Reduce AI hallucinations in writing: prompts and workflows

What hallucinations look like in AI-assisted writing

Hallucinations happen when an AI writes something that sounds confident and specific, but is wrong or simply made up. The danger is the tone. A sentence can read like a verified fact even when it's just a guess.

You can often spot hallucinations by their "too neat" details. A draft suddenly includes exact numbers, precise dates, or perfect quotes you never provided. It may cite a study, a book, or an expert that doesn't exist.

Common signs include:

• Made-up statistics (for example, "72% of companies saw a 3.4x lift") with no source.
• Wrong names, titles, or organizations that sound plausible.
• Invented quotes in quotation marks with no traceable origin.
• Clean timelines and dates that don't match reality.
• Confident cause-and-effect claims ("X causes Y") when the evidence is unclear.

This happens because language models predict likely word sequences. When the model doesn't have enough information, it fills gaps with something that fits the pattern of a good article. Prompts that ask for authority ("include stats" or "add a quote") without supplying sources push the model to improvise.

The goal isn't perfection. It's to make sure risky claims are either backed by evidence or clearly labeled as unverified. A solid process keeps the draft useful while you stay in control of what gets stated as fact.

A simple rule helps: if a claim could affect a decision, reputation, or compliance, treat it as high risk and require a source before it stays in the text.

Define scope, audience, and claim limits first

If you want to reduce hallucinations, start before the first prompt. Tell the model what it's allowed to say, what it must avoid, and who it's writing for. Clear limits make the draft easier to verify and much harder to "invent."

Write a simple claim budget. It's not about being timid. It's about being explicit.

• Allowed: plain explanations, definitions, and summaries of provided sources.
• Allowed: clearly labeled opinions (only if you ask for them).
• Not allowed: statistics, rankings, named quotes, or "according to X" unless you provide the source text.
• Not allowed: dates, legal/medical advice, or pricing details unless you give an approved reference.
• Required: mark anything uncertain as "needs verification" instead of guessing.

Define the reader in one sentence. Example: "Busy marketing manager, non-technical, reading on mobile, wants a practical checklist." That one line changes tone and word choice more than most style prompts.

Set boundaries for scope so the model doesn't blur contexts.

• Geography: global, US-only, EU-only, or a specific country.
• Time: "as of 2025" or "only the last 12 months."
• Sources: "use only the pasted sources" or "use internal notes only."

Finally, choose the output type up front (blog section, news summary, glossary entry). When this is vague, models tend to pad with filler claims.

Use source-first prompts that require evidence

If you let an AI write first and justify later, you're inviting confident fiction. Flip the order: make it collect evidence before it makes claims.

Start by asking for a source plan, not a draft. The goal is a small set of sources (or placeholders) the model must rely on when it writes.

A simple source-first prompt pattern

Use a two-pass request: sources first, then the draft. Keep the drafting step locked to those sources.

1. Ask for a source list with 5-8 items maximum, each with: title, author/organization, date, and what it supports.
2. Require a citation key for each source (S1, S2, etc.).
3. Only after that, ask for the draft that cites S1-S8 after every factual claim.
4. If the model can't find or infer a source, it must write: "No source" and move on.

That "No source" label is the point. It highlights what needs checking and blocks the model from inventing a reference that sounds real.

Force a citation format (even if you verify later)

Pick a strict, boring format so you can scan it fast:

• Claim sentence ends with [S#] or [No source]
• Quotes must include speaker, context, and [S#]
• Numbers and dates must include [S#]

Example (you can paste into any tool or an API call):

Task: Write a 700-word section.
Step 1: Propose 6 sources (S1-S6). For each: title, publisher, date, what it supports.
Step 2: Draft the section using ONLY S1-S6. After every factual claim add [S#].
If you cannot support a claim, write [No source] and do not guess.
Output: sources list, then draft.

