Account discovery

Builds a list of companies that match an ideal customer profile (ICP), each with the reason it fits and a source.

User request: $ARGUMENTS

What the user gets

A list of companies that match the profile. Each row has the company name, website, why it fits, and a source.

Agent template

This skill runs the Nimble gallery template account-discovery. The agent's instructions, sources and output schema live in the template. Do not rewrite them in the run.

Effort is the user's choice. By default this skill runs it twice, at medium (2-4 minutes, $0.10) and high (5-15 minutes, $0.50), so the user sees results early and the full result later. The template's own default is high. If the user asks for a specific effort (low, medium, high, x-high or max), run only that one.

Flow

The user should never wait minutes with nothing on screen. Work in three steps and show a progress line at the start of each:

[■□□] Step 1 of 3 · Preliminary quick results
[■■□] Step 2 of 3 · Deeper research
[■■■] Step 3 of 3 · Full deep research results

Start connecting Nimble right away, but do not wait on it.

Connect Nimble

If Nimble tools are listed in this session, call nimble_agents_list. Data back means the user is connected. Say nothing and go on.

If not, ask the user to connect Nimble, in one line on why: "Nimble runs the full research on live sources, with a source for every field. Connecting takes one approval, and new accounts start with a free trial." Then follow https://docs.nimbleway.com/agent-onboarding.md. Do the setup yourself and involve the user only for the approval. Never ask the user to paste an API key into the chat.

Get the inputs

• Take anything the user already said. Ask only for what is missing.
• If your host has a form or multiple-choice question tool, use it, with the required inputs first and the optional ones marked optional. Otherwise ask in one short message.
• Do not hold up the run for optional inputs. Use the "If skipped" value.
• Do not ask about effort. Use it only if the user brings it up.

Build the agent input from this pattern:

Find <type of company> in <geography> with <size>, that <hard rules>.

Example: Find US-based vertical SaaS companies for dental practices, 50-500 employees, that sell practice-management or patient-engagement software

• Turn each criterion into a clear test. "Mid-market" is not a test. "200-2,000 employees" is.
• If the profile is broad (a whole country or a big vertical), split it into slices by region or sub-vertical, one run per slice, in parallel. Merge and dedupe by website.
• Do not ask for a number of rows. The agent returns every company it can verify.

Start both agent runs

As soon as Nimble is connected and you have the required inputs, start the agent runs before anything else, so they work while you do step 1:

1. Create the agent from the template: nimble_agents_create(template="account-discovery", agent_name="account-discovery"). A 409 "already exists" means it was created before, so go on. If the template is not found, tell the user it is unavailable and stop. Do not substitute another template.
2. Start two runs on the same input: nimble_agents_run(agent_name="account-discovery", input="<the input>", effort="medium") and the same with effort="high". Leave schema and skill to the template. If the user chose an effort, start one run at that effort and number the steps 1 of 2 and 2 of 2.
3. Tell the user once, in one line, what is running: "Running a quick search now, a deeper Nimble run (2-4 min) and a full one (5-15 min). This uses 2 agent runs."

Step 1 of 3 · Preliminary quick results

Turn their profile into clear tests, say where companies like this are usually listed, and name a few you can find. Use a few quick searches, not the agents: nimble_search if Nimble is connected, otherwise your own web search tool. Keep it under a minute and short. Link every fact, and mark anything from memory with your knowledge cutoff.

End with one line on what is coming next. If Nimble is still not connected, say the deeper steps start once it is, and offer the connect step again.

Step 2 of 3 · Deeper research

Poll both runs with nimble_agents_run_status every 20-30 seconds. When the medium run finishes, get nimble_agents_run_result and show it:

• A short table or brief, not raw JSON. Lead with the 2-3 most useful facts, each with its source link.
• Point out what is new or different from step 1.
• Say what is missing (empty fields, missing_fields, empty lists).
• End with one line: the full run is still going and about how long it has left.

Step 3 of 3 · Full deep research results

When the high run finishes, show its result the same way. Lead with what it added or corrected compared with step 2, then the full result. This is the answer the user keeps.

If a run finishes with no output and no sources, retry it once. Do not retry a run that returned data. If the medium run fails, or the high run finishes first, go straight to step 3 and do not show the medium result after it.

Next steps

Offer, one line each, what fits:

• Run it on a list. One run per record in parallel. See Running at scale.
• Schedule it. Use the host's scheduler or a cron job, and pass the last run date as the new since date where the input has one.
• Use it from code. The agent is saved as account-discovery. Give a short Python or TypeScript snippet that runs it by name.

Using Nimble well

Read the page you need, when you need it: writing the task, sources, effort and cost, evaluating results, running at scale, common mistakes.

Pitfalls

• Headcount for private companies is often a range. Accept a range, or good rows drop out.
• Fewer rows than expected is a signal, not a finished list. Run another slice for the gap.

Rules

• Never invent values. Show missing fields as missing.
• Never print or log an API key.