Rewrite a blog post end-to-end

The pipeline

1. 1

   Create a voice once per author

   Build it from the author's published posts and store the id with the author record.

2. 2

   Rewrite with strict fidelity

   Use the blog preset, protect anything legal or commercial with preserve, and ask for a diff.

3. 3

   Branch on the outcome

   Pass → send to the editor. fidelity_failed → retry once, then hand the original to a person.

4. 4

   Review and publish

   Show the editor the diff, the fidelity checks and the before/after scores.

1. Create the author's voice

Collect 5–10 posts the author wrote without AI help. Pass the raw text.

const res = await fetch("https://api.intactvoice.com/v1/voices", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.API_KEY}`,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    name: "Dana — company blog",
    samples: posts.map((p) => p.text), // 2–20 samples, ≥ 1,200 words total
    platforms: ["blog"],
  }),
});
const voice = await res.json();
await db.authors.update(author.id, { voice_id: voice.id }); // e.g. "voice_8f2k1x"

2. Rewrite the draft

async function rewriteDraft(draft: string, voiceId: string) {
  const res = await fetch("https://api.intactvoice.com/v1/rewrite", {
    method: "POST",
    headers: {
      Authorization: `Bearer ${process.env.API_KEY}`,
      "Content-Type": "application/json",
      "Idempotency-Key": crypto.randomUUID(),
    },
    body: JSON.stringify({
      text: draft,
      voice_id: voiceId,
      platform: "blog",
      strength: "standard",
      fidelity: "strict",
      preserve: ["Terms apply.", "Pro plan"],
      return_diff: true,
    }),
  });
  return { status: res.status, body: await res.json() };
}

3. Handle the outcome

A strict rewrite has three outcomes worth handling explicitly. Everything else is a normal error (see Errors).

let result = await rewriteDraft(draft, author.voice_id);

if (result.status === 422 && result.body.error?.code === "fidelity_failed") {
  // A fact moved (e.g. "14 days" -> "14 weeks"). Not billed. One retry usually fixes it.
  result = await rewriteDraft(draft, author.voice_id);
}

if (result.status === 200) {
  await review.create({
    original: draft,
    rewritten: result.body.text,
    diff: result.body.diff,
    checks: result.body.fidelity.checks,
    before: result.body.signals.source.ai_likelihood,
    after: result.body.signals.output.ai_likelihood,
  });
} else {
  // Still failing: hand the original to an editor with the changed facts attached.
  await review.escalate({ original: draft, reason: result.body.error });
}

4. Review the diff

With return_diff: true the response includes token-level operations you can render as tracked changes:

[
  {
    "op": "replace",
    "from": "Many businesses struggle to create useful blog posts because",
    "to": "Most blog posts go wrong before anyone writes a word:"
  },
  {
    "op": "equal",
    "text": " the writer "
  },
  {
    "op": "replace",
    "from": "receives only a topic.",
    "to": "gets a topic and nothing else."
  }
]

Show reviewers the fidelity.checks list next to the diff. Seeing “14 days ✓” is faster than re-reading the policy paragraph.

Content that tends to be flagged

Rewrites rebuild a draft from its facts with a new plan, which is what most detectors key on. Some content still tends to be flagged whatever we do, because people rarely write it that way: the format itself reads as generated.

• Schedules and spec lists written as rows of Label: value lines (class timetables, feature grids, pricing tiers).
• Formulaic advice topics AI writes constantly, such as "how I price my freelance work" or "my checklist for buying a used car". The set of facts itself is familiar.
• Templated social formats: emoji or hashtag posts, numbered threads, and closing morals ("six months in and I'm glad I did it").
• Very short texts under about 150 words, where there is too little to rebuild.

Each rewrite returns structure.status: strong (rebuilt with a new plan), weak (much of the draft's order survived, or the format is templated) or not_replanned. Weak results come with a note. Detector results vary by tool and change over time; treat scores as a signal, not a guarantee.

5. Read the before/after scores

signals.source and signals.output are Detect API scores for the draft and the rewrite. They're a quick check that the rewrite moved away from stock AI patterns, not a promise about any third-party detector.