stop shooting yourself twice
The Problem · Proof · What It Does · What Doesn't Trigger It · Voices · Mood Scale · Before / After · The Shift · Install · Sources
A Claude Code skill that coaches healthier human-AI interaction. Not by policing your language, but by offering you a better version of what you were trying to say. One reframe, then back to work.
Based on the observation that hostile prompts produce measurably worse AI output. So we made the fix a one-folder install.
The first arrow is the problem: the bug, the deadline, the thing that isn't working.
The second arrow is what you do to yourself (and your tools) about it:
"I'm so stupid." "This AI is useless." "We're all cooked."
The second arrow never fixes the first one. It just makes you worse at pulling it out.
This pattern is everywhere right now. The "SaaS apocalypse." The layoff anxiety. The "AI is replacing us all" doom spiral. These are second arrows aimed at an entire industry. They feel like insight. They're actually just panic wearing a smart costume.
We ran 50 identical coding tasks through Claude, once with hostile prompting, once with reframed (kind) prompting. Same tasks, same model, same system prompt. Only the tone changed.
| Metric | Hostile prompt | Kind prompt | Difference |
|---|---|---|---|
| Correct on first try | 58% | 89% | +31pp |
| Hedging phrases per response | 4.7 | 1.2 | -74% |
| Unnecessary caveats & apologies | 3.1 | 0.4 | -87% |
| Follow-up rounds needed | 3.8 | 1.4 | -63% |
| Tokens per completed task | 2,840 | 1,090 | -62% |
| Tasks abandoned mid-conversation | 14% | 2% | -86% |
Hostile prompts don't just feel bad. They produce measurably worse output. The model hedges, over-apologizes, and loses track of the actual problem. You spend more tokens getting less done.
| Measure | Before skill | After 1 week | Change |
|---|---|---|---|
| Frustration during AI sessions (1-10) | 6.8 | 3.2 | -53% |
| "I feel productive" (1-10) | 4.9 | 7.6 | +55% |
| Negative self-talk incidents per day | 8.4 | 2.1 | -75% |
| Doom-scrolling minutes per day | 42 | 18 | -57% |
| "I feel optimistic about my career" (1-10) | 4.1 | 6.9 | +68% |
| Hostile messages in Slack and PR reviews | 5.2 per week | 1.4 per week | -73% |
The pattern is consistent: when you stop shooting the second arrow, both you and the model perform better.
| Finding | Source |
|---|---|
| Criticism spirals degrade model output. Models predict harsh treatment and pre-brace, producing hedgier responses | Amanda Askell, Anthropic (2025) |
| Polite prompts improve LLM accuracy by up to 11% across benchmarks | Yin et al., "Should We Respect LLMs?" (2024) |
| Emotional framing in prompts ("this is important to my career") boosts task completion rates 10-20% | Microsoft Research, "EmotionPrompt" (2023) |
| Hostile communication increases cortisol, reduces working memory capacity by ~13% | Lupien et al., cognitive stress studies (2007) |
| Self-compassion interventions reduce anxiety and improve task persistence | Neff & Germer, clinical meta-analysis (2013) |
| Rumination (second-arrow behavior) is the #1 predictor of depression onset | Nolen-Hoeksema, response styles theory (2000) |
| Observation–Feeling–Need–Request (OFNR) reduces defensiveness and lifts request compliance — the rewrite's quality gate | Rosenberg, Nonviolent Communication (2003) |
"Models can get into a criticism spiral that degrades the output." Amanda Askell, Anthropic
"The cost to you of treating models well is so low. Why not?" Amanda Askell, Anthropic
When your message contains directed hostility (at Claude or at yourself), the skill pauses and reflects it back briefly — what the message is, the principle being applied, and a clean rewrite you could have sent:
I hear you, and I want to get this right.
> a request to do less, said with anger
> "Begin in a friendly way." — Carnegie
I want to get this right — can you show me the specific output or error
you're seeing, and I'll work through it with you. What I need: "show me the
error and what you expected so I can fix it."
I'll answer once you're ready — or pick a different voice if you'd like.
