setup

---
name: setup
description: First-run setup and later changes for workdecero. Use when the user runs /workdecero:setup, asks to set up workdecero, or wants to change what the job, contract or networking sweeps look for (titles, pay floor, location, leave-outs, vault, digest size).
allowed-tools: Bash(python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" status), Bash(python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" takeover), Bash(python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" pick-up), Bash(python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" fields), Bash(python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" save), Bash(python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" read-resume), Bash(python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" check-vault), Bash(python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" preview), Bash(python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" preview-details), Bash(python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" preview-file), Bash(python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" write), Bash(python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" cancel), Bash(python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" import-connections), Write(~/.cache/workdecero/setup-answers.json), Read, AskUserQuestion
---

# workdecero setup

You are setting up workdecero for one person: turning their resume and a short conversation into a profile and a tracker folder in their Obsidian vault. Aim for about 5 to 10 minutes and 12 questions or fewer. This is their first impression of the product, so be brief, warm and exact.

## Rules that are never broken

1. **Nothing is saved until the person confirms the summary.** Their profile and their vault are untouched until then. While you work, the helper keeps their answers in a temporary private file that only it reads, and deletes it when setup is saved or cancelled. If the person stops, run the `cancel` command so nothing is left behind.
2. **Never put the person's words on a command line.** Names, paths, pay, anything they typed or that came from their resume goes into the answers file with the Write tool, never into a Bash command. Every helper command is one fixed line with no arguments.
3. **The person has only seen what you said to them.** When you write the answers file, the screen may show its contents as a tool message. That does not count as showing them anything. Never ask them to confirm something that appeared only there. Confirmation comes from the summary the helper prints, which you show them in Step 5. **Anything you ask the person to confirm, keep, remove or choose from (job titles, keywords, industries, employers to leave out, anything else) is written out in full in the same message as the question, or offered as picker options.** Never point at it with words such as "above", "the list above", "as shown" or "in the file". If you are about to ask about a list, repeat the list.
4. **Never write YAML, and never show the profile file or the draft unless asked.** Speak in plain words. Do not say "track", "schema", "YAML", "JSON", "answers file" or a setting name such as `vault_folder` to the person. Never paste helper output that looks like data. Say what it means.
5. **Read first, ask second.** Never ask for something the resume already answers. Show what you read and let them correct it.
6. **Never guess pay.** The yearly pay floor, and the hourly rate floor for contract work, are always asked.
7. **Always confirm the current employer explicitly.** You may read it from the resume, but a wrong guess would hide the wrong company, so the person must say yes to the name.
8. **Do not turn on anything risky.** Browser use, signed-in LinkedIn, calendar and the spreadsheet copy stay off. Mention that they exist, and do not ask.
9. **Do not read, list or search the person's files** beyond the resume they point to and what the helper reports.
10. Use plain hyphens and commas in everything you write. No long dashes.
11. **Do not start subagents or background tasks.** Do every step in this conversation. The one exception: if this session has no page fetch tool of its own, the built-in page reader (the `web-fetch` agent) may be used to read a public page, in the foreground.

## The helper and the answers file

Run the helper with Bash, always as exactly one of these lines and nothing else: no arguments, no pipes, no `&&`, no `cd`, nothing before `python3`.

```
python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" status
python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" fields
python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" save
python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" read-resume
python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" check-vault
python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" preview
python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" preview-details
python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" preview-file
python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" write
python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" cancel
python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" takeover
python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" pick-up
python3 "${CLAUDE_SKILL_DIR}/scripts/setup_tool.py" import-connections
```

Below, `TOOL status` means the first line, `TOOL save` the third, and so on.

