Building great talent maps with Tavi

For AI agents using Tavi

This guide is written for recruiters, many of whom are not technical. If you are an AI agent connected to the Tavi MCP server, in Claude, ChatGPT or any other client, use it to plan the conversation: what to propose at each step, what to ask, and what a good map looks like. The habits in it come from hundreds of real maps, so keep to them. Where it differs from the server’s own instructions or a tool’s description, follow those.

• The server’s instructions say which Tavi tool carries out each step. This guide stays at the level of the conversation.
• “Map all” loads the rest of a cluster beyond its first sample. It is where most of the cost is, so it comes last.
• When a user asks for a “deep search” to find missing companies (step 3), they want more research into companies and pools, not a paid deep search for one role brief. Confirm with the user before starting one of those.
• The guide’s advice on credits is for the user. Follow the server’s instructions on when to mention costs.
• A shortlist in this guide is a list in a project. Pipelines are saved deep searches, not shortlists.

Reading the “Say this to Tavi” prompts

Each step ends with things recruiters say at that point. They are written from the user’s side of the chat, so read them as requests you will receive:

• “Tavi” means you. Users talk to their assistant as Tavi. “Tavi, let’s build a smart sort…” is a request for you to do it with the Tavi tools.
• Treat each prompt as the intent behind it, not a script. “Let’s map all in the top three clusters first” means: agree the clusters and how many profiles to load, then load them.
• Offer the next habit when the user hasn’t asked. New users won’t know these prompts. Study the company at kickoff, show who a cut would remove before applying it, check which clusters are worth loading in full, offer an honest score out of 10 before the user shares the map, and offer insights and client questions at the end. Suggest the next one that fits rather than all at once.
• Be honest when scoring. The step 9 habit only works if the score names real gaps. Don’t round a 7 up to a 9.
• Some requests need tools Tavi doesn’t have: researching a company, reading LinkedIn profiles, building a spreadsheet or drafting messages. Use your own tools for them if you have them and the user agrees. If you don’t, say so plainly.
• Square brackets are placeholders, e.g. [client], [reason] or [list].
• Talk in the guide’s terms: clusters, map all, smart sorts and shortlists, not tool names or IDs.

Before you start: five things to know

The 12 steps at a glance

The steps fall into four phases: set up (1–2), build and refine (3–5), rank and review (6–9), and act and keep it live (10–12).

Step 1 · Kick off

Once Tavi is connected to your assistant (Claude, ChatGPT or another app), start a new chat and tell Tavi you want to build a map. Share the JD as a link, a PDF, or even rough notes. Then pass along anything else you know.

What to share at kickoff

No clear view on companies or titles yet? Leave it open. Tavi will propose them, and you confirm before it maps.

Study the company

Always ask Tavi to study the company you’re hiring for. What do they do? Who have they hired, and where did those people come from? Which companies keep feeding their team? The answers shape the companies you map, the titles you match and the bar you hold candidates to.

Get the companies and titles right from the start

Two things decide who ends up in the map: which companies you search, and which titles you match.

• Companies: go beyond the obvious names. Look at the client’s own team’s past employers, direct competitors, and companies selling to the same customer. Match the client’s business model, not just the industry: for a chain of small, leased urban fitness studios, big-box and franchise operators were the wrong pool.
• Titles: expand, clarify, then tighten.
  • Expand: most roles are described several ways. A deployed engineer is also a forward-deployed, solutions or implementation engineer. Add keyword forms and local-language titles.
  • Clarify: a title is a claim, not a job description. A “GTM Lead” was really a hands-on closer.
  • Tighten: use exact phrases and exclude off-function words like Sales, Analytics or Support.

Step 2 · Shape the map clusters

What Tavi comes back with

What a cluster is

Think of clusters as segments that candidates can be sourced from. Build them in layers:

Add the level layer when you have to decide which level to hire for. The talent density in each cluster often shapes the answer: in one search, the Head of pool was 1.7x deeper than the VP pool, so the client advertised the role at Head of.

