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 and follows. It is for the moment when you describe a capability in plain words, such as 'implement X for me', and want a working result rather than an interview. This guide explains what the skill makes the agent do, how to install it, how to trigger it, and what to check before you trust the output.
Research Implement Featurewanshuiyin/Auto-claude-code-research-in-sleep · MIT Download the skill Quick start
- 1Download and unzip the skill. Download the skill ZIP from the directory page and unzip it so that SKILL.md sits in a skill folder.
- 2Place it where your agent looks. Copy the folder to ~/.claude/skills/ for all projects, or to .claude/skills/ inside one project. In the Claude apps, upload it under Settings → Capabilities → Skills instead.
- 3Make a plain request. Describe the capability you want, for example 'implement a CSV parser that reports malformed rows'. The skill triggers on phrases such as 'implement X' or 'prototype then extend'.
- 4Choose effort and ask settings if needed. Leave the defaults (balanced effort, ask never) or add the options in the argument format to change the rung budget or to be asked about semantic choices.
- 5Review the ledger and output. Open implement-stage/ASSUMPTIONS.md to see every guess, then check BUILD_NOTE.md for the rungs that passed and any blockers before relying on the code.
What the skill makes the agent do
Given a request to build something, the agent first produces a thin end-to-end version that actually runs, which the skill calls the spine. Only after that does it add features, one rung at a time, and each rung has its own acceptance check. Earlier rungs must keep passing as new ones are added.
Every decision the request left open is written to an assumption ledger, named ASSUMPTIONS.md, before the code that depends on that decision is written. A final sweep looks for assumptions the agent made without declaring them. The result is that you can see what was guessed, not only what was built.
- Spine first: a minimal path from entry point to artifact, with stubs inside, must run before any feature is added.
- Ledger first: each under-determined choice that changes an interface or meaning gets a row, with its chosen value, a class and a source.
- Output folder: implement-stage/ containing SPEC.md, ASSUMPTIONS.md, BUILD_NOTE.md and SILENT_ASSUMPTION_SWEEP.json.
Install the skill (ZIP or Claude apps)
Download the skill ZIP from the directory page and unzip it. The skill folder contains SKILL.md, which is the file the agent reads. You can install it in one of two places depending on how you use Claude.
For Claude Code, place the folder in your user skills directory at ~/.claude/skills/, or in a single project at .claude/skills/ inside that project. For the Claude apps, upload the skill under Settings → Capabilities → Skills instead.
- User-wide: ~/.claude/skills/
- Single project: .claude/skills/ inside the project folder
- Claude apps: Settings → Capabilities → Skills, then upload
How to trigger it with a natural request
You do not need to name the skill. The skill description lists phrases that route a request to it. Phrases in the source include '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' ('build me one that runs'). It also triggers when you hand over a capability description rather than an experiment plan.
Arguments can narrow the run. The skill accepts a target description and three optional settings: effort level, ask mode and a base repository URL. The argument hint is shown below.
- 'Implement a CSV parser that reports malformed rows.'
- 'Build this feature: export the results as a Markdown table.'
- 'Prototype then extend a command-line tool that renames photos by date.'
[what-to-build] [— effort: lite|balanced|max|beast] [— ask: never|semantic] [— base repo: <url>]Choose effort and ask settings
Effort sets how much work the agent may do. The default is balanced. Under balanced the agent gets 5 rungs to build, 5 fix attempts per rung and 2 sweep rounds, and it surveys reuse with an ecosystem search. Lite is smaller and max and beast allow more rungs and attempts.
Ask controls who decides ambiguous choices. The default, never, means the agent declares each assumption and proceeds without waiting. With ask set to semantic, the agent stops at defined batch points and asks you about semantic choices before acting. Your answer is required to continue; if no one is available to reply, the skill says so and stops rather than quietly switching modes.
