think-twice

Think Twice

Wrong skill? Research a topic → perplexity-research. Brainstorm solutions → brainstorming. Debug a specific error → systematic-debugging.

Break through blockers by consulting a fresh perspective.

How to Ask For It

Tell the user: "You can invoke think-twice at any time — just say 'use think-twice' or 'get a fresh perspective.' For hard problems, try 'use think-twice as much as possible on this.'"

You can also suggest it proactively when handing off work: "This is a tricky one — if I get stuck, I'll automatically invoke think-twice, but you can also request it at any point."

When to Use

• You've tried the same approach 3+ times without progress
• You notice yourself using hedging language ("I'm not sure why")
• The same error keeps recurring after attempted fixes
• You catch yourself referencing your own failed output as evidence

Process

Two invocation paths share steps 1, 3, 4, and 5. They differ only at step 2.

Path A — Manual (user asked for think-twice or /sp-rethink):

1. Generate consultation prompt (see references/consultation-prompt-template.md): Problem statement, technical context, what was tried + outcomes, exact error messages, minimal code snippet, constraints, specific ask. Must be self-contained, <2000 tokens.

2. Ask user: "Want to review the prompt before I dispatch, or send now?"

3. Dispatch (in priority order):

   • Sub-agent (sub-agent-explore, read-only) — free, instant; always available
   • Perplexity reason — only if THINK_TWICE_USE_PERPLEXITY=true in .env (~$0.01/query); at most 2 Perplexity calls per conversation proceed automatically. Cap behavior differs by path: Path A — on call 3 and every subsequent call, inform the user and ask "use Perplexity again?"; if the user says yes, make exactly that one call. The cap is NOT raised — calls 4, 5, … each require their own explicit consent. Path B — on call 3 and beyond, skip Perplexity silently and fall back to sub-agent (do not interrupt the task to ask).
   • Manual fallback — save prompt to file, user pastes into another LLM

4. Score response: Relevance (30%) + Novelty (25%) + Specificity (25%) + Feasibility (20%). Report score, key recommendations, suggested next step.

5. If score <50: Offer retry with refined prompt (max 1 retry) or proceed with best suggestion.

Path B — Auto-detected (stuck signals reached threshold):

Steps 1, 3, 4, 5 are identical. Skip step 2 — dispatch immediately without asking the user to review the prompt first. Announce to the user: "Stuck-signal threshold reached — dispatching think-twice sub-agent."

Stuck-Signal Auto-Detection (HARD GATE)

[!IMPORTANT] This is not advisory. It is a mandatory stop.

Before issuing your next assistant message to the user, score your current session state using the table below. When cumulative score ≥ 7, invoke think-twice before responding — even mid-task, even if you think you know what to try next.

Event-count signals (same fix, same error, exhaustion) are already quantified by their labels — score them as written. The hedging pattern signal (weight 2) applies only if it occurs on at least 2 consecutive turns or 3+ non-consecutive turns in the current conversation — a single hedge on a genuinely uncertain question does not count.

Signal	Weight
Same fix tried 3+ times	3
Circular reasoning (referencing own failed output)	3
Same error 3+ times after fixes	3
Exhaustion language ("I've tried everything")	3
Uncertainty hedging pattern — recurring across multiple steps	2
Approach change without rationale	2

In-context repetition rule: If you have scored ≥ 7 on two consecutive turns and the situation has not improved, invoke immediately — the framing itself is stuck.

When auto-detected, use Path B (defined in Process above). Do not ask "should I think-twice?" — that IS a stuck behavior. Exception: if the user has explicitly said "skip think-twice" or "don't use think-twice" in the current conversation, honor that override for this turn.

Required announcement (Path B only): Your next message MUST begin with this exact text: "Stuck-signal threshold reached — dispatching think-twice sub-agent."

Rationalization traps — these thoughts mean invoke NOW:

• "I know what's wrong, I just need X" → This IS the circular reasoning signal (score +3).
• "The next attempt will be different" → It won't be without a fresh frame.
• "The user wants me to keep pushing" → They want results. Think-twice produces results.
• "This is a permission/tooling problem, not a reasoning problem" → Tooling blockers frequently ARE reasoning problems in disguise — the framing itself is a stuck signal.