What makes a good set of clusters

• One person, one cluster. No one should appear twice.
• Plain names. Every cluster name should tell you instantly who’s in it, e.g. “Head of – NYC restaurant software”.
• New title variants go into existing clusters. Add a new cluster only when the company type is clearly different.
• Give a dominant employer its own cluster if it floods one, so it doesn’t crowd out everyone else.
• Consolidate when there are too many. Grouping by distance from the client’s world works well: insiders first, then adjacent, then any-industry builders.
• Improve the map you have. Don’t start a v2.

Step 3 · Review the first build and fill the gaps

Tavi returns an initial build of the talent map, along with notes and questions for you. Study the map, read the notes and respond to the questions. Then push Tavi to fill the gaps. It will kick off deeper research and come back with proposals to fill them.

Where to look for more pools

Also look at cluster sizes. A cluster that’s much smaller than you’d expect is probably too tightly defined. A huge one probably needs focusing (step 4).

Step 4 · Refine and shape the map

If you’re seeing thousands of profiles, that’s okay. The aim is to reduce each cluster until it accurately reflects candidates who match the brief. Ask Tavi how to reduce and focus the pool, review its suggestions, and approve the ones that make sense.

This might take one round or a couple. It depends on how clearly you see the role, and how open you’d like to keep things. Work with Tavi until each cluster is the size you want, e.g. from 5,000 people to a focused 2,500.

Levers Tavi will typically suggest

Exclusions to apply everywhere

• The client’s current staff, across every cluster.
• Alumni, if the client doesn’t want rehires.
• Blacklisted employers, applied to current employer only.
• Your other clients, as source companies.
• Duplicates, so each person sits in one cluster.

Don’t filter blind. If you’re not sure what a cut would remove, ask Tavi to show you who it would drop first.

Step 5 · Map all profiles

Before you map all, check that you’ve:

• Locked in the clusters, including any new ones you added along the way
• Pushed Tavi to fill the gaps in companies, geographies and titles
• Shaped and refined each cluster to cut the noise and focus on the profiles that matter

What “map all” does

Up to now, each cluster has shown a sample of the profiles that exist. When you click “map all”, Tavi accesses all the profiles in each cluster. You choose the scope:

It may take a couple of minutes.

Step 6 · Rank profiles with smart sorts

You can rank profiles in whatever way you wish. A good starting point is how well each person matches the JD. From there, add what matters most for this role: how likely you are to land someone, or direct experience of the context you’re hiring for. Each smart sort reshuffles the ranking within every cluster. It’s a powerful way to cut and rank a large pool.

Smart sorts that work well

Keep sorts honest

• Use the same names on every map, e.g. “JD fit” and “JD fit + gettability”, so everyone knows what they’re looking at.
• Don’t reward brand names. Ask Tavi to leave employer reputation and school prestige out of the wording, or famous names float up regardless of fit.
• Measure tenure at the company, not the role. A recent promotion shouldn’t count against someone.

Step 7 · Review the rankings, share notes, ask questions

This is where your judgement goes into the map. Tavi turns each piece of feedback into a rule, re-runs the sorts, and the next pass is closer to how you’d rank people yourself.

What works

• Say what’s right, not just what’s wrong. Pointing to someone who’s exactly right tells Tavi what to look for.
• Ask why. “Why is this person ranked so high?” usually reveals a missing factor, or a keyword match that isn’t real evidence.
• Name the pattern, not just the person. “Too much enterprise experience” or “only one month in the role” each become a rule that catches everyone like them.
• Look across clusters, not just within them. If the strongest people keep coming from one cluster, ask for more companies like those. If a cluster rarely produces anyone good, cut or merge it.
• Ask for a line or two on each of the top 10. It’s a quick way to see where a sort is drifting.

Step 8 · Build and shape shortlists

Just tell Tavi in the chat which shortlists you want, and what to call them. Tavi builds an overall list of the best candidates from across the clusters. Once a shortlist is made, review it and keep giving Tavi feedback on the rankings and the order. It all helps with deeper refinement.