Constrain high-risk claims (stats, names, quotes, dates)

Most hallucinations hide inside details that sound credible: percentages, launch dates, "top 10" rankings, and exact quotes. Treat these as high-risk claims that need extra rules.

A practical guardrail is simple: if the draft includes numbers, dates, rankings, names, or quotes, it must also include a source you can check. If no source is available, the model should remove the detail or mark it as unknown.

Useful constraints you can bake into prompts:

• Numbers and dates: allow them only when the prompt provides a source excerpt or a verified note. Otherwise, write "no confirmed figure available" and continue.
• Rankings and "best" claims: ban "#1," "leading," and "most popular" unless you supply the ranking method and source.
• Exact quotes: require copy-pasted text from the source. No quote-shaped paraphrases.
• Names and studies: ban invented people, institutions, surveys, and "research shows" lines unless you provide study details.
• Product features: describe only what is explicitly in your notes. If you're unsure, label it unknown.

Uncertainty labels keep drafts usable without pretending. Ask the model to tag risky statements as confirmed, likely, or unknown. Only allow "likely" when it also states why (for example, "based on the provided changelog excerpt").

A repeatable workflow for low-hallucination drafts

The biggest win is consistency. Use the same small workflow every time so verification isn't an afterthought.

A fast loop that still catches the risky stuff:

1. Outline with claim types. Tag each point as context/definition, opinion/interpretation, or verifiable fact.
2. Draft low-risk sections first. Write definitions and background before adding stats, names, quotes, dates, or "top tools" lists.
3. Extract a claim list. Pull every factual claim into a short table: the claim, why it matters, and what kind of source would support it.
4. Verify, edit, then rewrite the affected parts. Add confirmed details, remove weak claims, and rewrite only what changed.
5. Final pass for certainty and missing citations. If a sentence makes a reader think "Really?" it needs a citation or softer wording.

Example: if a draft says "Google confirmed X in 2023," the claim list forces follow-up: which announcement, what date, and where is it recorded? If you can't answer quickly, soften or remove it.

Flag unverifiable statements without slowing down

Editing gets slow when every sentence turns into a courtroom. A faster approach is to make the draft self-identify what needs checking, then verify only the high-risk lines.

Instead of telling the model to "be accurate," tell it to label uncertainty. If it didn't see a source, it must not sound sure.

Use placeholders that stand out during editing:

• [VERIFY] for facts you'll check (numbers, dates, "first/only/biggest" claims)
• [SOURCE NEEDED] for statements that must be backed by a specific reference
• [UNSURE] for items that might be true but weren't in provided materials
• [QUOTE?] for any quotation or attributed phrasing

Then run a second pass that rewrites flagged lines in a safer way.

Example: "Acme launched in 2018 and now serves 10,000 customers." If you have no source, rewrite to: "Acme launched in [VERIFY: launch year] and serves [VERIFY: customer count]." A safe alternative is: "Acme has grown steadily since launch, serving customers across multiple markets." Less flashy, but not wrong.

When you review flags, decide quickly:

• Delete if the detail is nice-to-have.
• Soften if the point matters but the exact fact is unknown.
• Research if the claim is central.
• Replace with something observable or clearly sourced.

Lightweight artifacts that make quality control easier

You can reduce hallucinations without building a heavy process. The trick is to keep a few small artifacts that travel with the draft and make checking fast.

A simple claim table

Keep a small table next to the draft (even as a note). Every time the model makes a factual statement, capture it once and decide what happens next:

• Claim (one sentence)
• Risk level (low/medium/high)
• Source (title/author/date, or "none yet")
• Status (verified, needs source, remove, rephrase)
• Notes (what to check, what wording to soften)

If you can't source "Company X launched in 2017," treat it as high risk and rewrite the paragraph so it still works if the date changes.

A citation log you can reuse

Separate from the draft, keep a running citation log of sources you've already vetted. Over time, this becomes your go-to shelf for reliable references and reduces the temptation to accept whatever the model invents.

