Raw capture — raw/spec-driven-development-case-study.md
url: https://felipefontoura.com/articles/spec-driven-development-case-study/ title: "Spec-Driven Development Case Study: 13 Apps in 70 Days, Solo, with AI" author: Felipe Fontoura date_fetched: 2026-07-18 date_published: 2026-06-11
Author: Felipe Fontoura Publication Date: June 11, 2026 Tags: Spec-Driven Development, AI Agents, Case Study, Claude Code, Fintech
The article describes building a production-grade crypto payment platform — a PIX payment gateway for Brazil, an OTC exchange engine, and on-chain settlement on Bitcoin's Liquid Network — entirely solo over 70 days using AI agents guided by written specifications (Spec-Driven Development or SDD).
The brief was a complete, production-grade crypto payment platform — not a prototype. The author argues that the main objections to SDD come from toy projects where it seems like overhead, but notes that "A 10-hour solo project with no compliance pressure, no real balances, and no external deadline does not need a 28-file spec corpus."
He contrasts "prompt-and-pray" with specification: "it builds something that looks complete and passes a shallow review, then fails when a real edge case arrives."
Fontoura cites Anthropic's own guidance for Claude Code — a four-step loop of explore, plan, implement, commit — noting their reasoning that "letting Claude jump straight to coding can produce code that solves the wrong problem."
He references a 2025 METR randomized trial where experienced developers with AI were "19 percent slower than without it, despite expecting to be 24 percent faster." His conclusion: "Capability amplifies direction. It does not supply it."
The mechanism was a standing specification corpus of 28 markdown specs across 12 domains, plus 8 custom slash commands, all living in the repository and loaded as agent context each session.
Every capability began with a written specification covering requirements, design, and tasks. Agents executed against the spec. When output was wrong, the author fixed the spec, not the prompt.
The spec corpus included domains for auth, exchange pricing, database design, testing conventions, UI tokens, codebase conventions, and custom commands. The auth domain, for example, covered "five roles, nineteen scopes, eighteen granular permissions" plus machine-to-machine flows.
He explains that "A single developer does not hold a 13-app fintech in their head. The spec holds it. The developer holds the spec."
The four-step loop repeated per capability:
He calls the fourth step "where the discipline lives," warning that "Patch the code, leave the spec untouched, and the next time that module is regenerated or extended, the agent rebuilds from the spec and reintroduces the same error."
The OTC exchange pricing engine used a two-phase model with a peg rate and a VWAP window requiring at least five confirmed trades in 24 hours. The critical precision rule required "all money arithmetic in 8-decimal integers (satoshi precision)" with float anywhere being a test failure, not a lint warning.
The spec included a fallback chain across five market sources, asymmetric per-pair spreads, and an explicit out-of-scope list that stopped the agent from adding unrequested features twice.
The idempotency rule was frozen before any charge endpoint code existed: a UNIQUE(merchant_id, idempotency_key) constraint enforced at the database level. He notes "Without a spec, that constraint lives in your head. It falls out on the next regeneration."
Without specs (prompting): Float arithmetic and rounding drift, silent scope creep, edge cases surfacing in production, idempotency constraint held only in memory, no record of decisions.
With specs: 8-decimal integer math enforced by test failure, negative scope prevented unrequested features, edge cases caught at review, database-level unique constraint, every decision traceable in git.
The author states: "The agent's capability was identical in both columns. The spec is the only variable."
Fontoura is explicit about limitations:
His structural argument: "an AI agent has no memory between sessions, and the spec is the external memory you give it."
He concludes: "Without the specs, the agent is a fast bricklayer with no blueprint. With them, it is a senior engineer with perfect recall of every decision you made."
The article ends with links to further SDD resources and a newsletter tagline: "Don't Code, Specify. A weekly dispatch from where AI agents meet real production."
Final line: "The code wrote itself. The specs did not. That is where the seventy days went, and it is why they were enough."