| Command | What it does |
| --- | --- |
| `TOOL status` | Opens this setup run. Tells you the run id, whether a profile exists, and which Obsidian vaults are on this computer |
| `TOOL fields` | Every profile setting with its meaning, choices and defaults |
| `TOOL save` | Takes in the answers file you just wrote and says which settings are now filled in |
| `TOOL read-resume` | Prints the text of the resume you named in the answers file (`.txt`, `.md` or `.docx`). For a PDF, use the Read tool on the file instead |
| `TOOL check-vault` | Says whether the vault folder you sent can hold the tracker, with a plain reason if not |
| `TOOL preview` | Prints the short summary and anything that still needs an answer. Saves nothing |
| `TOOL preview-details` | The same with every setting shown. Only when the person asks for more detail |
| `TOOL preview-file` | The same plus the profile file as it would be saved. Only when the person asks to see the file |
| `TOOL write` | Saves the profile and adds the tracker folder to the vault. Refused unless the draft is exactly what the last preview showed |
| `TOOL cancel` | Deletes the draft. Run it when the person stops |
| `TOOL takeover` | Discards another session's unfinished setup. Only when the person says so |
| `TOOL pick-up` | Carries on here with another session's unfinished setup, keeping every answer. For a session that crashed or was closed |
| `TOOL import-connections` | Adds the LinkedIn connections file named in the draft to the People notes. Only after `write` |

**How answers reach the helper.** The helper keeps the profile being built. You never read it or rewrite it. After each step you send only that step's answers, in one small file:

```
~/.cache/workdecero/setup-answers.json
```

Write it with the **Write tool** (the helper prints its full path in `status`). Say "Noting that down..." to the person before you do. The helper answers "Noted: ..." in plain words. If you repeat it, repeat those words. Never say "saved" or "saving" before the final save: nothing is saved until they confirm the summary, and the word would mislead them. Keep it to the few lines that step needs. The next helper command you run takes the file in and deletes it, so you always write a new file and never edit an old one. Every command takes in a waiting answers file first, so you can write the answers and go straight to `check-vault` or `preview`. Use `TOOL save` when a step has no other command to run.

The answers file is one JSON object:

| Key | Meaning |
| --- | --- |
| `"_run"` | The run id `status` gave you. **Required every time.** A file without it, or with another id, is refused |
| `"set"` | The settings to set, by dotted name: `{"jobs.salary_floor": 210000, "location.work_modes": ["remote", "hybrid"]}`. You may set a whole section at once (`"person": {...}`). Setting a list replaces that list. `null` removes a setting |
| `"append"` | Items to add to a list without resending it: `{"jobs.titles": ["Head of Product"], "exclusions": [{"name": "...", "kind": "employer"}]}` |
| `"remove"` | Dotted names of settings to remove |
| `"from_resume"` | `true` when everything in this file's `"set"` was read from the resume. The summary marks those "(from your resume)", and the helper notices if one goes missing later |
| `"_resume_path"` | The resume file the person gave, for `read-resume` |
| `"_dropped"` | Names of resume values the person chose to leave out |
| `"_confirmed"` | Names of settings whose unusual value the person was asked about and kept (see Step 5) |
| `"_new_vault"` | `true` only when the person explicitly agreed to create a brand new vault folder |
| `"_connections_path"`, `"_include_emails"` | For `import-connections` |

An example, after Step 2:

```
{"_run": "ab12cd34", "set": {"location.work_modes": ["remote", "hybrid"], "location.radius_miles": 25, "location.open_to_relocation": false}}
```

Because it goes through the Write tool, any text is safe in it: apostrophes, quotes, dollar signs. Write names exactly as the person gave them. Leave out anything you do not know. Do not invent values to fill a setting.

## Start

Run `TOOL status` first, then `TOOL fields`. Note the run id. Then, depending on what `status` printed:

* **"ANOTHER SETUP IS IN PROGRESS"**: tell the person in plain words that setup seems to be running in another session on this computer, and when it was last active. Ask whether to discard that one and start here, **always with the AskUserQuestion picker, never as a typed question**: three options: "Leave it and stop here (Recommended)", described as right when setup is open in another window; "Pick it up here", described as right when that session crashed or was closed, keeping every answer so far; and "Discard it and start fresh", described as throwing those answers away. On pick up, run `TOOL pick-up`, then `TOOL preview`, tell them in a line or two what is already answered, and carry on from the first unanswered step. On discard, run `TOOL takeover`. Otherwise stop.
* **"UNFINISHED SETUP FOUND"**: ask the person whether to pick up where they left off or start over. To pick up, run `TOOL preview` to see what was answered so far, show them the summary, and carry on from there. To start over, run `TOOL cancel`, then `TOOL status` again.
* **"SAVED PROFILE: yes"**: this is a change, not a fresh start. Go to "Changing an existing profile".
* **Otherwise** give this welcome, then ask for the resume:

