Explore

Exploration

Orchestrate the exploration. Make decisions via AskUserQuestion. Delegate file operations to agents.

Delegate:

• Exploration → agents/explore-agent.md
• Log writing → agents/explore-log-writer.md

Work without stopping.

Input

$ARGUMENTS — the first question or exploration topic.

Phases

3 phases.

1. Init      → initialize variables
2. Loop      → user-driven Q&A with dispatch explore-agent (sonnet)
3. Finalize  → slug, dispatch explore-log-writer (haiku)

Phase 1 — Init

If $ARGUMENTS is empty — ask via AskUserQuestion: "What would you like to talk about?"

Treat $ARGUMENTS or the user's reply as the first question.

Read agents/explore-agent.md — use it across every Loop iteration without re-reading.

Initialize variables:

• EXPLORATION_SUMMARY = "" (cumulative chain of key findings)
• QA_LOG = [] (full log of all Q&A rounds)
• ITERATION = 0

Transition → Phase 2.

Phase 2 — Loop (user-driven Q&A)

Repeat until the user exits.

2a. Prompt enrichment

The orchestrator builds the prompt for the agent:

• Exploration topic = the user's first question
• Current question = the question from the current iteration
• Previous findings = EXPLORATION_SUMMARY

2b. Dispatch explore-agent

Dispatch via the Agent tool (model: sonnet). The agent file is already read in Init.

The agent returns structured output: RESPONSE_TYPE, ANSWER, DETAILS, SUMMARY, KEY_FILES, WEB_SOURCES. For an open question — also OPTIONS.

2c. Show the result

For RESPONSE_TYPE = answer — show ANSWER and DETAILS to the user.

For RESPONSE_TYPE = brainstorm — show ANSWER, DETAILS and OPTIONS to the user.

2d. Update state

Append the agent's SUMMARY to EXPLORATION_SUMMARY.

Append the full record to QA_LOG. Pass all fields without compression:

Q: <user's question>
CONTEXT: <why the question was asked — what led to it, what problem it solves, how it connects to previous questions>
ANSWER: <ANSWER from explore-agent — the main answer>
DETAILS: <DETAILS from explore-agent — in full, with file:line and code fragments>
SUMMARY: <SUMMARY from explore-agent — conclusion for the current question>
KEY_FILES: <list>
WEB_SOURCES: <list>
OPTIONS: <if brainstorm>

Build CONTEXT as the orchestrator — based on the flow of the conversation, previous questions and the user's current goal.

ITERATION++.

2e. Next step

AskUserQuestion — what's next:

• Ask another question
• Save and finish → go to Phase 3
• Continue without saving → end the skill, do not write the log

2f. Warning on a long session

At 20+ questions — warn: "The session is long (N questions), I recommend saving the results."

A new user question → return to step 2a.

Phase 3 — Finalize

Via AskUserQuestion offer 2-3 slug variants based on the exploration topic (prefix explore-, kebab-case, English, up to 40 characters). Add an "Other" option for manual input.

Dispatch explore-log-writer via the Agent tool (model: haiku).

Read agents/explore-log-writer.md. Pass to the agent:

• SLUG — the exploration slug
• TOPIC — the first question (exploration topic)
• QA_PAIRS — the full QA_LOG. Each record contains fields: Q, CONTEXT, ANSWER, DETAILS, SUMMARY, KEY_FILES, WEB_SOURCES, OPTIONS (if brainstorm). Pass all fields without compression.
• DATE — the current date

The agent will create the file .yoke/ai/<SLUG>/<SLUG>-exploration.md.

Print the result:

Exploration log: .yoke/ai/<SLUG>/<SLUG>-exploration.md
Questions explored: <ITERATION>

TodoWrite: "Exploration log written".

Rules

• Delegate file operations to agents — the orchestrator never touches files directly.
• Run phases without confirmations; use AskUserQuestion in Init and Loop.
• Wait for questions from the user.
• Accumulate findings in EXPLORATION_SUMMARY and pass to the agent as context.
• Mark phases via TodoWrite (Init, Finalize), not every question.
• Language: match the ticket/input language, or follow the project-level definition in CLAUDE.md / AGENTS.md.