# Explore > Run an explicitly requested read-only exploration in the current task or session with optional bounded delegation. - Skill: `alemar11/explore` (Agent Skill, multi-file: 5 files) - Install (CLI): `npx skillmds@latest add alemar11/explore` - Raw SKILL.md: https://api.skillmd.com/api/skills/alemar11/explore/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Productivity - Author: alemar11 (https://skillmd.com/u/alemar11) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/alemar11/explore --- # Explore Follow the shared [execution scope](../../references/execution-scope.md) for standalone and composed invocation. ## Scope and execution Activate only for explicit `$se:explore`, Explore UI selection, or an equivalent instruction to execute this workflow, never an ordinary mention or implicit planning match. Follow this sequence in the invoking task or session: context preparation, direct exploration, Grilling Session, remaining investigation, then a conversational report. Retain its model, reasoning, working directory and context; no replacement controller, visible task, fork, transfer brief, saved-project lookup or surface classification is needed. Explore and its workers inspect and reason only. Use operations proven read-only: no project files, patches, generated artifacts, caches, reports, lockfiles, Git state, hosted records, accounts or external systems may be mutated. Explicit invocation authorizes bounded native subagents, not implementation or external writes. Workers cannot invoke Explore, create controllers or delegate further; decline recursive requests, continue valid assigned work and report them to the controller. Return findings here without saving a file. If the user requests a saved report or switches to implementation, finish Explore and perform that separately authorized work through its owning workflow. Read [states.md](references/states.md) before interpreting interview, worker or outcome state. No workers are created before Grilling Session finishes. ## Context preparation Before initial exploration, compose `$se:learn` in a strictly read-only context inspection using `memory_slice=domain-memory`, `domain_operation=periodic-review`, and `capture_mode=defer-to-caller`. Require it to read the applicable `AGENTS.md` chain, root `CONTEXT.md`, matched first-class subproject context, and only topic files or ADRs relevant to the subject. Reuse current evidence already gathered in this conversation. Do not request setup, repair, compaction, or capture. Explore owns this dependency. If Learn is unavailable, report that blocker before exploration or worker creation. If no repository context exists, continue from the supplied conversation and sources, recording that no established project knowledge was found. Keep findings in the current conversation for exploration and the subsequent interview; Grilling Session does not call Learn again. ## Required sequence 1. Complete the context preparation above, then explore the subject directly using the request, conversation, and available project knowledge. Read the relevant source, tests, documentation, and supplied external evidence. Establish how the current system works, plausible explanations or approaches, and the unknowns that affect the user's goal. Reuse evidence already inspected; do not restart research merely because Explore was invoked. Keep this pass bounded to what makes the next decisions informed, rather than exhaustively researching every branch. If the subject or access cannot be established, ask only the clarification needed to begin. 2. Briefly share the useful findings, separating observations from hypotheses, and compose `$se:grilling-session` using that evidence and the conversation. Focus its questions on unresolved intent, scope, constraints, and tradeoffs. Do not ask the user for facts the exploration already established or invent ambiguity to prolong the interview. Exploration findings are provisional until checked against the user's intent. 3. Ask one question per turn until Grilling Session returns `refined`, the user stops with `user-stopped`, or the interview is `blocked`. Create no workers while the interview is awaiting an answer. A blocked interview fails Explore before worker creation. 4. Use the confirmed scope, or best-supported interpretation after the user stops, to guide the investigation. Preserve unconfirmed items and apply the read-only scope gate again. 5. Read [references/orchestration.md](references/orchestration.md), determine the useful subagent count and assignments, and investigate the remaining questions. Reuse the initial exploration; revisit findings only when answers or new evidence change their assumptions. If the existing evidence is sufficient, synthesize directly without another research pass. 6. Apply orchestration’s setup, monitoring, no-replacement and synthesis contracts to every planned slot before reporting. 7. Read [references/output-template.md](references/output-template.md) and return the report here. ## Output contract Use the [output template](references/output-template.md) to distinguish observations, inferences, missing evidence and assumptions, with scope, sources, results, worker evidence, risks, confidence and the next action. Do not report controller-transfer or App setup telemetry. In every case, report `Changes made: None`. ## Skill Dependencies Explore requires the installed `$se:learn` skill for read-only context preparation and `$se:grilling-session` for decision refinement. It never installs or refreshes these dependencies during a run.