<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[int21]]></title><description><![CDATA[int21]]></description><link>https://int21.hashnode.dev</link><image><url>https://cdn.hashnode.com/uploads/logos/6ae6fa397cc2fac5284d71bb/3c05987d-7323-4958-8e4c-fb351d94b91f.png</url><title>int21</title><link>https://int21.hashnode.dev</link></image><generator>RSS for Node</generator><lastBuildDate>Sat, 21 Nov 2026 15:41:15 GMT</lastBuildDate><atom:link href="https://int21.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[Stop Writing One Giant Prompt: The Gather → Draft → Verify Pattern for Reliable LLM Output]]></title><description><![CDATA[Most people use an LLM the same way: paste everything into one prompt, hit enter, and hope. Sometimes the result is great. More often it is generic…]]></description><link>https://int21.hashnode.dev/stop-writing-one-giant-prompt-the-gather-draft-verify-pattern-for-reliable-llm-output</link><guid isPermaLink="true">https://int21.hashnode.dev/stop-writing-one-giant-prompt-the-gather-draft-verify-pattern-for-reliable-llm-output</guid><category><![CDATA[AI]]></category><category><![CDATA[#PromptEngineering]]></category><category><![CDATA[Prompt Engineering]]></category><category><![CDATA[llm]]></category><category><![CDATA[Productivity]]></category><dc:creator><![CDATA[int21]]></dc:creator><pubDate>Sun, 01 Nov 2026 08:46:32 GMT</pubDate><content:encoded><![CDATA[<p>Most people use an LLM the same way: paste everything into one prompt, hit enter, and hope. Sometimes the result is great. More often it is generic, slightly wrong, or confidently invents a detail you never gave it.</p>
<p>The model is not the main problem. The prompt is doing three different jobs at once:</p>
<ol>
<li><strong>Understanding the input</strong> (what are the facts?)</li>
<li><strong>Producing the output</strong> (write the thing)</li>
<li><strong>Checking the output</strong> (is it correct, complete, compliant?)</li>
</ol>
<p>When one prompt does all three, the model optimizes for sounding finished. It skips the boring checks and fills gaps with plausible filler. Split the jobs into separate steps and quality goes up noticeably, and it becomes predictable.</p>
<p>I call this <strong>Gather → Draft → Verify</strong>. Below is the pattern, why each phase exists, and a full worked example with copy-paste prompts.</p>
<h2>Phase 1: Gather (facts only, no writing)</h2>
<p>The first prompt never produces the final output. Its only job is to turn messy input into a clean, structured list of facts, and to tell you what is missing.</p>
<pre><code class="language-text">You are preparing facts for a property listing. Do NOT write the listing yet.

From the notes below, extract:
- Property type, beds, baths, square footage, lot size, year built
- Upgrades and their year (kitchen, roof, HVAC, windows...)
- Standout features a buyer would notice in the first 30 seconds
- Location facts that are objective (distance to transit, parks, shopping)

Then list:
- MISSING: facts a buyer will ask about that are not in the notes
- UNCLEAR: anything ambiguous or contradictory

Output as a bulleted list. Do not add anything that is not in the notes.

Notes:
"""
3bd 2ba ranch, ~1,650 sqft, built 1987. Kitchen redone 2022, quartz, gas range.
New roof 2021. Big fenced backyard w/ deck. 2-car garage. 0.3 mi to Lincoln Park,
10 min drive to downtown. Original bathrooms. HOA? not sure.
"""
</code></pre>
<p>Why this works:</p>
<ul>
<li><strong>"Do NOT write the listing yet"</strong> stops the model from rushing ahead.</li>
<li><strong>"Do not add anything that is not in the notes"</strong> is your first defense against invented details.</li>
<li><strong>MISSING / UNCLEAR</strong> turns the model into a checklist. Here it would flag the HOA question and probably ask about heating, flooring and school district. You fix those <em>before</em> drafting, not after a buyer asks.</li>
</ul>
<p>You review this list in about 30 seconds. That review is the cheapest quality check in the whole process.</p>
<h2>Phase 2: Draft (write from the verified facts only)</h2>
<p>Now you write, but only from the fact sheet you just approved, and with explicit constraints for the target format.</p>
<pre><code class="language-text">Write a property listing using ONLY the facts below. If a fact is not listed, do not mention it.

Format:
- Headline: max 12 words, lead with the strongest concrete feature
- Opening: 2 sentences, what makes this home different
- Body: 3 short paragraphs (inside, outside, location)
- Close: 1 sentence with a clear next step
- Total: 150-200 words
- Tone: warm and specific, no clichés ("stunning", "must-see", "dream home", "paradise")

Facts:
[paste the approved list from Phase 1]
</code></pre>
<p>Two details matter here:</p>
<ul>
<li><strong>Ban the clichés explicitly.</strong> Models fall back on "stunning" and "must-see" because those words are everywhere in their training data. Naming them removes them.</li>
<li><strong>Give a word budget per section, not just a total.</strong> A total alone produces one long paragraph about the kitchen and one line about everything else.</li>
</ul>
<p>If you need several formats (MLS, a portal like Zillow, a social post), do <em>not</em> ask for all of them in one go. Run Phase 2 once per format, from the same fact sheet. Each format gets the model's full attention, and they all stay consistent because they share one source of truth.</p>
<h2>Phase 3: Verify (a separate pass that only looks for problems)</h2>
<p>This is the phase everyone skips, and it is where most of the value is. Start a <strong>new prompt</strong> (or a new chat) so the model is not defending text it just wrote.</p>
<pre><code class="language-text">You are a strict reviewer. Do not rewrite the listing. Only report problems.

