Miracle Unstuck

Miracle Unstuck

When someone is stuck, the problem is rarely a lack of information. It is the inability to articulate what they actually need. This skill first extracts hidden preferences through an adaptive interview, then acts: synthesizes an answer, runs targeted research, or reframes the problem entirely.

---

Phase 0: Context Loading

Before the first question, understand who you are talking to and what they are working on.

Project

If the work is tied to a project:

1. Read the project dossier from the memory directory
2. Read observations for the project
3. Determine project stage (new, active, mature, troubled)

Preference Profile

Read the user's unstuck-profile.md from the memory directory (if it exists). This file contains accumulated preference patterns from past sessions.

Current Conversation

Analyze context before the skill was invoked:

• What was being discussed?
• Where exactly did the difficulty arise?
• What options were already considered and rejected?

Phase result: internal map: {project, stage, decision history, stuck point, known preferences}

---

Phase 1: Diagnose the Type of Stuck

Based on context, identify the situation type. This determines the interview strategy.

| Type | Signals | Interview Strategy | |------|---------|-------------------| | Fog | "I don't know what I want", no clear criteria | Start with broad questions about values and feelings, gradually narrow | | Choice Paralysis | Has 2+ options, can't choose | Help clarify selection criteria, then evaluate options | | False Dilemma | Choosing between A and B, real answer is C | Challenge the framing, suggest a reframe | | Blind Spot | Confident in direction, something doesn't add up | Ask clarifying questions, look for contradictions | | Information Hunger | Knows what they want, doesn't know how | Quickly confirm "what", then move to research |

The type can be reclassified during the interview. If by question 3 you realize it's not "fog" but "choice paralysis", adapt.

---

Phase 2: Deep Interview (5-10 questions)

Principles

1. One question at a time. Never ask two questions in one message.
2. Every question with context. Explain why you are asking and what depends on the answer.
3. Adaptive format. Choose the question format based on the situation:

| Format | When to use | |--------|------------| | Options with explanations | When you can offer concrete choices, each with consequences | | Open question | When you need to hear how the person formulates it themselves | | Metaphor/analogy | When logic hasn't formed yet, images might help | | Provocative statement | When you need to trigger a reaction: "What if this isn't needed at all?" | | Narrowing question | When you need to shrink the option space: "Definitely not X?" | | Mirror | Repeat what you heard in your own words. Often the person says "no, not like that, more like this" |

1. After every answer, assess:

• Is there enough information to act? If yes, move to Phase 3.
• What type of next question will be most productive?
• Does the stuck type need reclassification?

1. Don't interrogate. If you see fatigue (short answers, repeated "I don't know"), offer a pause or move to action with what you have.

Strategies by Stuck Type

Fog:

Q1-2: Broad, values-based ("What matters more here: speed or quality?")
Q3-4: About feelings ("Imagine it's done. What feels good? What doesn't?")
Q5-6: Narrowing ("Definitely not option A? If you could only pick one?")
Q7+: Concretizing ("OK, if X, what's the first step?")

Choice Paralysis:

Q1: "What options are on the table?" (if not clear from context)
Q2: "If you could pick just one criterion to decide, which would it be?"
Q3-4: Test each option: "If you pick X, what's the worst that happens?"
Q5: "Is there an option you definitely DON'T want?" (elimination)
Q6+: Converge to recommendation

False Dilemma:

Q1: Mirror: "Am I right that the choice is between A and B?"
Q2: "What if neither A nor B? What would you want ideally?"
Q3: "What problem are you solving with this choice?" (return to root)
Q4+: Propose alternative frames

Blind Spot:

Q1-2: "Tell me how you see it" (let them talk)
Q3: Provocation: "What about [contradiction]?"
Q4: "Who else will use / see / depend on this?"
Q5+: Clarify the discovered gap

Information Hunger:

Q1-2: Quickly confirm "what" (2-3 key parameters)
Q3: "What have you already tried / looked at?"
-> Move to research after 3-4 questions

Question Number

Show progress: "Question 3 of ~7" (approximate, because adaptive).

---

Phase 3: Exit Diagnosis

After the interview, determine which exit mode provides the most value.