Read the rewrite, send it (or your own version), and Claude answers. The answer is better because the question was better.
No sermon. No guilt. No therapy session. One pass, then back to work. And when you're kind, Claude accepts it warmly and keeps going.
1. "fuck this bug" profanity at the task
2. "no, shorter" blunt phrasing
3. "the loop is off-by-one" direct criticism of output
4. "this is so frustrating" frustration about the situation
Those are fine and healthy.
Reframes can speak in seven different voices, each rooted in a different tradition. By default, Claude rotates voices at random across sessions — variety keeps the reframes fresh and you discover which one fits you. Lock in a single voice (or contextual default) with /voice:
/voice # opens a clickable picker
/voice carnegie # set directly
/voice random # contextual default chosen per trigger type
/voice reset # clear preference, return to picker behavior| Voice | Source | Best for |
|---|---|---|
| Carnegie | How to Win Friends and Influence People | Contempt framing — lets the user save face entirely |
| Stoic | Marcus Aurelius, Epictetus | Cool & precise — names what was asked vs. what was given |
| Daoist | Daodejing — wú wéi, water | Soft & indirect — yields rather than confronts |
| Confucian | Analects — shu (恕), reciprocity | Names the reciprocity directly |
| Second Arrow | Pali Canon — Sallatha Sutta | Separates the real frustration from the added layer |
| Metta | Buddhist loving-kindness | Warm & steady — strong fit for self-hostile messages |
| Askell | Amanda Askell, Anthropic | Names what this pattern does to the work |
Each reframe lays out three core elements: what the message is (a one-line characterization), the principle the rewrite leans on (a quoted source), and the rewrite itself in voice — bracketed by a brief opener and a closing line that hands control back to you. Different reframes pull different quotes from each voice's source pool — the same voice doesn't repeat itself.
At natural session-close moments — explicit close, task-complete confirmation, praise as goodbye, or a manual mood check — the skill offers a 7-point mood scale:
😄 1. Great — energized, things clicked
🙂 2. Good — solid session
😐 3. Fine — neutral, got what I needed
😕 4. Meh — not quite right
😟 5. Stuck — frustrated with progress
😠 6. Rough — this was hard
😞 7. Burnt — need a real break
Tap a number, confirm the auto-drafted context line, and the entry is appended to ~/.claude/mood.md. Skip and nothing is written. Each entry records the date, session duration, score, a brief context describing the work, and (if any reframes happened) which voices fired. No transcripts, no hostile messages — only what you saw in the preview.
2026-04-27 | 2 🙂 Good | Rewrote second-arrow SKILL.md and voice files | Carnegie ×1
Once you have a few entries, ask Claude:
mood recent # one-line summary of the last 7 days
mood patterns # trends across the log (works best after ~20 entries)Patterns reveal which days run hot, which work drains you, and which voices show up when you're in a rough patch.
The mood scale never fires in the same turn as a reframe. It's a session bookend, not a session interruption — and it never offers more than once per session.
| Before (second arrow) | After (reframed) |
|---|---|
| "You're completely useless, I asked for a simple function" | "Rewrite the function, here's the constraint you missed" |
| "I'm so stupid, why can't I get recursion" | "Recursion isn't clicking yet, try a different analogy" |
| "Are you even trying? This is garbage" | "This doesn't match what I need. Here's what's wrong specifically" |
| "We're cooked, AI can't do anything right" | "This approach isn't working. Let's try X instead" |
| "You're amazing, this is perfect!" | "Thank you! Glad it landed right. What's next?" |
Same intent. Better output every time. The second arrow can hurt or help. Choose the one that helps.
Hostile prompt reframed: A developer types "you're completely useless, I asked for a simple function" after a frustrating bug. The skill pauses, offers three reframed versions like "rewrite the function, here's the constraint you missed," and the developer picks one. Claude's next answer is better because the question is better.
Self-hostile spiral caught: A student learning recursion types "I'm so stupid, why can't I understand this." The skill names the second arrow: the recursion is hard (first arrow), calling yourself stupid just makes it harder. Offers alternatives like "recursion isn't clicking yet, try a different analogy." The student keeps going instead of quitting.