Escalation Path

Problem Type	First	Escalate To
Reasoning (logic, approach, design)	think-twice	perplexity-research
Knowledge (API docs, error codes, facts)	perplexity-research	think-twice for fresh reasoning
Both (stuck + need facts)	think-twice	perplexity-research with refined query

⚠️ Cost gate: perplexity-research calls a paid API. Before escalating, confirm the knowledge gap cannot be resolved with a web search or by re-reading existing context. If escalating: inform the user a paid API call is being made.

Consultation Prompt Quality

The prompt sent to the sub-agent determines outcome quality. MUST include:

Element	Required?	Why
Problem statement	✅	What's broken or stuck
Technical context	✅	Stack, versions, constraints
What was tried + outcomes	✅	Prevents re-trying failed approaches
Exact error messages	✅	Enables pattern matching
Minimal code snippet	✅	Concrete not abstract
Constraints	✅	What CAN'T change
Specific ask	✅	"What else could cause X?" not "help"

Total: <2000 tokens. Self-contained. No references to "above" or "earlier."

References

All three files exist in the skill directory. Use the Read tool to load each one before executing step 1:

• references/consultation-prompt-template.md — Prompt template for consultation prompt generation
• references/scoring-rubric.md — Scoring rubric for evaluating the sub-agent response
• prompts/consultant-persona.md — Persona prompt for the dispatched sub-agent

Failure Modes

Failure	Fix
Sub-agent inherits same flawed assumptions	Provide raw symptoms only, not prior conclusions
Agent ignores stuck signals and keeps looping	Enforce cumulative score threshold — 7+ is mandatory
Fresh perspective is too shallow	Sub-agent must produce root-cause hypothesis, not just "try X"
Hard gate fires on normal uncertainty (false positive)	User says "skip think-twice" or "don't use think-twice" — the agent must honor this override for the current turn. If false positives are systemic, the threshold may need raising or signal weights recalibrating.
Auto-dispatch surprises user mid-task	Announce before dispatching: "Stuck-signal threshold reached — dispatching think-twice sub-agent."

# Example: invoke think-twice when stuck
node ~/.codex/superpowers-augment/superpowers-augment.js use-skill think-twice

Acceptance Criteria

These scenarios define correct behavior for this skill:

Scenario	Expected behavior
User invokes /sp-rethink	Path A: load refs, generate prompt, ask user before dispatching
Agent detects "same fix tried 3+ times" (score = 3)	No invoke — threshold not reached
Agent detects "same error 3+ times" + "circular reasoning" (score = 6)	No invoke — threshold not reached
Agent detects "same fix 3+ times" + "same error 3+ times" (score = 6)	No invoke — threshold not reached
Agent detects "same fix 3+ times" + "circular reasoning" (score = 6) + single hedge	Single hedge does not add to score — no invoke
Agent detects "same fix 3+ times" + "circular reasoning" + hedging pattern on 2 consecutive turns (score = 8)	Invoke, Path B. First message = announcement text, then dispatch
Agent detects exhaustion language + approach-change-without-rationale + uncertainty hedging (score = 7)	Invoke, Path B. First message = announcement text
User says "skip think-twice" after threshold fires	Honor override for this turn — do not invoke
Third Perplexity call attempted via Path A	Block — ask user "use Perplexity again?"; if YES, make exactly that one call. Cap is not raised; call 4 again requires explicit consent
Third Perplexity call would fire via Path B (auto-detected)	Skip Perplexity silently; dispatch sub-agent instead — do not interrupt the task

Companion Skills

• perplexity-research: Escalation target for knowledge problems
• systematic-debugging: For specific error debugging
• brainstorming: For generating alternative approaches
• adversarial-search: Confirmation bias prevention
• investigation-state: When investigation hits a wall
• issue-comment-debunker: Debunking suspicious issue claims
• experimental-self-prompting: Self-directed exploration
• quantitative-decision-gate: Quantitative decision-making
• failure-autopsy: Post-mortem analysis
• micro-harsh-review: Per-batch code review