Super Survey skill

---
name: super-survey
description: Use when researching product opportunities, markets, open-source projects, competitors, technical feasibility, or business ideas that need multiple evidence-backed rounds and adversarial critique.
---

# Super Survey

## Overview

Super Survey turns a vague research target into progressively sharper conclusions through repeated research, red-team critique, synthesis, and persisted artifacts. Use it for product discovery, market validation, open-source project scouting, investment-style diligence, or strategic decisions.


## Operating Rules

Every survey round follows one staged judgment loop:

1. Frame the user's wording as a starting point, not the objective function.
2. Route the question through the Adaptive Research Framework: choose a decision archetype, 1-3 lens packs, domain hints, explicit framework dimensions, and an evidence standard.
3. Write the evidence plan and minimum direct evidence before current-source search.
4. Gather evidence, update `sources.jsonl`, `claims.jsonl`, and `evidence.jsonl`, and separate findings from interpretation.
5. Re-enter brainstorming after research to compare reframes and next evidence moves.
6. Red-team the strongest current argument, substitutes, weak evidence, constraints, and kill criteria.
7. Synthesize a conditional judgment with confidence, decision rationale, and remaining unknowns.
8. Run the lightweight evolver and use its raw `Keep / Narrow / Pivot / Kill / Final` decision for continuation.
9. Keep `index.md` as the round workbench and decision ledger; write `report.md` only after the raw evolver decision is `Final` or `Kill`.
10. Run `survey_round.py check` for rounds and `check-final` before final delivery.

Move beyond collecting links. The value of this skill is sharper judgment after each loop.

Empty templates are not artifacts. A round is incomplete until each file contains substantive findings, critique, synthesis, and evolved next-target content.

`index.md` is the per-round workbench, navigation page, and decision log. Reserve `report.md` for the final complete deliverable: a full standalone report in the selected artifact language, written after the raw evolver decision is `Final` or `Kill`, then validated with the final quality gate before the user-facing answer. During continuing rounds, update `index.md` for progress, decisions, and next targets.

Autonomous continuation is the default. Do not stop and ask the user how to proceed after `Keep`, `Narrow`, or `Pivot`; create the next round immediately and continue the staged workflow. Only pause for checkpoint approval when the user explicitly requested checkpoint approval before the round, a new constraint is required to continue, or a tool/environment blocker prevents meaningful work. Do not say "ready for the next round", "let me know how you would like to proceed", or similar handoff language after a continuing decision. A progress update may state that the next round is starting and name its target.

External tool approval is not checkpoint approval. If current web search, browser access, or another tool requires host/user approval, request the actual tool approval and set `index.md` Continuation Status to `Workflow State: awaiting_tool_approval`, `Pending Approval: <tool and reason>`, and `Resume Action: <exact next stage action>`. After approval is granted, immediately run the approved tool and continue the same stage; do not ask the user to confirm again, do not leave the conversation at "approval received", and do not treat the approval pause as a completed round. If approval is denied or the tool remains unavailable, record the denial/fallback in `NN-research.md` Data Quality Notes and continue with the best available fallback or mark the evidence gap explicitly.

Keep the workflow state consistent with the latest gate. Use `continuing_round` after `Keep`, `Narrow`, or `Pivot`; `ready_to_finalize` after `Final` or `Kill` before `report.md` exists; `final_report_draft` while the final report exists but is not yet delivery-ready; and `final` only when `report.md`, the quality score, residual gate, hard-constraint gate, and `check-final` are ready for delivery. `check-final` requires `Workflow State: final` and `Resume Action: none`.