Guide · Agent skill
How to use the interview agent skill to scope a Claude agent
Updated · written with AI from the project's own documentation and checked against our catalogue data
The interview skill is a scoping assistant for people who want to build a custom AI agent. It asks six short questions about the one job the agent should do, then produces a written build sheet that lists the planned tools, platform limits, deferred features and a test plan. This guide shows you how to install the skill, trigger it with a natural request, work through the questions, and read the output. No API key or live service is needed at this stage.
Interview to Agent Build Sheetalirezarezvani/claude-skills · MIT Download the skill Quick start
- 1Install the skill. Download the ZIP from this page and unzip it into ~/.claude/skills/, or into a project's .claude/skills/ folder. In the Claude apps, upload it under Settings → Capabilities → Skills.
- 2Start with a plain request. Say something like "help me scope an agent" or "I have an idea for an agent" and describe the idea in your own words.
- 3Answer the six questions. Answer job, trigger, inputs, actions, definition of done and recurrence. Use the recommended defaults when you are unsure.
- 4Run the planner and builder. Let the skill run interview_planner.py and build_sheet_builder.py to create plan.json and build-sheet.json in your output folder.
- 5Validate the sheet. Run primitives_validator.py. Fix any FAIL and run it again. Read every WARN before you move on.
What the interview skill makes the agent do
The skill sets the agent up as a scoping consultant. It asks you one question at a time, using choice prompts where the answer is a fixed option. It keeps your own words for the job and does not invent details you did not give.
Your answers become a build sheet. The build sheet contains a table of Claude Managed Agent primitives, a list of v1 and v2 deferrals with reasons, and an eval plan. The output is a plan. The actual build happens in later phases.
- Asks six intake questions: job, trigger, inputs, actions, definition of done, and recurrence.
- Maps the answers to primitives with interview_planner.py.
- Assembles build-sheet.json with build_sheet_builder.py.
- Checks platform limits with primitives_validator.py.
Install the interview skill
Download the ZIP for the skill from this directory page. Unzip it into your personal skills folder, ~/.claude/skills/, or into a project's .claude/skills/ folder if you want the skill available only inside one project.
If you use the Claude apps instead of Claude Code, open Settings, go to Capabilities, then Skills, and upload the skill there.
- Personal install: place the unzipped folder in ~/.claude/skills/.
- Project install: place it in the .claude/skills/ folder of that project.
- Claude apps: Settings → Capabilities → Skills, then upload the skill.
Trigger the skill with a natural request
You do not need a special command. The skill is meant to start when you say things like the phrases below. It also starts when the orchestrator routes a request to the interview phase.
Describe the idea in plain language. The more concrete you are about the job, the less the skill has to ask.
- "Help me scope an agent."
- "I have an idea for an agent."
- "What should this agent be?"
Answer the six intake questions
The skill walks through six slots in order. Each one maps to part of the build sheet. Answer with the specifics you actually have. If you are unsure, say so, and the skill will recommend a default and explain why.
Job is the single task the agent should do from start to finish. Trigger is what starts it: you asking, an event, or a schedule. Inputs are what it reads, such as files, a repo, memory, email, chat, GitHub or the web. Actions are what it does, such as drafting, writing, calling APIs or running code. Done is how you would grade a good run. Recurrence is whether it runs once, on request, or on a cadence.
- Job: one job only. Two jobs should become two agents.
- Trigger: on-demand, event or schedule.
- Inputs: files, repo, memory, email, chat, GitHub or web.
- Actions: draft, write, call APIs or run code.
- Done: a rubric of short, grounded grading lines.
- Recurrence: single pass, grade loop or cron loop.
Run the three bundled scripts
After the interview, the skill runs three Python scripts. You can also run them yourself. The first turns your answers into a plan file. Connectors you name as inputs become schema-true mock custom tools in v0, and real MCP servers are listed as v1 deferrals. Irreversible actions such as sending or publishing become v2 deferrals behind always_ask.
The second script assembles the plan into build-sheet.json. The third checks the sheet against platform limits and reports PASS, WARN or FAIL. A FAIL blocks progress, so fix the problem and run the check again. A WARN is advisory and should be read before you continue.
- interview_planner.py: answers to a primitives skeleton and deferrals.
- build_sheet_builder.py: assembles and normalises build-sheet.json.
- primitives_validator.py: validates against CMA limits (PASS, WARN or FAIL).
python3 scripts/interview_planner.py --job "Triage overnight support email" --trigger schedule --inputs "gmail,memory" --actions "label,reply" --dod "one label per email, grounded reason, no invented facts" --recurrence daily --out ./my-agent/plan.json
python3 scripts/build_sheet_builder.py --plan ./my-agent/plan.json --out-dir ./my-agent
python3 scripts/primitives_validator.py --sheet ./my-agent/build-sheet.jsonRead the output and what to do next
The build sheet is your handoff document. Check that the job is still one job, that the rubric lines are specific enough to grade a run, and that every deferral has a reason. Deferred items are not forgotten; they are versioned for later phases.
When the sheet passes validation, the skill records the plan in the goal state and moves to the next phase, stage-launch, which turns the sheet into payloads. Stage-launch is a separate skill and is not part of this one.
- Confirm the job is singular.
- Confirm each rubric line can be checked against a real output.
- Confirm connectors are mocked in v0 and real ones are listed as deferrals.
Safety notes: scripts and network
The skill runs scripts. The three Python scripts use only the Python 3 standard library. The skill does not use the network, so it does not call external services while it scopes your agent. It also needs no API key in this phase.
Because the skill runs local scripts, read them before you install it if you are working in a sensitive environment. Keep the output folder in a location you control, since the plan files are written to the path you give with --out or --out-dir.
- Requires Python 3.
- Runs scripts: yes.
- Uses network: no.
- No API key is needed in this phase.
What people use it for
- Scoping a support triage agent that labels overnight email, with a plan for mocked mail connectors in v0.
- Turning a founder's idea for a research or summary agent into a one-job build sheet with a grading rubric.
- Checking a planned agent against documented CMA limits before a developer starts wiring it up.
- Listing the features to defer to v1 or v2, such as real MCP servers or actions that need approval.
Frequently asked questions
What is the interview agent skill?
It is a Phase 1 skill for building a Claude Managed Agent. It interviews you about one job the agent should do and produces a build sheet with a primitives table, deferrals and an eval plan. It does not build or run the agent.
Do I need an API key to use it?
No. The source states that no API key is needed in this phase, because the output is a plan. The key is a concern for the later Phase 2.
What do I need to run it?
You need Python 3, since the scripts use only the standard library, and a Claude Code or other SKILL.md-compatible agent environment. The skill runs scripts and does not use the network.
Is the interview skill free to use?
The skill is released under the MIT licence, as stated on this page. The source does not mention any price, so check the repository for any terms that apply to your use.
What does it mean when the validator returns FAIL or WARN?
A FAIL blocks progress. Fix the sheet and run primitives_validator.py again until it passes. A WARN is advisory: the check still passes, but you should read and consider it before continuing.
How is this different from stage-launch?
The interview skill produces the plan as build-sheet.json. Stage-launch is the separate skill that turns that sheet into payloads. Use this skill first, then stage-launch.
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 if you have an agent idea but no defined scope yet, or if you want a structured plan and a limits check before a developer starts building. Pick something else if you need a finished, running agent now, or if you expect real email or chat integrations in v0, since connectors are mocked and real ones are deferred to a later version.
Download the skill