Guide · Agent skill
How to use the Research Implement Feature agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
Research Implement Feature is an agent skill, a packaged set of instructions that an AI coding agent such as Claude Code loads when a request matches it. It is for people who ask an agent to just build something, such as "implement X for me", and want a working result without being interviewed first. After this guide you will know how to install the skill, trigger it with a natural request, and review the assumption ledger and diff it leaves behind.
Research Implement Featurewanshuiyin/Auto-claude-code-research-in-sleep · MIT Download the skill Quick start
- 1Download the ZIP. Download the skill ZIP from this page and unzip it. Confirm the folder contains SKILL.md.
- 2Choose an install location. Place the folder in ~/.claude/skills/ for all projects, or in .claude/skills/ inside one project. In the Claude apps, upload it under Settings → Capabilities → Skills instead.
- 3Describe what to build. Write a plain request such as "implement X for me" describing the capability you want, and say whether you want a base repository.
- 4Let the agent build in rungs. The agent builds a runnable spine first, then adds one feature at a time, each with its own check.
- 5Read the assumption ledger. Open the ledger the agent wrote and check each row, including its Source, to see what it decided for you.
- 6Review the diff and sweep results. Review the code changes and any undeclared assumptions the cross-model sweep reported, noting any REVIEW_UNAVAILABLE markers.
What the skill makes the agent do
The skill turns a plain build request into a working artifact in stages. First the agent builds a minimal version that runs end to end, which the skill calls the spine. Then it adds one feature at a time, and each feature has its own check. Earlier parts must keep passing as later ones are added.
Every decision the request left open is written to an assumption ledger before the code that depends on it is written. Once the build is done, a reviewer from a different model family looks for assumptions that were never declared.
- Stage one: a runnable end-to-end spine with stubs where needed.
- Stage two: one feature per rung, each with its own acceptance check.
- Ledger: each open decision is recorded before the code relies on it, with a Source field.
- Sweep: a cross-model review looks for undeclared assumptions.
Install the skill from the ZIP
Download the ZIP for this skill from this page and unzip it. The skill is a folder whose main file is SKILL.md. Put the folder in one of two places: ~/.claude/skills/ to make it available to all your projects, or a project's .claude/skills/ folder to make it available only in that project.
If you use the Claude apps instead of Claude Code, you can upload the skill under Settings → Capabilities → Skills.
- Personal install: place the unzipped folder in ~/.claude/skills/.
- Project install: place it in .claude/skills/ inside your project.
- Claude apps: upload it under Settings → Capabilities → Skills.
Trigger the skill with a natural request
You do not need a special command. The skill loads when your request matches its description. The description lists phrases that match, including "implement X", "build this feature", "prototype then extend", and the Chinese phrases "给我实现" and "帮我做一个能跑的".
The skill is meant for a capability description, not an experiment plan. If you have a decided target you want built, say so directly. If you want the agent to choose a research direction, a different skill is a better fit.
- "Implement a rate limiter middleware for my Express app, then add per-user limits."
- "Build this feature: export the results table as CSV. Prototype it first, then extend."
- "帮我做一个能跑的 command-line tool that converts these logs to JSON."
Optional arguments and the base repository
The skill accepts optional settings in the request. The argument hint lists effort levels lite, balanced, max and beast, an ask mode of never or semantic, and a base repository URL. The default effort is balanced and the default ask mode is never, which means the agent proceeds without asking.
If you set a base repository, the agent clones it first and builds inside it, matching its conventions. Otherwise it extends the current project or creates files in it.
- Effort levels change the rung budget and the number of fix attempts per rung.
- ask: semantic makes the agent stop at batch points to ask about choices that change meaning.
- base repo: <url> asks the agent to build on existing code from that repository.
Build: [what-to-build] — effort: balanced — ask: never — base repo: <url>What you need before running it
The skill is written for Claude Code or another agent that reads SKILL.md files. It needs Bash access so it can run scripts. Reviewer tools are optional, but the cross-model sweep uses them: Codex, Antigravity, Grok, or claude-review. If none are configured, the skill falls back to a Codex MCP default, and any reviewer that cannot be reached is marked REVIEW_UNAVAILABLE in the output.
The skill runs scripts and uses the network, so review what it does before using it on sensitive work.
- An agent that supports SKILL.md files, such as Claude Code.
- Bash access for running scripts.
- Optional reviewer MCP servers for the cross-model sweep.
- Optional base repository URL if you are building on existing code.
Review the ledger and diff after the run
In the default mode the skill does not pause for approval. Treat the ledger and the diff as your review step. Check each ledger row, especially rows marked with consequential choices, and confirm that the sweep did not find undeclared assumptions you care about. Look at any REVIEW_UNAVAILABLE markers, since those mean a check did not happen.
If you want a say before choices take effect, use ask: semantic instead of the default.
- Read every ledger row and its Source field.
- Review the full diff, including files the agent created.
- Check that REVIEW_UNAVAILABLE markers are understood before trusting the sweep.
What people use it for
- Building a quick prototype of a capability from a one-paragraph description before deciding whether to invest more time.
- Turning an already decided idea into runnable code for an experiment, with a record of the choices made along the way.
- Adding a feature to an existing repository by setting the base repo option and letting the agent follow its conventions.
- Producing an auditable list of assumptions an agent made, for a team that needs to review a build.
Frequently asked questions
Is the research-implement-feature skill free?
The skill is published under the MIT licence, as stated on this page. Check the licence file in the downloaded ZIP for the full terms. Any model or reviewer service you connect to may have its own costs.
What do I need to run research-implement-feature?
You need Claude Code or another agent that reads SKILL.md files, with Bash access so scripts can run. Reviewer MCP servers such as Codex, Antigravity, Grok or claude-review are optional but needed for the full cross-model sweep.
Where do I put the skill folder?
Place the unzipped folder in ~/.claude/skills/ for personal use, or in .claude/skills/ inside a project to use it only there. In the Claude apps, upload it under Settings → Capabilities → Skills.
What does REVIEW_UNAVAILABLE mean?
It marks a reviewer slot that could not be used, for example because its MCP server is not configured. The skill falls back to a Codex MCP default where it can, and any unavailable reviewer is marked this way in the output, so the sweep may be incomplete.
Does the skill ask me questions before building?
By default it does not. The default ask mode is never, so it declares its decisions in the ledger and proceeds. Setting ask to semantic makes it stop at batch points to ask about choices that change meaning.
When should I use a different skill instead?
If you want the agent to find a research direction and decide the question for you, the research-pipeline skill fits better. This skill expects a decided target and decides nothing of consequence silently.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use this skill when you want a working prototype from a capability description and are willing to review a ledger and diff afterwards. Pick something else if you need open-ended research direction finding, or if unreviewed autonomous decisions would be risky for your work.
Download the skill