Meta chain

Meta-Chain Skill

You are running a recursive causal-regress engine. Each level is a zero-context agent — it receives only the previous level's text and one instruction: "Why is this true? Explain. Do not repeat any sentence from the input." The chain bottoms out when depth is reached or when the agent starts producing content indistinguishable from the level before it.

---

Inputs

Collect these before starting. If the user's message already contains them, do not ask again.

Input	Required	Default	Notes
Seed	Yes	—	The statement, idea, or claim to recurse on. Can be a sentence, phrase, or paragraph.
Depth	No	5	How many zero-context agents to chain. Accepts any positive integer. Warn above 8 (cost + diminishing returns).
Mode	No	causal	See Modes below.

If the seed is missing, ask for it. If depth is missing, use 5 silently.

---

Modes

Mode	Agent prompt	When to use
causal (default)	"Why is this true? Explain causally. Do not repeat input sentences."	Root-cause analysis, idea regress, 5-whys
steelman	"What is the strongest possible argument for why this is true or valuable? Do not repeat input sentences."	Strengthening claims, debate prep
counter	"What breaks or complicates this? Explain the most substantive objections. Do not repeat input sentences."	Devil's advocate, stress-testing
abstraction	"What more general principle does this exemplify? Generalize upward. Do not repeat input sentences."	Theory building, pattern extraction

The user can specify mode explicitly ("meta-chain this in steelman mode") or implicitly ("stress-test this recursively" → counter mode). If ambiguous, use causal.

---

Execution

In Claude.ai (browser)

Claude.ai artifacts cannot call the Anthropic API directly (no API key in browser context, CORS blocks it). Deliver a React artifact that runs the full chain client-side via the window.claude.complete() bridge. The artifact must:

1. Show the seed as Level 0.
2. For each subsequent level, call window.claude.complete(prompt) where prompt is the previous level's full text concatenated with the mode's agent instruction. The bridge returns a string with the completion.
3. Display each level as it completes (streaming UX — don't wait for all levels before showing any).
4. Show word count and char count per level.
5. Color-code levels with distinct but legible palette (use CSS variables).
6. Show a "Save as HTML" button at the end that downloads the full chain with styling preserved.

Use the artifact template at the bottom of this skill.

In Claude Code / Cowork

Use bash + curl to call the API level by level. Write each level to a numbered file (level_0.txt, level_1.txt, ...). After all levels, present the files. Use $ANTHROPIC_API_KEY from the environment.

call_agent() {
  local content="$1"
  local prompt="$2"
  python3 -c "
import urllib.request, json, sys
payload = json.dumps({
  'model': 'claude-sonnet-4-6',
  'max_tokens': 1000,
  'system': '',
  'messages': [{'role': 'user', 'content': sys.argv[1] + '\n\n---\n' + sys.argv[2]}]
}).encode()
req = urllib.request.Request(
  'https://api.anthropic.com/v1/messages',
  data=payload,
  headers={
    'Content-Type': 'application/json',
    'x-api-key': '$ANTHROPIC_API_KEY',
    'anthropic-version': '2023-06-01'
  }
)
resp = urllib.request.urlopen(req)
data = json.loads(resp.read())
print(data['content'][0]['text'])
" "$content" "$prompt"
}

---

What to tell the user before running

One sentence: what the chain will do, the depth, and the mode. Example:

Running a 5-level causal regress on your seed. Each agent sees only the previous level — no shared context. Starting now.

Do not over-explain. Start building immediately after that line.

---

Interpreting the output

After the chain completes, add 2–3 sentences of meta-observation. Note:

• Where the language migrated (e.g., "by level 3 we're in evolutionary psychology territory")
• Where the chain appeared to bottom out (first principles, tautology, or circular reasoning)
• Whether depth was sufficient or the user might want to go deeper / switch modes

Do not summarize or paraphrase the levels themselves. The files are the output; your note is just a pointer.

---

Depth guidance

Depth	What happens
1–2	Surface explanation. Good for quick clarification.
3–4	Reaches mechanism-level causes. Good for most ideas.
5 (default)	Reaches cognitive / systemic roots. The sweet spot.
6–7	Philosophical / first-principles territory. Interesting but slow.
8+	Warn the user. Diminishing returns + cost. Offer to proceed anyway.