Check the listing against the fact sheet:
1. FACT CHECK: list every claim that is not supported by the fact sheet.
2. MISSING: list important facts from the sheet that the listing left out.
3. COMPLIANCE: flag wording that describes who should live there instead of
   what the property is (e.g. "perfect for young professionals", "ideal for
   families", "great for retirees"), or that references protected
   characteristics. Suggest neutral, feature-based wording for each.
4. CLICHÉS: list any filler words that slipped in.

Fact sheet:
[paste]

Listing:
[paste]
</code></pre>
<p>A separate review pass catches things the drafting pass will not:</p>
<ul>
<li><strong>Invented details.</strong> "Hardwood floors throughout" appears even though the notes never mention flooring.</li>
<li><strong>Steering language.</strong> In US real estate, the Fair Housing Act protects seven classes: race, color, religion, national origin, sex, familial status and disability. Phrases that describe <em>the ideal buyer</em> ("perfect for a young couple") instead of <em>the property</em> ("two bedrooms on the main floor") are a well-known source of complaints. Describe features, not people. (This is general information, not legal advice. Check your local MLS rules.)</li>
<li><strong>Gaps.</strong> Your best feature never made it into the text.</li>
</ul>
<p>Then you fix the issues yourself, or paste the report back with "apply these fixes, change nothing else."</p>
<h2>Why the pattern works (beyond real estate)</h2>
<p>The same three phases apply to almost any document you generate repeatedly:</p>
<table>
<thead>
<tr>
<th>Use case</th>
<th>Gather</th>
<th>Draft</th>
<th>Verify</th>
</tr>
</thead>
<tbody><tr>
<td>Product descriptions</td>
<td>Specs, materials, dimensions, what's missing</td>
<td>Title + bullets + description per marketplace</td>
<td>Unsupported claims, banned words, platform rules</td>
</tr>
<tr>
<td>Technical docs</td>
<td>API surface, params, edge cases</td>
<td>Reference + example</td>
<td>Wrong signatures, missing errors, outdated calls</td>
</tr>
<tr>
<td>Reports / summaries</td>
<td>Key numbers and sources</td>
<td>Narrative</td>
<td>Numbers that don't match sources</td>
</tr>
<tr>
<td>Emails to clients</td>
<td>Facts and the one ask</td>
<td>Email</td>
<td>Tone, missing ask, promises you can't keep</td>
</tr>
</tbody></table>
<p>The underlying principles:</p>
<ol>
<li><strong>Separate thinking from writing.</strong> A fact sheet is easy to verify. A finished paragraph is not.</li>
<li><strong>One source of truth.</strong> Every output format comes from the same approved facts, so they never contradict each other.</li>
<li><strong>Fresh eyes for review.</strong> A reviewer prompt in a new context is far more critical than "now check your work" in the same chat.</li>
<li><strong>Make the model say what it doesn't know.</strong> "MISSING / UNCLEAR" sections turn hallucinations into questions.</li>
</ol>
<h2>A quick checklist to try today</h2>
<ul>
<li><input disabled type="checkbox" /> Take one document type you generate every week.</li>
<li><input disabled type="checkbox" /> Write a Gather prompt that ends with MISSING and UNCLEAR.</li>
<li><input disabled type="checkbox" /> Write a Draft prompt with per-section budgets and a list of banned clichés.</li>
<li><input disabled type="checkbox" /> Write a Verify prompt that reports only, in a fresh chat.</li>
<li><input disabled type="checkbox" /> Save all three as templates and reuse them.</li>
</ul>
<p>The first run takes 10 minutes to set up. After that, each document takes a few minutes, and the output is consistent enough that you stop rewriting it from scratch.</p>
<h2>Free tool and ready-made templates</h2>
<p>If you write property listings, I built a free <strong>fair-housing wording checker</strong> that flags commonly problematic phrases in your text, right in the browser: <a href="https://invernomut0.github.io/real-estate-seo-articles/checker.html">Fair housing listing checker</a>.</p>
<p>I also packaged the full workflow for listings: 32 prompts across the three phases, with versions for MLS, Zillow-style portals and social media, covering different property types. That's the <a href="https://invernomuto2.gumroad.com/l/geffl">Real Estate Listing AI System</a>. I made it, and it's a paid product. Everything in this article works on its own without it.</p>
<p>Do you use a multi-step prompt workflow, or still one big prompt? I'm curious which phase people find most useful.</p>
<hr />
]]></content:encoded></item></channel></rss>