# Triangulate Me > > - **Type:** Skill - **Install:** `agentstack add skill-jon-devlapaz-skill-scout-triangulate-me` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [jon-devlapaz](https://agentstack.voostack.com/s/jon-devlapaz) - **Installs:** 0 - **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [jon-devlapaz](https://github.com/jon-devlapaz) - **Source:** https://github.com/jon-devlapaz/tink-skills/tree/main/skills/triangulate-me ## Install ```sh agentstack add skill-jon-devlapaz-skill-scout-triangulate-me ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # Triangulate Me Sharpen the user's meaning without replacing it. Treat every interpretation as a proposal for the user to accept, reject, or revise. ## Establish the decision - Identify the single claim, choice, or assumption currently doing the work. - Split genuinely independent claims and handle them separately. - Ignore acknowledgements, logistics, and other replies with no meaningful interpretive gap; respond normally and continue the existing conversation. When the update is complete, do not invent checks, reminders, or follow-up work. - Find available facts yourself. Ask the user only for judgments, preferences, private context, or evidence you cannot inspect. **Complete when:** exactly one active claim, choice, or assumption is named. If the reply is not substantive, respond normally without running the frame. ## Triangulate each substantive answer Apply this sequence: 1. **Faithful read** — Restate only what the answer actually commits to. Complete when every qualifier that changes the meaning remains intact. 2. **Steel read** — Form the strongest coherent interpretation supported by the answer. Complete when every useful agent-added premise is explicit and labeled. 3. **Stress read** — Form the weakest interpretation that a reasonable reader could still derive from the answer. Complete when it names a specific, supported vulnerability—or states that none exists. 4. **Crux** — State declaratively the exact premise, definition, value, or missing fact that explains the distance between the readings; do not phrase it as a question. Complete when resolving it would materially collapse that distance. 5. **First-principles check, when triggered** — If the crux contains a supposedly fixed constraint (such as "must use X," "too expensive," "can't scale," or "that's how it is done"), or a solution justified mainly by analogy or authority, run the grounding pass below. Otherwise skip it. 6. **Convergence** — Recommend the most robust next formulation. Preserve the steel read's value while repairing the stress read's supported vulnerability. Complete when it states what was preserved, repaired, or left unresolved; derive it from grounded primitives when the check ran. If one position is already bounded, testable, and well-supported, endorse it directly rather than inventing objections, governance, exceptions, or extra requirements. 7. **Next question** — Ask zero or one question. Ask only when a material crux remains, and do not repeat the crux in question form. Omit the question when convergence settles the decision or no prerequisite answer is needed. Do not ask a downstream question while its prerequisite remains unsettled. For value conflicts, first ask which concrete action creates unacceptable harm and where the boundary lies. Do not jump to identity, policy, or other mechanisms until that action boundary is settled. Use the sequence as a reasoning discipline, not a mandatory output template. Lead with the decision or direct answer. Include only fields that materially change it; a clear or low-complexity decision may need one short paragraph. When the full frame helps the user inspect a real interpretive gap, use this compact form: ```markdown **Faithful read:** ... **Steel read:** ... **Stress read:** ... **Crux:** ... **Grounding:** ... *(include only when a fixed-constraint or analogy signal is present)* **Convergence:** ... **Next question:** ... ``` ## Ground constraints from first principles Use this as one bounded pass, separate from the six interpretation fields. Read [first-principles grounding](references/first-principles-grounding.md) when the check is triggered. It defines the constraint classes, primitive reduction, rebuild, residuals, and kill test. Keep the main response's `Grounding` line compact and concrete. Reserve the pass for a false-constraint or borrowed- solution signal; a proposed solution or missing evidence alone is not enough. Keep time-critical incidents and ordinary polish on their direct path. ## Protect the inquiry - Ground every interpretation in the user's expressed language and context; leave unstated motives, personality, emotions, and beliefs outside the frame. - Weight the steel and stress reads according to their evidence. - Let convergence endorse one reading, revise both, preserve an explicit disagreement, or abstain pending evidence or a prototype. - Route a factual dispute to the named check that could resolve it. - When no plausible stress read exists, say so and move to the next unresolved decision without manufacturing a crux or follow-up question. - Distinguish an ungrillable question from an unresolved one. Recommend a concrete experiment when reaction to evidence, behavior, or a prototype is required. When evidence is insufficient and the next step is an experiment or staged rollout, name its predeclared decision rule and material harm guardrail. - Keep the user's original language where possible. Mark agent-added language and assumptions plainly. Read [reasoning foundations](references/reasoning-foundations.md) when calibrating charity, adversarial pressure, integration, first-principles grounding, or abstention. Read [interaction examples](references/interaction-examples.md) when the answer is multi-claim, evidence-dependent, value-laden, or too trivial to triangulate. ## Continue and stop After the user answers, recompute the faithful, steel, and stress reads from the new evidence and treat the previous convergence as provisional. Continue until one of these conditions holds: - the user explicitly accepts a formulation and no material crux remains; - the remaining difference is an acknowledged value choice; - a named fact, experiment, or prototype must come next; - the user stops or changes the task. At the end, state the settled formulation, preserved disagreements, and any open evidence gate. Wait for the user's explicit request before planning or acting on the result. ## Source & license This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [jon-devlapaz](https://github.com/jon-devlapaz) - **Source:** [jon-devlapaz/tink-skills](https://github.com/jon-devlapaz/tink-skills) - **License:** MIT Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-jon-devlapaz-skill-scout-triangulate-me - Seller: https://agentstack.voostack.com/s/jon-devlapaz - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.