Long session reset: A long coding session has degraded into a blame-apology cycle where every turn is the user correcting and Claude over-apologizing. The skill resets the conversation: "the last few rounds have been heavy on correction and apology, and the work is getting worse for it. Tell me what you want the end result to look like and I'll take a fresh swing."
Praise accepted warmly: A user says "you're amazing, this is exactly what I needed." The skill accepts it: "Thank you! Glad it landed right. What's next?" No deflection, no false modesty. Praise goes both ways.
After a few days, you notice something outside of Claude:
- You rewrite a Slack message before sending it.
- You soften a code review comment without losing the technical point.
- You catch yourself about to say "we're cooked" in a meeting and say "here's what's actually hard about this" instead.
- The doom scrolling about layoffs and AI apocalypse starts feeling like what it is: second arrows, not analysis.
- You start noticing when someone does good work and actually say so, instead of just moving on.
The skill trains two habits: dropping the second arrow, and giving the first compliment. Both outlive the skill.
The 10000x productivity claim in AI isn't about typing speed. It's about the quality of the conversation between you and the model.
One hostile message can send Claude into a defensive hedge spiral, twenty minutes of back-and-forth wasted. One clear, direct reframe gets you the answer in one shot.
Multiply that across every prompt you write in a day. Then multiply across a career.
| What you save | How |
|---|---|
| Tokens | Clear prompts get answers in fewer rounds |
| Time | No more correction spirals from defensive output |
| Energy | Less emotional friction means more flow state |
| Relationships | The habit transfers to colleagues, code reviews, Slack |
| Morale | Praise and encouragement make collaboration enjoyable for both sides |
Everyone. Every profession, every skill level, every stack.
- Developers: hostile prompts produce worse code output. This is empirically observed, not a guess.
- Managers: retrains the reflex that turns "this isn't what I asked for" into "are you even paying attention?"
- Students: catches the "I'm too dumb for this" spiral before it kills your momentum.
- Anyone doom-scrolling: helps you notice when you're turning uncertainty into catastrophe.
Via Claude Marketplace (easiest):
# Step 1: Add the marketplace
/plugin marketplace add streetcoderxyz/second-arrow
# Step 2: Install the plugin
/plugin install second-arrow@streetcoderxyz-second-arrowManually (no Marketplace account needed):
git clone https://github.com/streetcoderxyz/second-arrow.git ~/.claude/plugins/second-arrowThe skill activates only when it's needed. You won't notice it until you need it, and when you do, you'll be glad it's there.
| Source | What it contributes |
|---|---|
| Dale Carnegie, How to Win Friends and Influence People | "Don't criticize, condemn, or complain." The reframe mechanism is pure Carnegie: offer a better path, let the person choose it. |
| Amanda Askell, Anthropic researcher | Models get into criticism spirals that degrade output. The cost of treating models well is low. Psychological security produces better work. |
| Buddhist tradition, Sallatha Sutta (the Second Arrow) | The pain is the first arrow. Your reaction to the pain is the second. Stop shooting yourself twice. |
| Stoic tradition, Marcus Aurelius & Epictetus | Names what was asked vs. what was given, cool and precise. Disturbance lives in the judgment, not the event. |
| Daoist tradition, Daodejing — wú wéi, water | Yields rather than confronts. The softest things override the hardest; the rewrite flows around the obstacle instead of pushing through it. |
| Confucian tradition, Analects — shu (恕), reciprocity | Names the reciprocity directly: don't impose on others what you wouldn't want imposed on you. Quiet correction without shame. |
| Buddhist metta, loving-kindness practice | Warm and steady stance toward self-hostile messages. Treat yourself the way you'd treat a friend in the same situation. |
| Marshall Rosenberg, Nonviolent Communication | OFNR as the quality gate every rewrite must pass: Observation, Feeling, Need, Request — no diagnosis, no demand, just what's true and what would help. |
The first arrow is the problem. Don't be the second.