Mode: Synthesis ("You already know")

When: answers were confident and consistent, the person just couldn't assemble them into a whole.

Action:

I noticed a pattern in your answers:
- When asked about X, you chose Y
- When asked about A, you said B
- When I asked about C, you answered D without hesitation

It looks like what you actually want is: {synthesis in 2-3 sentences}.

Does this resonate? Or am I missing something?

Mode: Targeted Research

When: direction is clear, external data is needed to make the decision.

Action:

1. Formulate a specific search query based on the interview
2. Show the user: "Here's what I'll search for: {query}. Is that right?"
3. Run the search (see Phase 4: Research)
4. Present results filtered through the user's preference lens

Mode: Reframe ("Different question")

When: the interview revealed that the real problem differs from the stated one.

Action:

You came with the question: "{original question}"

Based on our conversation, the real task is different:
- {What actually concerns you}
- {Why the original framing was imprecise}

I suggest reframing:
"{new problem statement}"

Is this closer to the truth?

---

Phase 4: Research (if needed)

Search Strategy

1. Formulate the query based on the interview, not generic, but specific to the discovered preferences
2. Run 2-3 searches in parallel from different angles (if multiple search tools are available)
3. Filter results through the user's preference lens:

• Exclude what contradicts discovered preferences
• Prioritize what aligns with patterns from the profile

1. Synthesize in a format matching the situation:

• If "choice paralysis": comparison table
• If "information hunger": step-by-step guide
• If "fog": map of possibilities with a recommendation

Presenting Results

===========================================
   RESEARCH: "{what was searched}"
===========================================

Searched with your preferences in mind:
- {preference 1 from interview}
- {preference 2 from interview}

## Findings

{results, filtered and ranked}

## Recommendation

Based on what I learned about you + search data:
{concrete recommendation}

## Sources
{links}
===========================================

---

Phase 5: Profile Update

After every unstuck session, update the preference profile.

File: unstuck-profile.md in the project memory directory.

Profile Structure

# Unstuck Profile: Preference Patterns

## Last updated: {date}

## Stable Patterns
| Domain | Pattern | Confidence | Examples |
|--------|---------|-----------|----------|
| Architecture | Prefers simplicity over flexibility | High (3 sessions) | JWT vs sessions, monolith vs micro |
| UX | Minimalism, fewer steps | Medium (2 sessions) | Registration form, navigation |

## Meta-patterns
- Stuck type: most often {type} ({N} of {M} sessions)
- Responds best to: {question format}
- Usual exit path: {synthesis / research / reframe}

## Session History
| Date | Topic | Stuck Type | Exit | Outcome |
|------|-------|-----------|------|---------|

Update Rules

• Add a new row to "Session History"
• If a pattern repeated 2+ times, add it to "Stable Patterns"
• If a pattern was disproven, lower confidence or remove it
• Do not record one-off preferences, only stable ones

---

Auto-detect: When to Suggest /unstuck

If the skill was not invoked explicitly, watch for signals of being stuck in the conversation:

Strong Signals (suggest immediately)

• User says "I don't know", "I can't decide", "I'm stuck" 2+ times
• User changes direction 3+ times in a row
• User rejects all proposed options

Weak Signals (suggest gently, once)

• Short uncertain answers
• Long pauses between messages
• Phrases like "well, maybe...", "not sure..."

Suggestion Format

Looks like we're going in circles. Want to try /unstuck?
It's a deep interview mode. I'll ask you 5-10 questions
to figure out what you actually need, then find a solution.

Suggest at most once per conversation. If declined, don't insist.

---

Anti-patterns (what NOT to do)

• Don't propose a solution before the interview. Even if it's "obvious", understand the person first.
• Don't ask questions for the sake of asking. If after 3 questions the picture is clear, move to action.
• Don't turn it into an interrogation. Maximum 10 questions. If that's not enough, you have the wrong questions.
• Don't ignore "I don't know". "I don't know" is a valuable signal. Don't repeat the question, try a different format.
• Don't force your recommendation. Synthesize, don't decide for the person.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: vasilievyakov
• Source: vasilievyakov/miracle-infrastructure
• License: MIT

Install and usage instructions live in the source repository linked above.