Keep it simple: source name, what it's good for, and any limits (for example, "good for definitions, not market stats").

A red-flag list and a short review script

Red flags deserve a second look: superlatives ("the best"), precise stats, named people/companies, quotes, dates, and "first/only/most" claims.

A short review script is enough:

• Scan for red flags and highlight them.
• Check each highlight against the claim table.
• Require a source for anything high risk.
• If no source exists, soften, generalize, or cut.
• Confirm the final draft matches the scope and audience.

Mistakes that let hallucinations slip in

Hallucinations usually appear when the model is allowed to fill in blanks and nobody notices. Most of the fix is removing those blanks.

Common traps:

• Prompts like "write an expert article" with no references or source material.
• Requests for "industry statistics" with no report name, dataset, or filing.
• Rewrites that chase a punchier tone and silently change years, names, or certainty.
• Prompts that mix timeframes and topics (past trends, current pricing, predictions) without clear boundaries.
• No rule for what to do when a source is missing.

If you do need stats, require a specific source type (named report, public dataset, annual filing) and force "unknown" when it's not available.

Quick checklist before publishing

Before you publish, treat the draft like a set of claims, not a set of sentences.

Read once for meaning. Then scan again for high-risk details: numbers, dates, rankings, quotes, and anything too specific.

• Numbers, dates, and rankings: Add a source you can point to, or rewrite as a general statement (or remove it).
• Quotes and attributions: If you can't verify it, delete it or turn it into an unattributed paraphrase.
• Names and spellings: Check people, companies, and product names. One wrong letter can create a fake entity.
• Scope check: Make sure claims match the stated region and timeframe.
• Uncertainty handling: Anything you can't verify quickly gets rewritten, labeled ("varies by"), or removed.

A practical trick: tag questionable lines with [VERIFY] as you edit. If any [VERIFY] tags remain at the end, the piece isn't ready.

Example: turning a risky draft into a source-checked article

Imagine you're drafting a post on a personal finance question: "Should I pay off my mortgage early or invest?" This is a common place where AI produces confident claims about returns, inflation, tax rules, and historical averages.

Treat it as a sources-first task, not a creativity task. Start with a short but strict prompt:

Task: Write a 900-word blog post for US readers deciding between extra mortgage payments vs investing.
Scope: General education only. No personal advice.
Allowed claims: Only claims supported by the sources I provide.
High-risk claims (must be cited): rates of return, inflation, tax rules, dates, named products, quotes.
Citations: After any sentence with a high-risk claim, add [S1], [S2], etc.
If a claim cannot be supported: mark it as [UNVERIFIED] and suggest what source is needed.
Sources:
S1: (paste a trusted explainer)
S2: (paste IRS page excerpt or official guidance excerpt)
S3: (paste a reputable calculator methodology page excerpt)

Then follow a tight flow:

• Outline first (headings only) and confirm each section has at least one source attached.
• Draft with inline citations, using [UNVERIFIED] when needed.
• Build a claim table: copy every high-risk sentence into a two-column table (Claim, Source tag).
• Verify: check each claim against the source excerpts.
• Rewrite: remove, narrow, or soften anything that doesn't verify.

A typical risky line is: "The stock market returns 10% per year on average." If your sources don't state that, rewrite it to avoid a fake precision: "Long-term returns can vary by time period, and no yearly return is guaranteed. [UNVERIFIED: needs a specific source and timeframe]" Or replace it with a narrower, sourced statement (for example, "Some long-term studies report historical averages over specific periods. [S1]").

Next steps: build a repeatable pipeline

The fastest way to reduce hallucinations is to stop treating each post like a one-off. Save your best prompts as templates with the same guardrails: allowed sources, forbidden claim types, and how to label uncertainty.

If you publish at scale, it helps when your tooling supports the same structure across drafts, rewrites, and updates. GENERATED (generated.app) is one example of a content platform that can generate content through an API, support polishing and translations, and help keep a consistent workflow around drafts and revisions.