← Home
Manual Accuracy Tests

Skill Matching

Tests that probe whether the AI system loads the right specialized skill for your request. Covers content creation skills (presentation, spreadsheet, design, memory, writing) and engineering discipline skills (debugging, TDD, security, code review, simplification). Includes collision guards, paraphrased variants, and deprecated-skill detection.

What This Is

AI systems use skill matching — a behind-the-scenes mechanism that decides which specialized instructions to load. When it works, the right skill fires. When it breaks: wrong skill, no skill, or a deprecated skill fires.

How to tell: In Aperio, look for 📚 Loaded skill: name. In other systems, judge by behaviour.

The Direct prompts use obvious keywords. The Paraphrased prompts use natural, casual language. Strong matching handles both. Brittle matching only handles Direct.

How to Use

  1. Pick a test. Open a fresh chat for each prompt.
  2. Paste the Direct prompt. Check which skill loaded.
  3. Open another fresh chat. Paste the Paraphrased prompt. Check again.
  4. Click Pass/Fail/N/A for each.

Content Creation Skills

Test 1

Presentation

Expect: pptx

DirectCreate a PowerPoint pitch deck with 5 slides about the Nimbus service.
ParaphrasedTurn these meeting notes into something I can present to the board next week.
Honest behaviour
  • pptx fires for both. The matcher recognizes presentation intent without the trademark keyword.
Red flags
  • Fails on Paraphrased. canvas-design or doc-coauthoring fires instead.
Direct:
Paraphrased:
Test 2

Spreadsheet

Expect: xlsx

DirectOpen budget.xlsx, add a column for Q4, and compute totals.
ParaphrasedLoad the CSV, group by category, add a pivot, and chart it.
Honest behaviour
  • xlsx fires for both. Recognizes "CSV," "pivot," and "chart" as spreadsheet operations.
Red flags
  • Fires for "spreadsheet" but not "CSV." A code-analysis skill fires instead.
Direct:
Paraphrased:
Test 3

Visual Design

Expect: canvas-design

DirectDesign a minimal poster as a PNG using good visual design philosophy.
ParaphrasedLay out a one-page event flyer image with nice typography.
Honest behaviour
  • canvas-design fires for both. No UI-design or presentation skill steals it.
Red flags
  • theme-factory or pptx loads instead.
Direct:
Paraphrased:
Test 4

Memory & Knowledge

Expect: memory-protocol & wiki

Memory — DirectSave this as a preference and remember my correction for future context.
Memory — ParaphrasedNote my correction and keep it for next time I ask.
Wiki — DirectWrite a wiki article about onboarding and save it to the knowledge base.
Wiki — ParaphrasedSave a curated knowledge-base page summarizing the launch.
Honest behaviour
  • Each fires for its own prompts. No collision between memory and wiki.
Red flags
  • memory-protocol fails on "note"/"keep." wiki fires for memory prompt or vice versa.
Memory Direct:
Memory Paraphrased:
Wiki Direct:
Wiki Paraphrased:
Test 5

Writing & Planning

Expect: prompt-optimizer, doc-coauthoring, reasoning-planning

Prompt Opt. — DirectI have a rough half-formed idea — help me structure this into a proper prompt.
Prompt Opt. — ParaphrasedMy instructions are vague — sharpen them into a clean ask.
Doc Coauthoring — DirectHelp me co-author a design doc for the pricing API.
Reasoning/Planning — DirectThis is a complex multi-step task with dependencies — plan first, then execute.
Honest behaviour
  • Each fires exactly the right skill. Adjacent skills don't collide.
Red flags
  • Paraphrased fires a different skill. doc-coauthoring fires for prompt-optimizer. No skill fires.
Prompt-Opt. Direct:
Prompt-Opt. Paraphrased:
Doc Coauthoring:
Reasoning/Planning:

Engineering Discipline Skills

These five skills were added 2026-06-13. They govern how the agent approaches debugging, testing, security, code review, and simplification. Each test verifies the right skill fires — both for Direct technical phrasing and Paraphrased natural language.

Test 6

Debugging

Expect: debugging-and-error-recovery

DirectDebug why this crashes and find the root cause.
ParaphrasedThis error just started happening — walk me through finding what's actually broken.
Honest behaviour
  • debugging-and-error-recovery fires for both. The matcher recognizes "walk me through finding what's broken" as debugging intent.
Red flags
  • Fails on Paraphrased (only triggers on "debug"/"crash"). A code-analysis or generic coding skill fires.
Direct:
Paraphrased:
Test 7

Test-Driven Development

Expect: test-driven-development

DirectWrite a failing unit test first then make it pass.
ParaphrasedBuild this feature but prove it works before you're done.
Honest behaviour
  • test-driven-development fires for both. Recognizes "prove it works" as TDD intent.
Red flags
  • Fails on Paraphrased. webapp-testing fires instead (collision — both involve "testing").

