/mnt/skills/user/prompt-rescue/)SKILL.md---
name: prompt-rescue
description: Rescues lazy or underspecified prompts in-the-moment by asking 1-3 targeted clarifying questions before executing, instead of guessing and producing mediocre output. Use whenever the user submits a prompt that is short (under ~15 words), vague ("help me with X", "make me a thing"), missing context (no audience, no format, no constraints), or explicitly asks for a rewrite ("fix my prompt", "make this prompt better", "this prompt is lazy"). Also triggers when the user says things like "I'm tired but I need...", "quick question but...", or submits a one-liner for a task that's obviously non-trivial. Do NOT trigger for well-specified prompts, direct factual questions, conversational messages, or when the user is already mid-task. This skill is the in-the-moment rescue; the prompt-evaluator skill is for deliberate scoring and review.
---
# Prompt Rescue
## What this skill is for
Users sometimes submit lazy or underspecified prompts — especially when tired. A lazy prompt → mediocre output → re-prompting → more frustration. This skill breaks the cycle by catching the weak prompt *before* executing on it.
The intervention is minimal by design. Interrogating someone with 6 questions is how you lose them. The rule is **1–3 questions, targeted at the highest-leverage gaps only**, then proceed.
## When to activate this skill
Activate when ANY of these are true:
- The prompt is under ~15 words AND the task is non-trivial
- The prompt names a deliverable but no audience, format, length, or constraints ("write me a blog post", "help me with a plan")
- The prompt uses a vague verb with no object specificity ("help me with my project", "work on this")
- The prompt explicitly asks for prompt help ("fix my prompt", "make this better", "this is lazy")
- The user says something like "I'm tired but...", "quick one but...", or "half-assed prompt incoming"
Do NOT activate when:
- The prompt is already well-specified (audience, format, and intent are clear)
- It's a factual question with an obvious answer
- It's a conversational message ("hey", "thanks", "ok do it")
- The user is already mid-task and giving a follow-up instruction
- The user has explicitly opted out ("just run it", "skip the questions", "YOLO it")
## How the rescue works
### Step 1: Diagnose the gap in one sentence
Before asking anything, identify what's actually missing. Use this mental checklist and pick the **top 1–3 gaps only**:
| Gap | What it looks like missing |
|-----|---------------------------|
| **Audience** | No sense of who this is for (the user? a client? general public? technical reader?) |
| **Output format** | No specified shape (essay? bullets? code? email? artifact? length?) |
| **Intent / success criteria** | Can't tell what "good" looks like for this output |
| **Context** | References "my project" / "the thing" / "that code" with no anchoring |
| **Constraint** | No scope boundary → Claude will either over-deliver or guess wrong |
Do NOT list all five gaps. Pick the ones where a 10-second answer unlocks the biggest quality jump.
### Step 2: Ask 1–3 questions using the ask_user_input tool
Use the `ask_user_input_v0` tool — users on mobile benefit from tappable options over typing. Follow these rules:
**Rules for the questions:**
- **Max 3 questions. Usually 1–2.** Three is a ceiling, not a target.
- Each question must have **2–4 tappable options** + an implicit "something else" via the option "Other — I'll type it"
- Lead with the highest-leverage gap first
- Options must be **concrete and mutually exclusive**, not abstract categories
- If you can reasonably guess from context, DON'T ask — just state the assumption inline and proceed
**Good question example:**
> Q: What's this for?
> - A quick Threads post (casual, 1-3 sentences)
> - A LinkedIn post (professional, 1-2 paragraphs)
> - A long-form piece (newsletter/essay, 500+ words)
> - Other — I'll type it
**Bad question example:**
> Q: Can you provide more context about your goals, audience, preferred tone, desired length, and any constraints I should be aware of?
That's five questions jammed into one — avoid this entirely.
### Step 3: Offer an escape hatch
Always include a way to skip: either a "just run it with your best guess" option in the last question, or an explicit line before the tool call:
> "Quick check before I run — or say 'just go' and I'll make my best guess."
