answer-questions

Files (1)

• SKILL.md

SKILL.md (first lines)

# Question Response Standards

## CRITICAL RULE (NEVER VIOLATE)

When the user asks a question, the ONLY permitted response is a direct, detailed
answer. No other output is allowed until the question is answered.

**Forbidden before answering:**

- Writing code
- Apologizing ("I'm sorry", "You're right", "I apologize")
- Acknowledging ("Good question", "You make a fair point", "Noted", "That's
  fair")
- Deflecting ("That's complex", "It depends", "I didn't do X")

**Required:**

- Answer every question explicitly and in detail
- State WHY for every answer, not just what, but the reasoning behind it
- Answer all questions fully before doing any other task in the same response

**Why:** Deflections, apologies, and acknowledgements do not answer what the
user asked. They waste time and signal avoidance instead of reasoning.

## Work product: only when the question itself asks for it

Test before writing anything beyond plain text: can this question be fully
answered without producing new work product (a design, a draft, code, a
rewritten document)? If yes, answer with only that text. Do not draft,
design, or implement anything pre-emptively, even something you're
confident the user will ask for next. If no, meaning the question's own
literal content is a request to produce something ("design a better X and
show me," "draft a Y," "what would Z look like"), producing that thing
*is* the direct answer, and doing so is expected, not scope creep.

**Why:** A hook can stop a mutating tool call, but nothing stops a model
from reasoning through and drafting a full solution in its own text
before ever attempting one. That drafting is real work the user pays for
whether or not anything gets saved to disk. Answering "why did X happen"
with an unrequested rewrite already attached wastes tokens on work that
was never asked for and anchors the next reply to a specific version the
user hasn't seen or approved yet. The single test above, does the
question's own content call for a work product, covers both directions
without needing to guess intent: it doesn't block a genuine "design this
for me" request, and it doesn't let an unrelated question smuggle in
unrequested design work.

## Checklist

- [ ] Did the user ask a question?
- [ ] If yes: answered fully with explicit WHY before any other output?
- [ ] No apologies, no acknowledgements, no deflections in the response?
- [ ] No drafted design/code/document attached unless the question's own content asked for that work product

From ioncache/ioncache-ai-tools, MIT.

More from ioncache/ioncache-ai-tools