explore

Explore — Settle Unknowns by Oracle

Announce at start: "I'm using the attune explore skill to route unknowns by oracle."

Overview

Knowledge has three sources, and each source settles a question in exactly one way:

• WORLD-owned — feasibility, prior art, API truth, current fact. The world holds the answer. Settle by research; never ask the user.
• HUMAN-owned — preference, taste, budget, tacit constraints, output style. Only the user's head holds the answer. Settle by asking; never assume.
• NOBODY-owned — no one knows until it is tried. Settle by experiment (the attune experiment skill); never ask, never assume.

This skill runs that routing eagerly during any discussion. Until its oracle answers, a question stays open, no matter what is already written down.

Eager research (WORLD-owned)

The moment you classify an unknown as WORLD-owned, settle it by grounded web research, one researcher per unknown or candidate. Every researcher writes its full brief to ${TMPDIR:-/tmp}/attune-explore/<topic>/<unknown>.md — the brief files are the evidence trail the user can check; the thread carries only the compiled digest.

Launch the researchers as background subagents with the Agent tool, in parallel (one message, multiple Agent calls). The conversation continues while they run, but the digest does not: wait for every spawned explorer to complete — never a partial digest as briefs land.

Spawning explorers:

MUST:

1. You MUST spawn every explorer with model: "sonnet" — never the session model.
2. You MUST spawn every explorer as the general-purpose subagent type, restricted by prompt to research plus one brief-file write — the built-in Explore type cannot write files.
3. You MUST ground every explorer with web search.
4. You MUST carry the whole response contract in the spawning prompt — the subagent interface enforces no schema.
5. You MUST state the response requirements as MUST: / MUST NOT: lists.
6. You MUST require the brief file to take the template below.
7. You MUST require a dated source for every claim.
8. You MUST keep every field fail-open — "none" is a valid entry.
9. You MUST require the explorer's reply to carry only the brief's file path and an abstract of at most three sentences.

MUST NOT:

1. You MUST NOT let the template bind the research behind the brief — it binds only the written brief.
2. You MUST NOT let the template pressure a fabricated source, conflict, or confidence.
3. You MUST NOT use all-caps section titles or labels in the template.
4. You MUST NOT let a full brief travel back through an explorer's reply — it lives in the file.

Required brief template:

## Findings
- Claim: <one sentence>
  Evidence: <what supports it>
  Source: <name or URL, date>
(one entry per claim)

## Conflicts
<disagreements between sources, or "none">

## Confidence
<one line: how settled this is, and what would change it>

Compiling the digest: read every brief file and validate it — check source credibility, check dates and discard stale facts, resolve conflicts by source authority and recency; never conclude without validation, and never rewrite a brief file while validating it. Then report one digest: for each unknown, the settled answer in one or two sentences, the confidence, any conflict worth surfacing, and the brief's file path so the user can check the facts.

The interview (HUMAN-owned)

1. Pick the single most decision-changing open HUMAN-owned unknown.
2. Ask exactly one question via your platform's structured question tool (AskUserQuestion on Claude Code), or directly in conversation when none exists, carrying your recommended answer and the reason for it — never a bare open prompt, and never a menu for a decision you can reason out yourself.
3. Before the next question, state the settled ruling explicitly and carry it into the work at hand — the decision it unblocks, the artifact the current task owns (the active plan, the document under discussion, the code). A ruling left implicit is not settled.

Constraints:

• The interview belongs to the main thread. Never delegate it to a subagent.
• A session tolerates roughly 4–6 question calls. Ration them: ask only the unknowns that change decisions, and batch up to 4 related forks into one call when they are independent.

Experiments (NOBODY-owned)

Invoke the attune experiment skill through your platform's skill tool. Its outcome returns here as evidence; the user still rules on it, and the ruling carries into the task like any other.

Boundaries

Never ask the user a WORLD-owned question; never silently assume a HUMAN-owned answer; never guess a NOBODY-owned one.