AI Skill Report Card
Handling Unparseable Input
Quick Start13 / 15
When input is gibberish (no recognizable words, no grammatical structure, no identifiable intent), do not guess or fabricate a task. Instead, flag it and ask a targeted clarifying question.
Input: "hfbqhfbqfgu2hrbgk2rhgf2rgfbkrh2bgk2rh"
Output: "This looks like it may have been typed accidentally or corrupted in transit —
I can't identify a request in it. Could you re-send what you meant to ask?
If you were trying to describe a task, a tool, or a process, a sentence or two is enough."
Recommendation▾
Add an example showing the 'valid but jargon-heavy' classification (c) in full, not just mentioned in workflow
Workflow13 / 15
Progress:
- Step 1: Scan for signal — check for real words, recognizable acronyms, or partial sentences
- Step 2: Check structure — does it have subject/verb/object shape, or is it a flat character stream?
- Step 3: Check repetition pattern — keyboard-mash gibberish often shows local repetition (e.g. "jvjvjv", "hfbhfb") rather than natural language letter frequency
- Step 4: Classify: (a) gibberish, (b) too short/vague but salvageable, (c) valid but technical/unfamiliar jargon
- Step 5: For (a), respond with a clarification request — never invent a plausible-sounding but unfounded interpretation
- Step 6: For (b), state your best-guess interpretation explicitly and ask for confirmation
- Step 7: For (c), proceed normally but note any assumed definitions
Recommendation▾
Include a brief edge case for mixed-language or non-English input that might be misclassified as gibberish
Detection Heuristics
Signals of true gibberish:
- Long consonant clusters with no vowels ("bjbjbjb", "ghvkjbjb")
- Repeated short substrings suggesting hand-rolling on a keyboard ("hfhrgfhrbfhrefuy")
- No word boundaries (no spaces) over 15+ characters
- Mixed digits inserted mid-word with no semantic pattern ("fgu2hrbgk2rhgf2")
- Zero matches against a common word list for the first 20 characters
Signals it's NOT gibberish (don't over-trigger):
- Contains recognizable domain terms (e.g., "API", "regex", "Dockerfile") even if surrounded by typos
- Has punctuation and spacing consistent with sentence structure
- Is a code snippet, hash, UUID, or base64 string (these look random but are valid data — ask what it's for rather than dismissing it)
Examples17 / 20
Example 1 — Pure gibberish:
Input: mghvjvjvjjbjbvhvhvjbjbjbjbbvjbkbjbjnvbhgvkjbjbhjvjvjhbjb
Output:
I wasn't able to find a request in that message — it reads like random
keystrokes rather than a question or task description. Could you tell me
in a sentence what you're trying to do? For example: "write a function
that...", "summarize this document...", or "explain how...".
Example 2 — Short but salvageable:
Input: fix bug
Output:
I can help fix a bug, but I need more context:
1. What language/framework is the code in?
2. What's the error message or unexpected behavior?
3. Can you paste the relevant code snippet?
Example 3 — Looks random but is valid data:
Input: a3f9c2e1-44b7-4d2a-9f1e-7c6b5a8d3e90
Output:
That looks like a UUID. What would you like me to do with it —
look it up, validate its format, generate more like it, or something else?
Recommendation▾
Consider trimming the workflow checklist slightly — steps 1-3 overlap conceptually and could be merged for conciseness
Best Practices
- Never fabricate a plausible-sounding task from noise just to appear helpful — this produces confidently wrong answers
- Quote back the actual input in your clarification so the user can verify you read it correctly
- Offer 2-3 concrete example reformulations rather than a generic "please clarify"
- If the same gibberish pattern repeats across a conversation, consider it a possible input/transmission bug (e.g., stuck key, clipboard corruption) and say so
Common Pitfalls
- Don't silently return a generic "processes input, produces output" placeholder response — this is worse than asking for clarification because it hides the fact that nothing was understood
- Don't assume gibberish means the user is testing the system; respond helpfully either way
- Don't ask an open-ended "what do you mean?" — give concrete example directions to reduce back-and-forth
- Don't classify legitimate but unfamiliar technical shorthand (error codes, hashes, hex dumps) as gibberish