Lesson 3 · 18 min
Listing pages, news article bodies, page summaries: the agent writes from a source, in batches. The risk is that it embellishes and makes things up. Here you call the model through the subscription, in read-only mode, read the first JSON in the reply, and run everything through a check that discards the made-up item.
Start the lesson
Goal: call Codex or Claude through the subscription to extract from a source and discard, without fixing, any item that fails the check.
Tap each block.
The model can rewrite and organize; the check ensures it didn't create facts.
Tap each step.
INEMA's news text generator calls Codex with the request through standard input, a read-only sandbox, an ephemeral session, and output to a file. Close the input after sending the request: otherwise the process keeps waiting.
In Claude Code, the same request runs in the project folder, with AGENTS.md forbidding writes outside the output folder.
On 10/02/2026, Codex replies with duplicated JSON broke the batch reading. The smallest fix was to read only the first complete object (the primeiroJson function).
It wasn't switching models or rewriting the generator: a ten-line safeguard.
Of the 159 news articles, 113 got a text body. The 46 without text stayed that way because the source requires a login (25), has little text (18), or returned a 404 (1), and they stay with noindex, out of the sitemap.
Nobody wrote the missing parts by hand: better one page fewer than a page with a made-up fact.
The first round of listing pages used an authorized API: 40 listing pages extracted and validated for US$ 0.031. The following batches moved to Codex through the subscription.
The check is the same on both paths; what changes is who pays for the call.
pedido.txt = extractive prompt (course 2, lesson 5) + the source text. valida-ficha.mjs is a template that uses the validar_extracao.mjs library.
codex exec --skip-git-repo-check --ephemeral --sandbox read-only -C "$PWD" -o saida.txt - < pedido.txt node valida-ficha.mjs fonte.txt saida.txt
// passes each item of the listing page through the check (uses validar_extracao.mjs as a library)
import fs from 'node:fs'
import { apoiada, numerosNaFonte, tituloNaFonte } from './validar_extracao.mjs'
const [fonteArq, fichaArq] = process.argv.slice(2)
const fonte = fs.readFileSync(fonteArq, 'utf8')
const ficha = JSON.parse(fs.readFileSync(fichaArq, 'utf8'))
let validos = 0
for (const m of ficha.estrutura || []) {
const ok = tituloNaFonte(m.titulo, fonte) && numerosNaFonte(m.detalhe || '', fonte)
console.log(`${ok ? '✓' : '✗ descarta'} módulo: ${m.titulo}`)
if (ok) validos++
}
for (const campo of ['aprende', 'publico', 'constroi']) {
for (const f of ficha[campo] || []) {
const ok = apoiada(f, fonte) && numerosNaFonte(f, fonte)
console.log(`${ok ? '✓' : '✗ descarta'} ${campo}: ${f}`)
}
}
console.log(validos >= 3 ? `ficha enriquecida (${validos} módulos válidos)` : `não publica como enriquecida (${validos} módulos válidos)`)function primeiroJson(txt) {
const ini = txt.indexOf('{')
let prof = 0, str = false, esc = false
for (let i = ini; i < txt.length; i++) {
const c = txt[i]
if (str) { if (esc) esc = false; else if (c === '\\') esc = true; else if (c === '"') str = false; continue }
if (c === '"') str = true
else if (c === '{') prof++
else if (c === '}' && --prof === 0) return JSON.parse(txt.slice(ini, i + 1))
}
throw new Error('JSON incompleto')
}$ node valida-ficha.mjs fonte.txt saida.txt ✓ módulo: Módulo 1: Preparar o barro ✓ módulo: Módulo 2: Centralizar no torno ✓ módulo: Módulo 3: Abrir e levantar a peça ✓ módulo: Módulo 4: Acabamento e secagem ✓ aprende: Amassar, tirar o ar e pesar as bolas de argila. ✓ publico: quem nunca sentou num torno e quer fazer as primeiras peças utilitárias. ✓ constroi: copos e tigelas simples, prontos para esmaltar. ficha enriquecida (4 módulos válidos)
Excerpt. The source was the sample page of a wheel-throwing ceramics course, with 4 modules and "8 encontros de 3 horas".
$ node valida-ficha.mjs fonte.txt saida-livre.txt ✗ descarta módulo: Módulo 1 — Prepare o barro para suas primeiras peças ✗ descarta módulo: Módulo 2 — Encontre o centro no torno ✗ descarta módulo: Módulo 3 — Dê forma ao seu primeiro copo ✗ descarta módulo: Módulo 4 — Cuide do acabamento e da secagem ✓ aprende: Posicionar mãos e corpo para centralizar o barro no torno. ✗ descarta constroi: Seu primeiro copo modelado no torno. não publica como enriquecida (0 módulos válidos)
Excerpt. Same source, but the prompt was "escreva a ementa de forma atraente". The model rewrote every title and, in module 4, added up "totalizando 24 horas": the number is not in the source.
Pick an item on your site that has its own page with text.
Save the page text in fonte.txt, build the pedido.txt with the extractive prompt, and run the Codex call (or the same request in Claude Code). Repeat with a free-form request, "de forma atraente" (in an attractive way), to compare.
Done when: you have saida.txt and saida-livre.txt, both in JSON, and no API key was used.
Copy validar_extracao.mjs and valida-ficha.mjs into the folder and run them on both outputs. Write down what was discarded and why.
Done when: a note lists each discarded item with the reason (title, words, or number outside the source), and only the listing page with 3 or more valid modules is marked for publishing.
The explanation appears after your answer.
Question 1 of 4
What is the read-only sandbox option in the Codex call for?skill
Question 2 of 4
In the "attractive" prompt, the model wrote "totalizando 24 horas". The source said 8 sessions of 3 hours. What does the check do?
Question 3 of 4
The model's responses came back with duplicated JSON and the batch broke. What was the smallest fix?
Question 4 of 4
46 news items were left without body text because the source required login, had little text, or returned a 404. What was done with them?
Choose an option. The explanation appears after your attempt.