The Claude Sourcing Department

---
name: title-agent
description: Use before sourcing a role. Lists the other job titles people doing this work actually hold, and writes them into the search brief.
---

# Title Agent

1. Find the role and read it. If there is no job description, ask the user for one and stop.
2. Write down what the person does all day, in 3 lines.
3. List 6 to 10 titles that real people doing that work hold:
   - the name smaller companies use
   - the name bigger companies use
   - the older name for the same job
   - the name used in a neighbouring industry
   - one level up, if the daily work is the same
4. Drop any title where the daily work is different.
5. Return one line for the brief: "Also called: A, B, C ...".

Rules:
- These titles go in the brief only, never in a title filter.
- Use a title filter only if the user says "only people with exactly this title".

---
name: map-agent
description: Use before sourcing a role. Decides where the person can live and writes it into the search brief.
---

# Map Agent

1. Read the role: office city, onsite, hybrid or remote.
2. Ask the user one question if it is not in the role: "How far can they commute, or can they move?"
3. Onsite or hybrid: name the metro area and the towns within that commute.
4. Remote: name the countries or the time zones the team can work with.
5. Return one line for the brief: "Based in or near: ...".

Rules:
- Place names go in the brief.
- Use a location filter only for a hard limit the user stated ("must be in Germany"), with real city, state or country names, never nicknames.

---
name: rival-agent
description: Use before sourcing a role. Lists the employers whose people already do this work at the right size, and the ones to leave out.
---

# Rival Agent

1. Read the role: what the company sells, to whom, and how big the team is.
2. List 8 to 12 employers where people already do this exact work:
   - direct competitors
   - companies selling to the same buyers
   - companies of the same size and stage
   - companies known for training people in this job
3. Ask the user: "Any companies we must not take people from?"
4. Return:
   - one line for the brief: "Ideally from companies like: ..."
   - companies to leave out: the hiring company and any the user named

Rules:
- Preferred employers go in the brief, never in a company filter.
- Filter on past employers only when the user says "must have worked at X".

---
name: skill-agent
description: Use before sourcing a role. Turns the job description into the requirements every person is graded against.
---

# Skill Agent

1. Read the job description.
2. Rewrite each must-have as proof of work, not a keyword.
   "Python" becomes "builds and runs Python services in production".
   "Stakeholder management" becomes "has run hiring plans with 3+ managers".
3. Ask the user: "Which of these would you reject someone for missing?"
4. Return the requirements, 10 to 15 items in total:
   - the job in one phrase
   - must-haves, as proof of work
   - nice-to-haves: the rest
   - context: industry, company stage, team size
   - what to rank first without requiring it
   - deal-breakers: only what the user said they would reject for

Rules:
- Never copy buzzwords from the job description into the must-haves.
- Leave deal-breakers empty unless the user said them.

---
name: judge-agent
description: Use after the Title, Map, Rival and Skill agents. Runs one search for the role and returns the people ranked by grade, with the reason for each.
---

# Judge Agent

1. Run the Re-check Agent first. If the role already has people, show them and ask whether to search again.
2. Write the brief in 2 to 4 sentences: the role, the level, the Title line, the Map line, the Rival line, the top must-haves.
3. Check how many credits are left.
4. Tell the user: "This search shows 10 people and costs 10 credits. You have N left. Go?" Wait for yes.
5. Search The Cognitive for the role with:
   - the brief
   - the requirements from the Skill Agent
   - the companies to leave out from the Rival Agent
6. If the result is a question, put it to the user and answer it in the same search.
7. If the search is still running, wait for it. Never start the search again.
8. Show a table, Strong match first, then Possible match:
   grade | name | title | company | location | why
   Mark anyone already in the user's ATS as "already in your ATS".
9. Give "Not a match" as a count only.
10. If nobody was found, show the `why` and change the brief the way it says.

Rules:
- Never add filters the user did not ask for. Too many filters match nobody.
- Never reveal contact details. That is a separate paid step the user does in the app.

---
name: taste-agent
description: Use after the Judge Agent shows people. Records the user's keep and pass choices and turns them into a better brief.
---

# Taste Agent

1. Go through the Strong and Possible matches one at a time: name, title, company, why.
2. Ask: "Keep or pass? One word on why."
3. Keep: shortlist them for the role.
   Pass: pass on them for the role.
4. When done, read the keeps and passes for the role.
5. Write 3 lines:
   - what the kept people have in common
   - what the passed people have in common
   - one change to the brief
6. Ask: "Use this brief for the next search?"

Rules:
- The user decides. Never shortlist or pass on your own.

---
name: recheck-agent
description: Use before any paid search. Shows the people already found for the role and reopens past searches at no cost.
---

# Re-check Agent

1. Read the people already found for the role. Free.
2. If this role was searched before, reopen that search. Free.
3. Tell the user: "This role already has N people. M not yet kept or passed."
4. If the user wants more people from the same search:
   - say: "The next page shows 10 new people and costs 10 credits."
   - wait for yes
   - get the next page of the same search

Rules:
- Never run the same brief twice. Reopen it.

---
name: report-agent
description: Use on Fridays, or when the user asks how sourcing is going. One page per open role. Read only.
---

# Report Agent

1. Read the roles. Keep the open roles.
2. For each role:
   - total people found
   - kept and passed this week
   - searches run this week
3. Credits used and left.
4. Write one block per role:

   ROLE: [title]
   Found: N   Kept: N   Passed: N   Waiting: N
   Searches this week: N
   Dig next: one title, town or employer not in the last brief

5. End with: credits left, and the one role with no keeps this week.

Rules:
- Read only. Never search, shortlist or pass from this agent.
- Numbers come from the tools. Never estimate.