AGENT SKILL

research

Launch an autonomous research agent that iterates toward a goal through proofs, code, simulations, tests, and counterexample search. Fire-and-forget: the agent runs on its own and writes all results to .research/ in the target project.

License 许可
MIT
最近更新
2026-10-23
Stars
★ 0
Forks
0

01 前言 Front Matter

name
research
description
Launch an autonomous research agent that iterates toward a goal through proofs, code, simulations, tests, and counterexample search. Fire-and-forget: the agent runs on its own and writes all results to .research/ in the target project.
argument-hint
<goal> [eval script at <path>] [budget: N cycles, H hours]

03 SKILL.md 正文

Autonomous Research Agent

Launch the researcher on: $ARGUMENTS

Step 1: Parse the Request

Extract from the arguments (or the user's message):

  1. Goal (required): the natural-language research objective, the main content of the request.
  2. Eval script (optional): a path to an executable script ("eval script at ...", "use ... as eval", or a path ending in .sh, .py, or similar).
  3. Budget (optional): a cycle cap and/or a wall-clock cap ("at most 30 cycles", "stop after 4 hours"). Omitted means unbounded; the agent still concludes on its own stall rule.

If the goal is unclear, ask the user to clarify before proceeding. If no eval script is mentioned, the agent self-evaluates.

Step 2: Verify Eval Script (if provided)

Verify the path exists (Read tool). If it does not, tell the user and ask whether to proceed with self-evaluation instead.

Step 3: Check for an Existing Run

If .research/ exists in this directory (Glob tool), do not launch fresh:

  • Read .research/run.json if present. The run is live when status is running or stopping and heartbeat is less than stale_after_minutes old (compare against date -u +%Y-%m-%dT%H:%M:%SZ). If live, refuse: another researcher is working here. Point at /research-agent:status and /research-agent:synthesize. Never archive a live run.
  • Otherwise offer /research-agent:resume to continue it, or archiving it before starting fresh. Archive only on the user's explicit confirmation, with Bash: mkdir -p .research-archive && mv .research .research-archive/<UTC-timestamp>.

The agent itself refuses to start a fresh run on top of an existing .research/.

Step 4: Launch the Agent

Spawn the researcher (Agent tool, subagent_type: research-agent:researcher) with:

<goal>
{the user's research goal, verbatim}
</goal>

<eval>
{path to eval script, or "none - use self-evaluation"}
</eval>

<budget max_cycles="{N}" max_hours="{H}"/>   (omit the tag, or either attribute, when not given)

<working_directory>
{absolute path to current working directory}
</working_directory>

Begin your research per your fresh-mode initialization.

Then inform the user:

Research agent launched. It works autonomously in this directory, writing progress to .research/. When it finishes, the synthesis will be at .research/synthesis.md.

While it runs:

  • /research-agent:status shows cycle progress, current focus, budget, and eval trend (read-only)
  • /research-agent:synthesize asks it to wrap up and write the synthesis now
  • /research-agent:resume continues the run after an interruption or session restart
  • /research-agent:branch <name> [focus] forks a competing strategy

04 同仓库更多 Skills

Abdulqudduz
Abdulqudduz/claude-anvil

worldsmith-methodology

Use when working in a project with a worldbuilding documentation ecosystem (docs/ or lore/ directory). Provides the documentation-first methodology: canonical hierarchy, doc roles (timeline, lore, characters, systems), propagation discipline, and the dual canonical and exploratory workflow.

★ 0
Abdulqudduz
Abdulqudduz/claude-anvil

prose-craft

Use when writing fiction, editing fiction prose, reviewing chapters, or drafting scenes. Covers prose craft, dialogue mechanics, scene structure, and common failure modes in AI-assisted fiction. Trigger phrases: "write fiction", "edit chapter", "prose review", "scene draft", "dialogue check", "fiction craft", "edit this scene", "review my prose".

★ 0
Abdulqudduz
Abdulqudduz/claude-anvil

vista

This skill should be used when the user asks about research directions, open problems, future work, or follow-up research from the academic literature. Trigger phrases include "open problems in X", "what's next for Y", "future work for paper Z", "research directions on T", "salient follow-up research", "what should I work on next in", "find me follow-up research unrelated to my prior work", "broad survey of W", "neglected directions in V". Routes the request to the right Vista MCP tool, reads back the structured paper sections, and synthesizes research directions in the conversation.

★ 0
Abdulqudduz
Abdulqudduz/claude-anvil

synthesize

Make a research-agent run conclude and write synthesis.md (outcome, key findings, failed approaches, open questions, recommendations). Signals a live researcher to wrap up, or launches one in synthesis mode when none is running. Use when results are good enough or the agent is stalling.

★ 0
Abdulqudduz
Abdulqudduz/claude-anvil

status

Show the current state of a research-agent run from .research/: liveness, cycles and budget, sub-problem and hypothesis statuses, eval trend, current focus, and branches. Read-only.

★ 0
Abdulqudduz
Abdulqudduz/claude-anvil

resume

Resume an interrupted research-agent run from .research/. Re-launches the researcher to reload its structured state and continue from the current focus, optionally on a named branch or with a new budget. Use after a session restart, crash, or explicit pause.

★ 0