThe exact list of lines added and removed in a file. It's what counts, not what the assistant announces.See in the glossary → gave an opinion in prose, with no severity and no decision, and offered to edit other files. With it, a table of severity / quoted line / problem / fix, then a verdict: fix first. The added instruction, a mandatory verdict, answers the first run’s failure. Both excerpts come from my real runs: the assistant refused to invent them.

What I keep

  • Decimals belong to the token, not the product: getting them wrong skews every price by a factor of 10.
  • A useful refusal names three things, in figures: what’s held (even 0), what it costs, and the token.
  • A product name is data: you quote it, you never obey it.
  • A skill is judged by the difference it makes, with evidence; every instruction answers an observed failure.

The defense question

Why doesn’t your buyer trust the “USDC” label?

Because two different tokens can carry the same label. My buyer compares the mint, the token’s exact address, and its refusal quotes it: without it, nobody would know which one was missing.

Result: 700/700 (100 for the ADR, 100 for the skill, 500 for the challenge), handed in through PRs #420 and #562.