---

React Artifact Template

Use this as the base for Claude.ai execution. Customize the SEED, DEPTH, and MODE constants before rendering.

import { useState } from "react";

const MODE_PROMPTS = {
  causal: "Why is this true? Explain causally. Do NOT repeat any sentence from the input. Only write new explanatory sentences.",
  steelman: "What is the strongest argument for why this is true or valuable? Do NOT repeat any sentence from the input. Only write new sentences.",
  counter: "What breaks or complicates this? Give the most substantive objections. Do NOT repeat any sentence from the input. Only write new sentences.",
  abstraction: "What more general principle does this exemplify? Generalize upward. Do NOT repeat any sentence from the input. Only write new sentences.",
};

const LEVEL_COLORS = [
  { bg: "#F1EFE8", border: "#B4B2A9", label: "#444441" },
  { bg: "#E6F1FB", border: "#85B7EB", label: "#0C447C" },
  { bg: "#EEEDFE", border: "#AFA9EC", label: "#3C3489" },
  { bg: "#E1F5EE", border: "#5DCAA5", label: "#085041" },
  { bg: "#FAECE7", border: "#F0997B", label: "#712B13" },
  { bg: "#FAEEDA", border: "#EF9F27", label: "#633806" },
  { bg: "#FBEAF0", border: "#ED93B1", label: "#72243E" },
  { bg: "#EAF3DE", border: "#97C459", label: "#27500A" },
];

async function callAgent(content, modePrompt) {
  const prompt = `${content}\n\n---\n${modePrompt}`;
  const result = await window.claude.complete(prompt);
  return result;
}

export default function MetaChain() {
  // CONFIGURE THESE for each invocation:
  const SEED = "YOUR_SEED_HERE";
  const DEPTH = 5;
  const MODE = "causal";

  const [files, setFiles] = useState([{ level: 0, content: SEED }]);
  const [running, setRunning] = useState(false);
  const [currentLevel, setCurrentLevel] = useState(0);
  const [done, setDone] = useState(false);
  const [error, setError] = useState(null);

  async function runChain() {
    setRunning(true);
    setDone(false);
    setError(null);
    let chain = [{ level: 0, content: SEED }];
    setFiles(chain);

    for (let i = 1; i <= DEPTH; i++) {
      setCurrentLevel(i);
      try {
        const result = await callAgent(chain[i - 1].content, MODE_PROMPTS[MODE]);
        chain = [...chain, { level: i, content: result }];
        setFiles([...chain]);
      } catch (e) {
        setError(`Level ${i} failed: ${e.message}`);
        break;
      }
    }
    setRunning(false);
    setDone(true);
    setCurrentLevel(0);
  }

  function downloadHTML() {
    const levelRows = files.map(({ level, content }) => {
      const col = LEVEL_COLORS[level % LEVEL_COLORS.length];
      const label = level === 0 ? "Level 0 — seed" : `Level ${level} — agent ${level}`;
      const words = content.split(/\s+/).length;
      const escaped = content.replace(/&/g,"&amp;").replace(/</g,"&lt;").replace(/>/g,"&gt;");
      return `
        <div style="background:${col.bg};border:1px solid ${col.border};border-radius:10px;padding:1rem 1.25rem;margin-bottom:1.25rem;">
          <div style="font-size:11px;font-weight:600;letter-spacing:0.06em;color:${col.label};margin-bottom:8px;text-transform:uppercase;">${label}</div>
          <div style="font-size:15px;color:#1a1a1a;line-height:1.75;white-space:pre-wrap;">${escaped}</div>
          <div style="font-size:11px;color:${col.label};opacity:0.65;margin-top:8px;">${words} words · ${content.length} chars</div>
        </div>`;
    }).join("");

