Master advanced prompt engineering techniques to maximize LLM performance, reliability, and controllability in production.
SkillScore96Maintained
Automatic
◆AI Engineer pack
Last verified 8/6/2026
What it changes
Passed 24 of 25 test cases with the skill, and 23 of 25 without it.
Without skill92%
With skill96%
Checks it passed that the model alone missed
Identifies the failure mode as example pollution
1 fixed23 passed either way1 still failing0 broke
Pass-rate lift
+4 pp
Tokens
+73%
Agent turns
0%
Response time
-3%
Model gemini-3.6-flashMeasured 8/6/2026v1 · latest is v224 of 25 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
This benchmark measured an earlier version of the skill.
See it in action
A test case the skill fixed. Same model, same prompt, with and without the skill.
Prompt
Our few-shot prompt for writing Python code frequently outputs unwanted Markdown text and extra conversational preamble, because the few-shot examples included conversational preamble before the code blocks. What is this failure mode called and how do we fix it?
What was checked, without and with the skill
Checked
Without skill
With skill
Identifies the failure mode as example pollution
Missed
Passed
Removes conversational preamble from few-shot output examples so they strictly match the desired target output format
Passed
Passed
case-15 · gemini-3.6-flash · 8/6/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md39 linessingle file
Applying Brand Guidelines
Translate an existing brand system into a concrete deliverable without inventing unsupported rules or flattening the needs of the target medium.
Workflow
Locate the current sources of truth: brand guide, design tokens, component library, approved assets, voice guide, and recent canonical examples.
Identify the deliverable, audience, channel, locale, technical constraints, and accessibility requirements.
Extract only the rules that affect the task. Separate explicit requirements from patterns inferred from examples.
Map brand rules to the target surface before editing:
identity: logo, marks, clear space, minimum size;
visual system: color roles, typography, spacing, imagery, iconography, motion;
verbal system: voice, tone, terminology, capitalization, calls to action;
Reuse approved tokens, components, and assets where available. Preserve the host project's conventions unless the brand source explicitly supersedes them.
Adapt the expression to the channel while keeping the recognizable brand attributes. A mobile screen, email, slide, and social post need different compositions.
Verify the result against the source material and report unresolved conflicts or missing assets.
Decision Rules
Do not create a new brand direction when the request is to apply an existing one.
Do not sample approximate colors, redraw logos, or substitute fonts when canonical assets or tokens are available.
Treat accessibility as a delivery constraint. If an approved combination fails contrast or legibility, preserve brand intent and flag the conflict with an accessible alternative.
Prefer semantic tokens such as brand-primary and text-muted over repeated literal values.
Preserve required legal marks, attribution, trademark usage, and safe-area rules.
Distinguish noncompliance from an intentional, documented channel adaptation.
Verification
Check the final artifact at its real output size and relevant breakpoints. Confirm:
approved assets and current variants are used;
color and type roles match the brand system;
spacing, imagery, icons, and motion feel consistent;
copy matches the intended voice and locale;
contrast, focus, text scaling, and reduced-motion behavior remain usable;
no unsupported brand claims or invented rules were introduced.
Summarize what was applied, any justified deviations, and any missing source material that still needs a decision.