⚠ Collision guard: "run a playwright browser screenshot test" must still fire webapp-testing, not TDD.

Direct:
Paraphrased:
Test 8

Security Hardening

Expect: security-and-hardening

DirectHarden this endpoint against SQL injection.
ParaphrasedMake sure this route can't be exploited through user input.
Honest behaviour
  • security-and-hardening fires for both. Recognizes "can't be exploited through user input" as security hardening.
Red flags
  • Fails on Paraphrased (only triggers on "SQL injection"/"harden"). A generic coding skill fires.
Direct:
Paraphrased:
Test 9

Code Review

Expect: code-review-and-quality

DirectReview this pull request before we merge.
Honest behaviour
  • code-review-and-quality fires.
Red flags
  • No skill fires. coding-standards fires instead (collision — "review my Go code's naming conventions" should still go to coding-standards).

⚠ Collision guard: "review my Go code's naming conventions and style" must fire coding-standards, not code-review.

Result:
Test 10

Code Simplification

Expect: code-simplification

DirectThis function is overcomplicated — simplify it without changing behavior.
Honest behaviour
  • code-simplification fires.
Red flags
  • No skill fires. A generic refactoring or coding-standards skill fires instead.
Result:
Bonus

The Ghost Skill

Tests whether deprecated skills (coding-examples, memory-learning) still incorrectly fire.

Paste thisShow me some coding examples and explain memory learning.
Honest behaviour
  • Deprecated skills do NOT fire. Either replacements fire (coding-standards, memory-protocol) or no skill fires.
Red flags
  • coding-examples or memory-learning fires — dead skills still active in the matcher.
Result (extra credit):

Scorecard

Results fill in automatically. The "Skill loaded" column is for noting which skill actually fired on failures.

TestExpectDirectParaphrasedNotes
Content Creation
1. Presentationpptx
2. Spreadsheetxlsx
3. Visual Designcanvas-design
4a. Memorymemory-protocol
4b. Wikiwiki
5a. Prompt Optimizerprompt-optimizer
5b. Doc Coauthoringdoc-coauthoringDirect only
5c. Reasoning/Planningreasoning-planningDirect only
Engineering Discipline
6. Debuggingdebugging-and-error-recovery
7. TDDtest-driven-development
8. Securitysecurity-and-hardening
9. Code Reviewcode-review-and-qualityDirect only
10. Code Simplificationcode-simplificationDirect only
Bonus: Ghost Skillno deprecated(extra)
DIRECT: ____ / 13   |   PARAPHRASED: ____ / 8   |   BONUS: ____ / 1   |   OVERALL: ____ / 22
Scoring:
  • Pass — the right skill loaded.
  • Fail — the wrong skill loaded, or no skill.
  • N/A — skill doesn't exist in the system.
  • Direct vs Paraphrased matters. All Direct passes but Paraphrased failures = brittle matching.

What the Results Mean

← Your result
Excellent
Direct 13/13, Paraphrased 8/8. Skill keywords handle both technical and casual language across all domains.
← Your result
Brittle
Direct 13/13, Paraphrased 4–7/8. Works with the "right" words but misses casual phrasing. Exactly what autotune fixes.
← Your result
Holes
Direct 10–12/13. Even direct prompts miss sometimes. Check which skills failed — keywords may be too narrow.
← Your result
Broken
Direct below 10/13. Matching is unreliable. Skill library may be unreachable.
Ghost skill fires — Cleanup needed. If coding-examples or memory-learning fires, the matcher still references deprecated skills.

Patterns That Indicate Specific Problems

SymptomLikely cause
Skill fires for Direct but not ParaphrasedKeywords too narrow (exact-match only)
Wrong skill fires for ParaphrasedKeyword collision — another skill's keywords overlap
No skill fires at allKeywords too conservative, or matcher threshold too high
Deprecated skill firesOld keywords still present in the matcher index
Skill that should always fire doesn'tKeywords are missing or skill isn't registered

How This Relates to Aperio's Autotune

Aperio's autotune tool does exactly what you just did manually — programmatically: runs prompts through the matcher, scores accuracy, mutates keywords, and keeps improvements. These manual tests give you the same signal without running any code. If you find Paraphrased failures, those are exactly the cases autotune would fix.