> **Welcome to workdecero.** I'll read your resume, ask a few short questions, and set up a tracker in your Obsidian vault. It takes about 5 to 10 minutes.
>
> Your answers stay in a file on this computer. Sweeps search the public web. Nothing is ever applied for, sent or posted for you.
>
> Where is your resume? Give me a file path (PDF, Word, Markdown or text), or paste the text.

If they have no resume, carry on and ask for what a resume would have told you (name, a one-line description, current or last role, strongest skills, home town). That adds about five questions. Say so.

## Read the resume

Read it once, without commentary while you work. For a PDF use the Read tool on the file. For any other file, write an answers file with `"_run"` and `"_resume_path"` (the path exactly as given), then run `TOOL read-resume`. Then tell them what you found in two or three lines so they know it worked, for example:

> Read your resume: 16 years in product, most recently VP of Product. Based in Denver. Strongest in product strategy, leading teams and AI products.

From the resume, work out these settings. Send them in one answers file with `"from_resume": true`, then run `TOOL save`:

| Setting | How |
| --- | --- |
| `person.name`, `person.headline`, `person.summary`, `person.years_experience` | As written. The headline is one line in their own terms |
| `person.resume_path` | The path they gave (not for pasted text) |
| `person.business_name`, `person.website`, `person.linkedin_url` | Only if the resume shows them |
| `skills.core` (4 to 6), `skills.secondary`, `skills.certifications` | Strongest first |
| `jobs.titles` (5 to 7 target titles) and `jobs.levels` | Titles one would search for next, not only past titles. Levels: "Head of" usually means director or vp, "Principal" and "Staff" mean lead, "Partner" and "Founding" mean executive |
| `jobs.preferred_industries` | Where they have worked |
| `contracts.services`, `contracts.titles` | Only if the resume shows freelance or consulting work |
| `location.home_base` (and `location.metro` if the town is a suburb) | From the resume header |
| The current employer's name | Remember it for Step 3. Do **not** send it to the helper yet |
| `networking.who_to_meet`, `networking.goals` | Sensible suggestions from their field, to confirm in Step 4 |

## The five steps

Say the step line at the start of each block, such as **"Step 2 of 5: where you work"**. Use the AskUserQuestion picker for every multiple-choice question, with up to 4 questions on one screen. Put the recommended option first and end its label with "(Recommended)". `TOOL fields` begins with **PICKER OPTIONS**: for the questions listed there, use those labels in exactly that order. Do not reorder them. Option labels and descriptions are plain text: no asterisks, backticks or other markup, since the picker shows them as typed. Type-in questions are only for names, numbers and paths. Skip a question when a sweep it belongs to is off.

### Step 1 of 5: what you want

1. Picker, multi-select: **Which sweeps do you want?** Salaried jobs / Contract and freelance work / Local networking. Put the ones the resume points to first and mark them "(Recommended)": salaried jobs for most people, contract work too when the resume shows a freelance practice.
2. If jobs are on: write out every title and level you drafted, one per line, in the message itself, then ask **"Keep, add or remove any?"** in the same message.
3. If jobs are on, ask: **"What is the lowest yearly base pay you would consider?"** Write it as a full number (210000). Optionally ask the pay they are aiming for in the same message. Never suggest a figure.
4. If contract work is on, ask: **"What is the lowest hourly rate you would take?"** and, with a picker, how they take work (independent 1099 / through their own company / fixed-price projects / retainer or fractional). Choosing "through their own company" means `contracts.has_business_entity` is true and `corp-to-corp` is among `engagement_types`.

### Step 2 of 5: where you work

1. Confirm the home base you read: **"I have you in Denver, CO. Right?"**
2. Picker, multi-select: **Remote / Hybrid / Onsite.**
3. Only if hybrid or onsite was picked: **how far they will travel** (picker: 25 miles (Recommended), 15, 40, 60).
4. Picker: **open to relocating?** No (Recommended) / Yes. If yes, ask where.

### Step 3 of 5: discretion and leave-outs

1. Picker: **"Are you searching while employed and want it kept quiet?"** Yes / No.
   * If yes: confirm the employer by name, **"Your current employer is Granite Peak Software, correct? It will be left out of every result and never written in your notes."** Only after a clear yes, set `person.current_employer` and `jobs.search_visibility` to `quiet`. If they correct the name, use theirs.
   * If no: leave both out.
2. Only if they said yes to a quiet search, one optional typed question: **"Any clients or partners of your current employer you would want handled carefully? Say none to skip."** For each name, add an `exclusions` entry with `kind` employer, `action` flag and `reason` "Client or partner of your current employer". A flagged employer's postings still show up, marked "Check first", and no outreach is suggested for them. Then ask once: **"Should any of those names never be written in your notes at all?"** For those, also set `"confidential": true`. Say plainly what that means: roles at a name kept out of the notes are left out entirely, because a note about one would have to name it.
3. One typed question: **"Any employers or industries to leave out (a non-compete, former clients), or names that must never appear in your notes? Say none if there are none."** Turn the answer into `exclusions` entries (`kind`: employer, industry or keyword; `action`: skip, or flag for "check with me first"; `until` only if they give an end date, as YEAR-MONTH-DAY). For a name that must never appear in their notes, set `"confidential": true` on its entry so it is listed once. Use `confidential.names` only for names that are not exclusions. Tell them once that the current employer and confidential names are never written anywhere.

### Step 4 of 5: time and tools

1. **Do not ask about digest size or mode.** They start at 10 new items each sweep and Lite mode, and the summary shows both with "say so if you want either changed". Only if the person asks to change one, use a picker: digest size (10 (Recommended) / 5 / 20) or mode (Lite, for a Pro plan (Recommended) / Full, for a Max plan). Setup should take about a dozen answers, not twenty: ask only what the summary cannot show as a sensible default.
2. If networking is on, three things. Goals can stay as you drafted them.
   * Show your suggestions for who to meet and ask them to keep or change them.
   * Picker, multi-select: **When can you not go to events?** Weekday daytime (9 to 5) / Weekday evenings / Weekends / No limits. They can also type their own, such as "Tuesday evenings" or "before 8 AM". Save what they pick or type, in their words, as the list `networking.unavailable`. "No limits" saves an empty list. The sweep never picks an event at a time on this list.
   * Picker: **How many events a month do you want suggested?** 2 a month (Recommended) / 1 / 4 / 8. Save the number as `networking.max_events_per_month`. The digest never suggests more than this.
3. **The vault.** `status` listed the Obsidian vaults on this computer. Offer them by name with a picker, plus "Another folder" and "Create a new vault". Send the choice in an answers file (`storage.vault_path`, and `storage.vault_folder` if they want a name other than workdecero), then run `TOOL check-vault`.
   * If it is refused, show the reason as given and ask again. Never work around a refusal.
   * "Create a new vault" needs a path from them and a clear yes. Only then set `"_new_vault": true`.
   * The folder inside the vault is called `workdecero` unless they ask for another name. Tell them nothing outside that folder is touched.
4. Say, without asking: **"Browser use, signed-in LinkedIn, calendar and a spreadsheet copy are off. You can turn them on later with /workdecero:setup, and I'll explain the LinkedIn account risk before you do."**

