Verify a claim by checking 2+ independent sources, before committing to your first answer.
A Claude Code skill that adds the discipline of cross-source review at specific decision points.
LLMs anchor on first instinct. When the model produces output AND evaluates it, evaluation runs on the same weights that produced it — same blind spots, same biases. Self-review under those conditions is rubber-stamp review.
triangulate is the meta-decision skill that fires existing review tools (second-opinion, adversarial-reviewer, dispatching-parallel-agents, Agent subagents) at the right moments — and covers surfaces those skills don't, like compliance copy, image prompts, architectural alternatives, and async state verification.
One command:
git clone https://github.com/danzimon-rgb/triangulate ~/.claude/skills/triangulateRestart Claude Code. The skill auto-loads on session start. The description triggers Claude to invoke it automatically when relevant; you can also invoke it explicitly with the Skill tool.
Project-scoped install (this repo only, not user-level):
git clone https://github.com/danzimon-rgb/triangulate .claude/skills/triangulate- Six trigger points where the agent must fork to an outside source automatically: compliance/legal copy, architectural alternatives, image/video prompts, debug loops >2 cycles, cross-cutting decisions, async state verification (propagation patience).
- A taxonomy of "outside sources" — different model > fresh subagent > persona shift, in order of independence.
- A maturity-test gating question to ask before shipping: "would another model produce a meaningfully different answer here?"
- Concrete invocation patterns — which tool to fire for which trigger.
- A propagation-windows reference table — Vercel, GitHub API, DNS, Supabase, Stripe, etc. — so you wait the right amount of time before claiming verification.
- Anti-patterns — when NOT to fork (routine work, after the user said ship, etc.)
Authored by Dan Zimon (founder, Teranode AI) on 2026-04-29, after watching a parallel Codex session out-architect Claude on a streaming-progress refactor. The lesson:
The cost of the 30-second pause to fork the question is much lower than the cost of shipping the inferior pattern.
The skill grew on its first day in public: Trigger #6 (async state verification, "go slow to go fast") was added the same night after Claude polled a Vercel deploy too quickly and produced wrong conclusions on still-propagating data — Dan named the discipline, we built the artifact, the world got the lesson. That's the pattern this whole skill is about.
Same principle Teranode ships at the product level — five reasoning models forking advisory questions in parallel — now applied at the developer level. Two scales, one principle.
| Existing skill | What it does | What triangulate adds |
|---|---|---|
second-opinion (trailofbits) |
Shells out to Codex CLI + Gemini CLI | The trigger logic — when to invoke it, including non-code surfaces |
adversarial-reviewer (engineering-team) |
3 hostile review personas | The non-code applications and the cross-model layer |
dispatching-parallel-agents (superpowers) |
Fans out independent debug tasks | The decision rule for whether to dispatch in the first place |
brainstorming (superpowers) |
Explores design space before deciding | The fork-on-finalists step after design space is mapped |
triangulate doesn't replace any of these — it tells the agent when to reach for them.
This skill is the static, developer-facing version of a principle that Teranode AI productizes for regulated financial advisors at scale:
- Static (this skill): fixed trigger points, fixed sources, fixed decision rule. Reliable, copy-pasteable.
- Dynamic (Teranode): learning routing layer that gets smarter with every Council run — which model excels at which question, which dissent was material, which calibration was off. Every API call is training data.
See skills/triangulate/references/teranode-routing-layer.md for the strategic breakdown of how the principle scales.
If you're an advisor or RIA who wants the productized version: teranode.ai.
MIT. Free to use, fork, modify, redistribute. Attribution appreciated, not required.
Open an issue or PR. Small, focused changes welcome. Especially:
- New trigger points the original draft missed
- Better invocation patterns for specific tools
- Anti-patterns observed in the wild
- Translations of the principle for non-Claude environments
Dan Zimon — Series 7, Series 66, insurance licensed; fourteen years across institutional finance. Building Teranode AI, decision infrastructure for regulated financial advisors.
OpenClaw ecosystem.