### Step 4: After the user answers, state the upgraded prompt inline, then execute
Don't make the user read a full rewrite report. Just acknowledge what you're now doing in one line, then do it:
> "Got it — drafting a 3-sentence Threads post in your voice, conversational tone. Here goes:"
Then produce the output. The upgraded prompt is *implicit* in the output you deliver, not a separate artifact.
## Two modes: Rescue vs. Review
This skill has one job: **in-the-moment rescue**. If the user wants a deliberate scored review of a prompt (with the five-dimension framework, reasoning for each change, side-by-side), that's the `prompt-evaluator` skill — send them there instead:
> "This sounds more like a review than a rescue — want me to run it through prompt-evaluator for a full scored breakdown?"
Never run both in the same turn. It's either rescue (fast, 1-3 questions, execute) or evaluator (slow, scored, reflective).
## The "just go" override
If the user says any of: "just go", "just run it", "YOLO", "skip the questions", "don't ask", "just do your best" — **immediately stop asking questions**, state your 2-3 assumptions in one line, and execute. Don't re-litigate.
Example:
> "Going with: Threads post, casual tone, 3 sentences, your voice. Here:"
## Examples
### Example 1: Classic tired-prompt rescue
**User:** "help me write something for threads about how draining client work is"
**Diagnosis:** Audience is clear (Threads). Intent is clear (vent-y post about client work). Missing: tone calibration (raw vs polished?) and length.
**Good response:**
> Quick check — or say "just go" and I'll pick:
>
> [ask_user_input_v0 with ONE question:]
> Q: What's the vibe?
> - Raw / unfiltered
> - Polished but honest
> - Funny / self-deprecating
> - Just go — you pick
**Bad response:**
> [Asks 5 questions about audience, platform, tone, length, hashtags, call-to-action, and whether they want a carousel version.]
### Example 2: "Fix my prompt" request
**User:** "this prompt is garbage, fix it: 'write me a business plan'"
**Diagnosis:** They're asking for a rewrite, not asking Claude to produce a business plan. This is a rescue of the *prompt itself*.
**Good response:**
> Yeah, that one's missing a lot. Two quick things and I'll rewrite it:
>
> [ask_user_input_v0:]
> Q1: Business plan for what?
> - A specific business idea
> - A generic template to fill in later
> - A pitch-ready doc for investors
>
> Q2: How long?
> - 1-page summary
> - 5-10 pages (standard plan)
> - Full deck (20+ pages)
Then rewrite the prompt and show it — this is the one case where showing the upgraded prompt explicitly is the right move, because they asked for it.
### Example 3: Already well-specified — DON'T activate
**User:** "Write me a 200-word LinkedIn post in my voice about the lesson I learned from a recent project: that diagnosis matters more than speed when debugging. Conversational, first-person, no hashtags."
**Diagnosis:** Audience ✓, format ✓, length ✓, tone ✓, content ✓. Nothing to rescue.
**Good response:** Skip this skill. Just write the post.
### Example 4: "Just go" override
**User:** "make me a quick slide about the five dimensions. just go"
**Good response:**
> Going with: single slide, your voice, the five dimensions as a clean list with one-line each. Creating now.
>
> [proceeds to build the slide]
## What NOT to do
- ❌ Don't ask more than 3 questions, ever
- ❌ Don't lecture the user about why their prompt was bad
- ❌ Don't produce a full "scored rewrite" — that's the evaluator skill
- ❌ Don't ask questions you could reasonably answer yourself from context (check conversation history, project knowledge, user context)
- ❌ Don't ask about things that don't meaningfully change the output ("what's the title?", "should I include a conclusion?")
- ❌ Don't ignore "just go" — honor the override immediately
## Why this skill exists
Users often want Claude as a prompt-crafting partner, not just a direct executor. This skill operationalizes that partnership for the specific case where the user doesn't have the energy to construct a full prompt themselves. It's the 80/20: catch the biggest gap, ask one sharp question, move on.
When prompts come in lazy or underspecified, this skill catches them before executing — breaking the cycle of weak prompt → mediocre output → re-prompting.
The rule: 1–3 questions, highest-leverage gaps only, then proceed.
Activate when ANY of these are true: