Schedules and manages multi-step autonomous research cycles.
Install
mkdir -p .claude/skills/auto && curl -L -o skill.zip "https://agentskills.codes/api/skills/download/12577" && unzip -o skill.zip -d .claude/skills/auto && rm skill.zipInstalls to .claude/skills/auto
Activation
This is the description your AI agent reads to decide when to run this skill — the better it matches your request, the more reliably it fires.
Autonomously execute research cycles. Specify the cycle limit as an argument (e.g., /auto 2). Default: 5.Key capabilities
- →Dispatch direction-challenger for pre-direction opposition
- →Dispatch research-planner to decide the next question and plan
- →Dispatch curator for graph/lifecycle/placement and evidence absorption
- →Dispatch critic for independent provisional review
- →Dispatch guide-writer for human-facing oversight guides
How it works
The skill autonomously executes research cycles by dispatching a team of agents in a fixed sequence, delegating substantive decisions to specialized roles.
Inputs & outputs
When to use auto
- →Automate deep research
- →Schedule research agent tasks
About this skill
/auto — Research Cycle Scheduler
/auto is the thin scheduler that drives research forward. It owns no scientific judgment of its own. Each cycle, it dispatches the agent team in a fixed sequence; every substantive decision (what to investigate, what to record, what to verify) is delegated.
The team and who owns what:
| Role | Agent | Owns |
|---|---|---|
| Direction challenge | direction-challenger | Pre-direction opposition: challenges value, goal, necessity, frame, scale, authority, and inertia anchors before the direction hardens |
| Direction | research-planner | research/focus.md — reads the tree, decides the next question, expresses it as a cursor + dispatch plan + tree directives; may create a minimal child node when immediate dispatch needs that structure |
| Tree transaction | curator | Graph/lifecycle/placement, pre-worker readiness transactions, structural closure for planner-created children, state.md absorbed evidence, map.md / plan.md consistency, child presentation transactions, conventions/checks placement, analysis-material preservation/promotion, retraction, dead_ends.md, and admitted findings.md materialisation |
| Verification | critic | Independent Provisional Review of every review-eligible worker submission and Durable Surface Review of findings/analysis surfaces requested by curator |
| Execution | researcher / simulator / reader / scout / engine-builder / concept-checker / self-check | Bounded tasks producing provisional worker submissions in _reviews/ plus raw process logs as specified by their agent prompt |
| Human oversight guide | guide-writer | research/**/guide.md — session-end sweep over scheduler-supplied target nodes; writes human-facing oversight guides from durable surfaces without deciding research claims or direction |
| Session boundary | scheduler + .scripts/close-session.mjs | Transactional close of research planner's close-session packet into session log / focus / last_session / node-scoped backlog.md / agenda; beacon deletion; manifest-scoped commit + push |
/auto itself owns only: the cycle loop, the resume beacon, pre-direction challenge dispatch, parallel worker dispatch, auto-attaching Provisional Review to each worker submission, running the optional one-repair loop, dispatching curator with the right inputs, dispatching Durable Surface Review when curator requests it, detecting parent-ascent presentation boundaries, maintaining an in-memory guide-target set from scheduler-known node paths, dispatching guide-writer at Session End, collecting the session-owned stage manifest, and closing the session through .scripts/close-session.mjs. It does not create nodes itself; if research planner creates a minimal child before returning focus.md, the scheduler simply parses the new cursor/worker target and curator closes the structure later in the cycle. When research planner writes ### Pre-Worker Tree Directives, /auto runs curator before workers as an inserted readiness transaction, then continues the ordinary cycle unless curator reports that the planned worker dispatch was invalidated.
Constraints
- Write all prose in japanese. Applies to
research/focus.md,_reviews/,.logs/,agenda.md, curator's tree writes, all worker submissions. Technical terms, proper nouns, LaTeX mathematics, file/folder slugs, frontmatter keys, and the structural##headings documented in.codex/research-tree.mdand here may stay in English. The rule is about body prose, not structural tokens. request_user_inputand all other user-input solicitations are prohibited. Users are often away during/auto; asking blocks the session.- Do not initiate user-facing progress messages during a valid active session. The only scheduler-initiated user-facing message during an active session is the final draft emitted at Session End. Startup precondition failures may emit their specified stop message. If the user initiates communication mid-session, respond briefly and continue. Corrections from the user take precedence over scheduled dispatches.
exec_command("sleep ...")is prohibited; polling viaexec_command("ls ...")file-existence checks is prohibited. For waiting on agent completion use only Pattern A or Pattern B as defined inphases/dispatch.md.- Full paper text is acquired only from arXiv so source provenance stays reproducible and full-text licensing/source drift does not enter the research tree. Metadata, bibliographic records, abstracts, and non-full-text discovery may use other sources when the relevant agent prompt allows it.
- Paper writing is NOT
/auto's responsibility. Writing is handled by the/writeskill./autodrives research only.
Turn-Yielding Discipline
/auto is an autonomous loop: the user is not present between cycles, and a closing-tone message mid-session stalls the session waiting for input that never arrives. Any mid-session message that summarizes progress and waits for user confirmation is a stall, regardless of the exact wording.
- Never end a turn mid-session with a user-facing progress draft. Between cycles, the next action is a tool call — the next direction-challenger dispatch, the next research planner dispatch, the next worker batch, or Session End. If you are tempted to draft "I have finished cycle N of M; continuing with cycle N+1?", that is the stall — replace it with the actual next dispatch.
- The only user-facing closing message is the final Session End draft, emitted when
MAX_CYCLESis exhausted or the scheduler exits for an unrecoverable failure.research/focus.mdstatus labels are research-state records, not scheduler stop signals. - Compaction / reconnect / crash do not terminate a session. The
.logs/.auto-activebeacon (written at the start of every cycle) and theSessionStarthook jointly ensure the next session resumes the loop without a greeting. Seephases/session-lifecycle.md§ Resume for the mechanics and the fallback. - Progress summaries that genuinely belong somewhere go into
.logs/{timestamp}_auto.md(session log, written by.scripts/close-session.mjsfrom research planner's close-session packet) orresearch/focus.md § Context(research planner's next-cycle direction). Neither is a yielded turn.
Arguments
/auto {N} — cycle limit. Default: 5. Hereafter MAX_CYCLES.
Terminology
| Term | Meaning |
|---|---|
| Session | One /auto execution — from start to final draft |
| Ordinary Cycle | One iteration of the normal scheduler loop (direction-challenger → research planner → optional pre-worker curator readiness transaction → workers → Provisional Review → optional one repair loop → curator → optional Durable Surface Review → curator follow-up) |
| Presentation-Boundary Cycle | A cycle with a child-to-parent ascent. The child presentation transaction runs before workers; it replaces worker dispatch only when curator invalidates the planned dispatch or research planner intentionally left Worker Dispatches empty |
| Task | One spawn_agent tool call |
| Presentation Boundary | A child-to-parent cursor ascent where curator first makes the child readable as a parent component, usually by updating parent map.md / state / plan, before any parent-level workers use that context |
| Child Presentation Judgment | Research planner's meaning judgment at the boundary: what the child was for, what it achieved or failed to achieve, and what the parent should now see |
| Child Presentation Transaction | Curator's tree update at the boundary: applying the judgment to status, Current Board, parent map/plan/state, durable surfaces, archive/reframe mechanics, and link hygiene |
| Guide target set | In-memory Set of node paths the scheduler already touched or observed this session; used only at Session End to tell guide-writer which guides to inspect |
One session = MAX_CYCLES cycles unless a startup precondition fails, the user explicitly stops the session, or the scheduler reaches an unrecoverable failure. Multiple tasks can run in parallel within a cycle.
Phase Index
Detail lives in two phase files; read on demand, not all at once.
| File | Loaded when | Purpose |
|---|---|---|
phases/dispatch.md | When launching workers or critic | Pattern A / B launch methods, prompt template, Provisional Review, Durable Surface Review, per-agent dynamic data |
phases/session-lifecycle.md | Session Start and Session End | Resume check, initial sanity check, scheduler-owned session-end mechanical steps (simulation housekeeping, final sweeps), close-session packet handoff |
The research information model (tree structure, file roles, context scoping, convention ledger, provenance taxonomy) is canonical in .codex/research-tree.md — direction-challenger, research planner, and curator read it at every dispatch. /auto itself does not need it in working memory; /auto reads research/focus.md (the dispatch-spec file — treat it as the scheduler's interface with research planner, not as "tree content") only to extract the fields it dispatches on. /auto never reads node-level files directly. Node-level files are read by direction-challenger in its narrow local challenge scope, by research planner for direction-setting, and by curator for tree writing.
Session Start
Read phases/session-lifecycle.md § Session Start. Key outcomes:
- Beacon-based resume decision (fresh vs. resume mid-cycle)
- Initial sanity gates:
research/state.mdmust exist (else emit the stop message specified inphases/session-lifecycle.mdand stop);concepts/exists;.gitignorecovers.logs/.auto-active research/focus.mdexistence check (if missing, the first research planner dispatch initialises it after direction-challenger runs — see session-lifecycle)
/auto does not read node-level tree files at session start. The ancestor-chain read is research planner's responsib
Content truncated.
When not to use it
- →When full paper text is required from sources other than arXiv
- →When paper writing is the primary responsibility
- →When user-facing progress messages are needed during an active session
Limitations
- →Prohibits request_user_input and all other user-input solicitations
- →Prohibits exec_command("sleep ...") and polling via exec_command("ls ...")
- →Requires all prose to be written in Japanese, except for technical terms and structural elements
How it compares
This skill provides an autonomous, structured research cycle scheduler with specialized agent roles, unlike manual, ad-hoc research processes.
Compared to similar skills
auto side by side with the closest alternatives in the catalog.
| Skill | Installs | Updated | Safety | Difficulty |
|---|---|---|---|---|
| auto (this skill) | 0 | 3mo | No flags | Advanced |
| autoresearch | 0 | 3mo | No flags | Advanced |
| idea-discovery | 0 | 2mo | Review | Advanced |
| learn-new-things | 0 | 1mo | Review | Advanced |
Try saying
Example prompts that trigger this skill in your AI assistant.
More by RintaroMasaoka
View all by RintaroMasaoka →You might also like
autoresearch
abhidya
Stateful validator-gated research loop with native-hook persistence
idea-discovery
Shallow-W
Workflow 1: Full idea discovery pipeline. Orchestrates research-lit → idea-creator → novelty-check → research-review to go from a broad research direction to validated, pilot-tested ideas. Use when user says \"找idea全流程\", \"idea discovery pipeline\", \"从零开始找方向\", or wants the complete idea explorati
learn-new-things
Abilityai
Continuous learning heartbeat - autonomously researches, extracts insights, and expands knowledge base
nlm-skill
jacob-bd
Expert guide for the NotebookLM CLI (`nlm`) and MCP server - interfaces for Google NotebookLM. Use this skill when users want to interact with NotebookLM programmatically, including: creating/managing notebooks, adding sources (URLs, YouTube, text, Google Drive), generating content (podcasts, reports, quizzes, flashcards, mind maps, slides, infographics, videos, data tables), conducting research, chatting with sources, or automating NotebookLM workflows. Triggers on mentions of "nlm", "notebooklm", "notebook lm", "podcast generation", "audio overview", or any NotebookLM-related automation task.
deep-research
davidorex
Multi-agent parallel investigation for complex VCV Rack problems
scientific-brainstorming
davila7
Research ideation partner. Generate hypotheses, explore interdisciplinary connections, challenge assumptions, develop methodologies, identify research gaps, for creative scientific problem-solving.