What works

• Fewer lists, named exactly. One main list is often enough. Split by level or location only when the client’s question needs it, e.g. Top 100 London vs Top 100 rest of UK.
• No quotas per cluster. Take the best across all clusters, and skip duplicates.
• Keep the reviewed order. Once people have been checked, the list should reflect that, not just the raw sort.
• Flag before outreach: the client’s customers, competitors, and anyone who’d need a visa.
• Don’t pad. If only 62 people truly fit, ship 62.
• Archive old lists, so only the current ones show in the map.

Step 9 · Score the map and refine

Your assistant and Tavi are honest about the state of a map, including what’s weak. So check yourself against what they think. Maps often start around 6 or 7 out of 10. The fixes that take them to a 9 are usually concrete:

• New clusters for employers that were missed
• Profile reads that cut false positives
• Merging or rebuilding thin clusters
• A signal the first sort missed, e.g. people who built US teams for European scale-ups
• A final check of the top of each shortlist against the client’s latest criteria

Ask at every stage, not just at the end: after the first build, after the gap-filling, and before you share anything.

Bring in client feedback

The best rankings come after the client reacts to real people. Paste their feedback in raw, then tell Tavi which point matters most.

• Turn new criteria into a sort. A CEO’s “hypergrowth plus complex implementations” became the main ranking for one search.
• Rank, don’t filter. A signal like fast growth should move people up, not remove people who aren’t on a list.
• Write in the realism, e.g. “they won’t attract people from major AI labs” or “they want a startup vibe”.
• Keep unknowns open. Don’t let Tavi assume an answer the client hasn’t given. Save the question for the next meeting.

Step 10 · Capture map insights and learnings

The map is the evidence. What the client remembers is the answer it gives, e.g. “Head of had 1.7x more strong fits than VP overall, and 2.1x outside NYC, so advertise the role as Head of.”

What to ask for

• 10 core insights + 10 questions for the client, before every client meeting. Good questions also raise the map’s score: the client’s answers often close the last gaps.
• An answer to the question from step 1, with exact numbers in a table, e.g. VP vs Head of, or pool sizes by country.
• A “companies you may not know” list, for clients who want to learn the market.
• Five short points on how the map was built, for the meeting.

Pressure-test every insight: does it deserve its place, and is it about this client?

Package it in a workbook

Assistants that can create files, such as Claude and ChatGPT, can put the whole map into one clean Excel workbook. Settle on a template once, then point your assistant at your last good one.

• Mirror the map. The order in the workbook should match the order in Tavi.
• Put the LinkedIn link right next to each name.
• Build it once, at the end. Finish the reviews and re-ranks first.

Step 11 · Kick off outreach

11A · Email sequences

• Get the list out of the map. Your assistant can put the shortlist into a spreadsheet, with each person’s cluster, LinkedIn link and notes, ready to import.
• Connect your ATS. An integration with your ATS keeps candidates tracked in one place and lets you kick off sequences from there.
• Personalise from the map. Each person’s cluster and profile notes tell you why they’re a fit, which is the best opening line you’ll get.

11B · LinkedIn outreach

An assistant that can control a browser can help with LinkedIn outreach, but use it with caution.

Step 12 · Keep the map live

Signals and map intelligence

Tavi watches the map for you. It notifies you about meaningful shifts and updates in the map on a daily basis, by email, by text or in the map itself, whichever you prefer.

Ways to keep using the map

• Revisit the shortlist as people move, so it stays current.
• Re-run the smart sorts whenever new profiles come in, so everyone is ranked.
• Reuse it for the next role. The company study, clusters and sorts carry over to the next search for the same client or team.
• Use it as talent intelligence, not just for hiring: where the talent pools are, who’s growing, and which companies feed which.

Beyond LinkedIn: GitHub, research hubs and other sources

Some of the best candidates have thin LinkedIn profiles, or none at all. For the right roles, ask your assistant to look where their work actually lives, then bring what it finds back into the map.

People with a LinkedIn profile go into the map as their own cluster, e.g. “GitHub contributors”, so they’re ranked with everyone else. People without one go on a separate sheet in the workbook.

Adapting the steps by role type

The 12 steps hold for every search. What changes is where the talent lives, what makes someone gettable, and what to watch for.