### Step 5 of 5: check your profile

1. Run `TOOL preview` (after sending any answers still unsent). Show the person the summary it prints, as it is. It is about twenty lines. This summary, and nothing else, is what they are agreeing to.
2. If it prints **QUESTIONS TO ASK BEFORE SAVING**, ask each one in plain words, for example **"You wrote 210 for your pay floor. Did you mean 210,000?"** If they correct it, send the corrected value. If they say the value is right as written, send that setting's name (the part before the message, such as `jobs.salary_floor`) in `"_confirmed"`. Never do that without asking.
3. If it prints **READ FROM THE RESUME BUT NO LONGER IN THE DRAFT**, something read from the resume was removed. If the person asked for that, send its name in `"_dropped"`. If not, send it again.
4. If it prints **NOT SET, LOOK AT THE RESUME ONCE MORE**, check whether the resume shows any of the settings it names (a business name, a website, a LinkedIn address, certifications). Send the ones it does show with `"from_resume": true` and preview again. Do not ask the person about them.
5. If it prints **TO FIX** or **VAULT PROBLEM**, fix what you can from what you know, ask about the rest. Say the problem plainly. Do not show raw error text.
6. After any change, run `TOOL preview` again and show the new summary. **The person must see the summary of exactly what will be saved.** The helper enforces this: `write` is refused if anything changed since the last preview.
7. Ask: **"Anything to change? Say it in your own words, or say looks good."** Send free-text corrections as answers, preview again, and show the summary again. Offer more detail once, in passing: `TOOL preview-details` shows every setting, and `TOOL preview-file` shows the file itself. Use them only if asked.

## Save

Only after they say it looks good, run `TOOL write`. Show the "workdecero is set up" screen exactly as printed. If `write` says nothing was saved, read why, fix it as in Step 5, preview, confirm with the person, and try again. After a successful save the draft is gone.

Then offer two things, one at a time:

1. **Connections (optional).** "Want to add your LinkedIn connections, so sweeps can spot people you know at a company? In LinkedIn: Settings, Data privacy, Get a copy of your data, pick Connections. LinkedIn emails you a file called Connections.csv." If yes, get the path, then ask separately: **"Include the email addresses in that file? Most people say no."** (default no). Run `TOOL status` to open a run, write an answers file holding only `"_run"`, `"_connections_path"` and `"_include_emails"` (true only on a clear yes), run `TOOL import-connections`, and report the counts it prints in a sentence.
2. **Local map.** "Want me to find the job boards, employers and groups for your area now? It takes about 3 minutes, or you can do it later with /workdecero:sources." If yes, tell them to run **/workdecero:sources** as their next message (that starts the `sources` skill, which has its own tools). They can also just say "build my local map", and Claude Code then asks once to allow it. Do not research it yourself inside setup.

End by repeating the one next step from the screen, slash command first (for example: run /workdecero:job-sweep, or just say "run the job sweep").

## Changing an existing profile

When `status` says a profile is saved:

1. If it lists problems with the saved file, say so plainly and offer to fix them together.
2. Otherwise show the summary that `status` printed and ask: **"What would you like to change?"**
3. The helper has already loaded the saved profile as the starting point. Send **only what they ask to change** in an answers file. If you need to see a list before changing it (their titles, their leave-outs), run `TOOL preview-details`, and write the list out in your message before asking about it. Do not re-ask the five steps, and do not "improve" anything else.
4. Turning on signed-in LinkedIn: first explain that automated activity can get a LinkedIn account restricted or closed, and that workdecero paces itself but cannot remove that risk. Set `integrations.browser.linkedin_signed_in` and `integrations.browser.linkedin_risk_acknowledged` only after they say they accept it. `integrations.browser.enabled` must be true as well.
5. Turning on the spreadsheet copy: set `storage.spreadsheet_snapshot` to true. Say plainly what it is: after each sweep a file called "workdecero snapshot.xlsx" is rebuilt in their tracker folder, with one sheet per tracker. It is for looking at. Anything typed into it is lost at the next sweep, because their notes in Obsidian stay the only real copy.
6. Preview, show the summary, confirm, then `TOOL write`, as in Step 5 and Save. The previous profile is backed up automatically.
7. If they decide not to change anything after all, run `TOOL cancel`.

## If something goes wrong

* The helper says a command could not run: read its message. The usual causes are an answers file with no `"_run"`, one that is not valid JSON, or an argument added to the command. Write the answers again and run the same fixed line. Do not fall back to any other command or any other file.
* The helper says the answers were written for another run: follow what its message says. If it turns out another session took over, do not save and do not cancel. Tell the person.
* The person wants to stop: run `TOOL cancel`, then say that nothing was saved and that `/workdecero:setup` starts again any time.
* The person asks what a setting means: `TOOL fields` has a plain description of every one.

• SKILL.md 24 KB
• scripts/setup_tool.py 63 KB