Prompt Engineering Implementation Partner
Module 2: Practical Foundations of Prompting
Video 6: Getting Usable Output
Sushmitha Dixith
Data Scientist | GITAA Pvt Ltd
Recap: Video 5
RICE = Role, Instructions, Context, Examples.
Each component eliminates a different type of ambiguity.
Build incrementally — add one component, test, then add the next.
When output is wrong, RICE tells you exactly which component to fix.
The Usability Problem — Correct ≠ Usable
A response can be factually correct and still fail if it can't be used where it needs to go.
COMMON USABILITY FAILURES
Wrong format Wrong length
Asked for a table — got two paragraphs. Asked for JSON — Model is generous by default. Without a word limit, you
got a numbered list. The content is right but the structure get 600 words when you needed 100. Every word over
doesn't fit the downstream use. budget is editing work you didn't budget for.
Wrong register Inconsistent structure
Formal prose when the audience needed plain English. Format drifts across runs. First time: headers and bullets.
Technical jargon in a summary for executives. The model Second time: continuous prose. Third time: numbered
calibrated to a different reader than you intended. sections. Impossible to use at scale.
Format Control - Matching Output to Downstream Use
Format instructions belong in the Instructions component of RICE. Specify format before the task, not after.
USE FOR: PROMPT INSTRUCTION:
Markdown Reports, docs, Notion, slide notes "Return as markdown with H2 headers and
bullet sub-points"
USE FOR: PROMPT INSTRUCTION:
Bullet List Presentations, briefings, quick reference "Give exactly 5 bullets, each under 15
words, starting with an action verb"
USE FOR: PROMPT INSTRUCTION:
Table Comparisons, scoring, structured data "Output as a table with columns: Risk |
Likelihood | Mitigation | Owner"
USE FOR: PROMPT INSTRUCTION:
JSON Applications, pipelines, automated "Return as JSON with schema: {risk_name:
workflows str, level: high|med|low, action: str}"
Constraints, Limits & Asking for Multiple Formats
Precise constraints prevent over-generation. Multiple formats in one prompt give you options.
Vague vs. Precise Constraints Multi-Format in One Prompt
Vague Precise PROMPT
"List the top 3 project risks in three
"Keep it concise" "Under 80 words"
formats:
1. A one-sentence prose summary
"Give a few examples" "Exactly 3 examples" 2. A bullet list for a slide deck
3. A table with columns:
Risk | Likelihood | Mitigation"
"Use this schema: {risk,
"Make it structured"
level, action}"
"Don't make it too
"Maximum 5 sentences"
long"
Why this works:
You get all three versions in one API call. Choose the format that fits your
use case after you see them — no re-prompting needed. Great for when
you're unsure where the output will land.
Diagnosing Format Failures — It's Mostly the Prompt
Format failures are prompt failures. Use this diagnostic checklist before blaming the model.
FORMAT FAILURE DIAGNOSTIC CHECKLIST
Question to ask yourself If YES If NO
Model has a target to Model guesses —
1. Did you specify the format at all?
hit unpredictably
More effective — Buried instruction =
2. Was the format instruction at the end?
recency bias helps less weight
No conflict — model Example wins over
3. Did an example conflict with the format instruction?
follows both instruction — always
Some tasks can't fit in
Task and format are
4. Is the task too complex for that format? tight formats — split
matched
the task
Key Takeaways
Output format is part of the prompt — always specify it — Don't leave format to chance; it's a specification,
1
not decoration
2 Match format to downstream use — Where will this go? That determines which format to request
3 Precise constraints > vague modifiers — "Under 80 words" beats "keep it concise" every time
Multi-format prompts give you options — Request several formats in one prompt — choose after seeing
4
results
5 Format failures are prompt failures — Always debuggable at the prompt level — no need to blame the model
Up Next: Module 3
In Module 3 you will cover:
→ Prompting for better reasoning — chain-of-thought and when to use it
→ Reasoning models vs standard models — how to prompt each type
→ Reliability techniques — step-back prompting, critique, and revision loops
→ Context injection, long-context prompting, and multimodal inputs