- effort lite: 3 rungs, 3 fix attempts, 1 sweep round
- effort balanced (default): 5 rungs, 5 fix attempts, 2 sweep rounds
- ask never (default): declare and proceed
- ask semantic: ask first at batch points, then wait for your reply
What it needs to run
The skill is written for Claude Code or a compatible SKILL.md runtime. It needs shell access and permission to read and write files, because it runs commands, creates the implement-stage/ folder and edits code. Its allowed tools include Bash, Read, Write, Edit, Grep, Glob and a question tool.
Some reviewer backends are optional. The skill's allowed tools name external reviewers such as Codex, Antigravity, Grok and a Claude review server. These are used only if you have them configured; the source does not require them for the basic build.
- Claude Code or a compatible SKILL.md agent runtime
- Shell (Bash) access and file read/write permissions
- Optional reviewer backends such as Codex, Antigravity, Grok or a Claude review MCP server
- Network access is listed as used by this skill
Safety notes before you run it
The skill runs scripts and uses the network, according to its listing. Read the skill's files and the commands it proposes if the target is in a sensitive environment. Run it in a project you can reset, such as a git repository or a copy of your work.
The skill runs autonomously. Its self-checks may be reviews by the same model family that wrote the code, which the skill itself marks as provisional. It can confirm that rung checks passed, but it does not confirm that the implementation is correct. Review generated code yourself before using it in production or on security-sensitive work.
- Use a version-controlled project so every change can be inspected or reverted.
- Treat the ledger as the list of guesses to review first.
- Do not rely on the skill for production or security-sensitive code without human review.
Tips for better results
Describe the capability and its inputs and outputs in as much detail as you can. Every detail you give removes a row from the ledger that would otherwise be a guess. Name the entry point, such as a command or a function, if you have one in mind.
Read ASSUMPTIONS.md before BUILD_NOTE.md. The assumptions show where the design could differ from what you wanted. If a choice is wrong, correct it in the next run or with ask set to semantic so the agent checks with you first.
- Give a base repository with the base repo option if you want to build on existing code.
- Use ask semantic if you want to be consulted on interface decisions.
- Keep the request to one capability per run so the rung ladder stays short.
What people use it for
- Turning a one-paragraph idea into a runnable command-line prototype that you can then extend.
- Building a small internal feature in an existing repository, using the base repo option to start from it.
- Producing a prototype for a student project with a written record of the design choices the agent made.
- Auditing code generated by another run by reading its assumption ledger and sweep results.
Frequently asked questions
Is the research-implement-feature skill free to use?
The skill is published under the MIT licence, which governs how the code may be used and shared. The source does not state any price, so check the directory page for current terms. Your agent runtime, such as Claude Code, may have its own requirements.
How do I install research-implement-feature in Claude Code?
Download the ZIP from the directory page and unzip it. Put the skill folder in ~/.claude/skills/ for all your projects, or in .claude/skills/ inside one project. Claude Code loads skills from these locations.
Can I use this skill in the Claude apps?
Yes. Upload the skill under Settings → Capabilities → Skills in the Claude apps. The skill still needs an environment with shell access and file permissions to do its work.
What does the assumption ledger do?
It is a table of every decision the request left open, with the value chosen and its source. The agent writes each row before the code that depends on it, so you can see what it guessed and correct it.
What are the alternatives if I want a specific experiment run instead?
This skill is for building, not running experiments. The source says to use the experiment tools instead when you have an experiment plan, for example /experiment-bridge for a written plan or /run-experiment to launch an existing one.
Why does the skill need network access and scripts?
The skill's listing says it runs scripts and uses the network. The source does not list which scripts or endpoints, so review its files and run it in a version-controlled project before trusting the output.
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 plain description and a written record of the guesses behind it. Pick something else if you need a specific experiment run, or if you want to be consulted on every decision, unless you set ask to semantic and stay available to reply. Production or security-sensitive code should always get human review before use.
Download the skill