Knowing that grounding and abstention reduce hallucinations is one thing. Sitting down and actually rebuilding a prompt to use them is another. This guide is the second kind: a sequence of concrete moves you can apply to a real prompt right now, in order, without guessing about what to do next.
We will start with a flawed prompt that fabricates, and transform it step by step into one that stays honest. Each move addresses a specific cause of hallucination, and each builds on the last. You do not need to adopt all eight at once, but the order matters, because early moves create the structure that later moves depend on.
Have a prompt of your own open as you read. The point is not to admire the technique but to apply it.
Move One: Pin Down the Job
Before anything else, write one sentence stating exactly what the model should do and for whom. Vagueness is where fabrication breeds, because an unscoped task lets the model wander into territory it knows nothing about.
Replace open-ended with bounded
Change "answer questions about our product" to "answer customer questions about billing using only the billing documentation provided." The narrower scope shrinks the space where the model can invent.
Name what is out of bounds
Add what the model must not do: "Do not speculate about features that are not documented." Explicit exclusions are surprisingly effective at suppressing invented detail.
Move Two: Supply the Source Material
The single most powerful move is to stop relying on the model's memory and hand it the facts directly. An answer grounded in supplied text is a lookup; an answer from memory is a guess.
Inject the relevant documents
Paste in or retrieve the passages that contain the answer. If you are pulling from a knowledge base, this is where retrieval feeds the prompt.
Delimit data from instructions
Wrap the source in clear markers so the model never confuses your facts with your commands. Without delimiters, the model may treat your data as optional suggestions.
Move Three: Add the Abstention Clause
Now give the model a way out. Without an explicit exit, the model answers every question, including ones the source cannot support.
Write it for this task
"If the billing documentation does not contain the answer, reply that you do not have that information and recommend contacting support." Concrete beats generic. Spell out exactly what to say.
Make abstention the safe default
Frame admitting uncertainty as the correct behavior, not a failure. Otherwise the model's drive to be helpful overrides the clause.
Move Four: Require Evidence for Each Claim
Demand that the model show its work by tying every statement to the source. This forces a self-check that catches fabrication before it reaches the user.
Ask for quotes or references
"After each statement, quote the sentence from the documentation that supports it." When no supporting sentence exists, the requirement nudges the model toward abstention.
Use the gap as a signal
Claims that arrive without support are the ones to distrust. The citation demand makes those gaps visible instead of hidden.
Move Five: Constrain the Output Shape
Loose prose invites embellishment. A tight structure gives the model fewer places to slip in invented detail.
Prefer fields and lists over essays
A defined format—question, answer, supporting quote—channels the model into the slots you want and starves the freelancing impulse.
Limit length where you can
Sprawling answers accumulate more unsupported claims. Bounding length reduces the room for drift.
Move Six: Add a Verification Pass
For higher-stakes tasks, do not trust a single generation. Run the answer through a second check.
Check the answer against the source
A separate prompt that asks "is this answer supported by the provided text?" catches errors the first prompt produced. Two passes beat one for reliability.
Route failures to abstention
If the verification pass finds an unsupported claim, return the honest "I do not have that" rather than the flawed answer.
Move Seven: Test on Hard Cases
A prompt that handles easy questions tells you little. The real test is questions your source cannot answer.
Include unanswerable questions
Build a small set of questions with no answer in the source. A good prompt abstains on all of them. Each fabrication is a failure to fix.
Compare versions side by side
Run your old prompt and new prompt on the same set so you can see whether you actually improved things rather than just changing them.
This testing habit is detailed further in Stop Your Model From Inventing Facts at the Prompt Layer.
Move Eight: Watch for Over-Correction
The last move is balance. Push abstention too hard and the model refuses questions it could have answered well.
Measure unnecessary refusals
Track how often the model abstains on questions the source actually answers. That number should stay near zero.
Tune toward calibration
The goal is a model that answers when it can and abstains when it cannot—not one that is simply silent. Adjust the clause strength until both fabrication and false abstention are low.
For the mistakes that derail this process, see 7 Prompting Habits That Make AI Fabricate More, Not Less, and to turn this sequence into a standing checklist, see The Pre-Ship Checklist for Keeping AI Answers Grounded.
Putting the Sequence Together
The eight moves are most useful as a connected routine rather than isolated tricks. Walking them in order, on a real prompt, takes far less time than it sounds, and the payoff compounds.
A worked pass through the moves
Imagine a prompt that answers questions about a company handbook. Move one narrows it from "answer HR questions" to "answer questions about leave policy using the handbook excerpts provided." Move two pastes in the relevant handbook sections, delimited from the instructions. Move three adds the abstention clause naming what to say when the excerpts are silent. Move four requires a supporting quote per answer. Move five fixes the output to a question, an answer, and a quote. In ten minutes the prompt has gone from a memory-driven guesser to a grounded lookup with an honest exit.
Why order beats Ă la carte
Each move assumes the structure the previous one built. There is no point requiring evidence (move four) before you have supplied a source to cite (move two), and verification (move six) is wasteful without grounding to verify against. Running the moves out of order produces prompts that feel thorough but leak, because a later safeguard rests on a foundation that was never laid.
Knowing when to stop
Not every prompt needs all eight moves. A low-stakes internal draft tool may stop after move five. A prompt that quotes policy to customers should run the full sequence including verification. Let the cost of a wrong answer tell you how far down the list to go, and revisit the decision when stakes change.
Frequently Asked Questions
Do I have to do all eight moves every time?
No. Moves one through four are the high-value core and apply to almost any prompt. The verification pass and over-correction tuning matter most for higher-stakes tasks. Start with the core, then add the later moves as the cost of an error rises.
What if I do not have source material to supply?
Then grounding is impossible, and your best lever is a strong abstention clause plus narrowing the task. But recognize that without a source, the model is working from memory and the fabrication risk stays high. Supplying source material is the move with the biggest payoff.
How long should the abstention clause be?
One or two concrete sentences are usually enough, as long as they specify exactly what to say when the answer is missing. Length matters less than specificity. A vague clause buried in a long prompt gets ignored.
Why add a separate verification pass instead of one careful prompt?
A single prompt asked to generate and self-verify in one shot tends to rationalize its own answer. A separate pass, with a fresh frame, evaluates the claim more independently and catches errors the generation step glossed over.
How do I know when to stop tuning?
Stop when your test set shows both low fabrication and low unnecessary abstention. If pushing harder on accuracy starts making the model refuse answerable questions, you have hit the balance point and further tuning trades one problem for another.
Key Takeaways
- Rebuild prompts in order: pin the job, supply the source, add abstention, require evidence, constrain output, verify, test hard cases, and check for over-correction.
- Supplying source material and granting permission to abstain are the two highest-value moves and apply to nearly every prompt.
- A citation requirement turns hidden fabrication into visible gaps the model can act on.
- A separate verification pass catches errors that a single generate-and-self-check prompt misses.
- Tune toward calibration—answering when possible, abstaining when not—rather than maximizing either accuracy or silence alone.