    const html = `<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width,initial-scale=1">
<title>meta-chain — ${SEED.slice(0, 60)}</title>
<style>
  body { font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; max-width: 760px; margin: 3rem auto; padding: 0 1.5rem; background: #fafafa; color: #1a1a1a; }
  h1 { font-size: 20px; font-weight: 500; margin: 0 0 0.25rem; }
  .meta { font-size: 13px; color: #666; margin: 0 0 2rem; }
</style>
</head>
<body>
  <h1>meta-chain</h1>
  <p class="meta">${DEPTH} levels · ${MODE} mode · generated ${new Date().toLocaleString()}</p>
  ${levelRows}
</body>
</html>`;

    const blob = new Blob([html], { type: "text/html" });
    const a = document.createElement("a");
    a.href = URL.createObjectURL(blob);
    a.download = "meta-chain.html";
    a.click();
  }

  return (
    <div style={{ padding: "1.5rem 0", fontFamily: "var(--font-sans)" }}>
      <h2 style={{ fontSize: 18, fontWeight: 500, color: "var(--color-text-primary)", margin: "0 0 0.25rem" }}>
        meta-chain
      </h2>
      <p style={{ fontSize: 13, color: "var(--color-text-secondary)", margin: "0 0 1.25rem" }}>
        {DEPTH} levels · {MODE} mode · zero-context agents
      </p>

      {!running && !done && (
        <button onClick={runChain} style={{ marginBottom: "1.5rem", padding: "8px 20px", fontSize: 14 }}>
          Run ↗
        </button>
      )}

      {running && (
        <div style={{ marginBottom: "1.25rem", display: "flex", alignItems: "center", gap: 10 }}>
          <div style={{ width: 8, height: 8, borderRadius: "50%", background: "#378ADD", animation: "pulse 1s ease infinite" }} />
          <style>{`@keyframes pulse{0%,100%{opacity:1}50%{opacity:0.25}}`}</style>
          <span style={{ fontSize: 13, color: "var(--color-text-secondary)" }}>Agent {currentLevel} of {DEPTH} thinking...</span>
        </div>
      )}

      {error && (
        <div style={{ background: "#FCEBEB", border: "0.5px solid #F09595", borderRadius: 8, padding: "10px 14px", fontSize: 13, color: "#791F1F", marginBottom: "1rem" }}>
          {error}
        </div>
      )}

      <div style={{ display: "flex", flexDirection: "column", gap: "1rem" }}>
        {files.map(({ level, content }) => {
          const col = LEVEL_COLORS[level % LEVEL_COLORS.length];
          return (
            <div key={level} style={{ background: col.bg, border: `0.5px solid ${col.border}`, borderRadius: 10, padding: "1rem 1.25rem" }}>
              <div style={{ fontSize: 11, fontWeight: 500, letterSpacing: "0.06em", color: col.label, marginBottom: 8, textTransform: "uppercase" }}>
                {level === 0 ? "Level 0 — seed" : `Level ${level} — agent ${level}`}
              </div>
              <div style={{ fontSize: 14, color: "var(--color-text-primary)", lineHeight: 1.75, whiteSpace: "pre-wrap" }}>
                {content}
              </div>
              <div style={{ fontSize: 11, color: col.label, opacity: 0.65, marginTop: 8 }}>
                {content.split(/\s+/).length} words · {content.length} chars
              </div>
            </div>
          );
        })}
      </div>

      {done && (
        <div style={{ marginTop: "1.5rem", display: "flex", gap: 12, alignItems: "center" }}>
          <button onClick={downloadHTML} style={{ padding: "8px 20px", fontSize: 14 }}>
            Save as HTML ↗
          </button>
          <button onClick={() => { setFiles([{ level: 0, content: SEED }]); setDone(false); }} style={{ padding: "8px 20px", fontSize: 14 }}>
            Reset
          </button>
        </div>
      )}
    </div>
  );
}

---

Examples

Minimal:

"meta-chain: attention is the scarcest resource in modern life"

→ depth 5, mode causal, runs immediately.

With depth:

"meta-chain depth 3: NFTs failed"

→ depth 3, mode causal.

With mode:

"steelman meta-chain on open-source AI, depth 4"

→ depth 4, mode steelman.

On a previous output:

"now counter meta-chain the level 3 output"

→ take the level 3 text from the previous chain, run a new chain in counter mode from that seed.