How to Write Prompts to Track for Devtool Companies

Direct answer

Write prompts to track from slots. List your buyers' words for each slot (interface, category, user group, coding agent, job, qualifier) and fill them into six sentence types. Then keep only the prompts a real person would type. A good set covers every product line, leaves your brand out, and keeps buyer questions apart from coding-agent tasks.

A prompt to track is a fixed question or task you run on a schedule to see whether answer engines and coding agents mention, recommend or cite your product. A good set is realistic, covers everything you sell, and keeps its wording fixed, so any change in results comes from the agents.

The examples use two kinds of devtool company: a search company like TinyFish, which sells web search, fetch and web agent APIs, and a document parser like Reducto, which turns PDFs and scans into structured data for LLMs. They're illustrations, not those companies' own prompts.

1. Split buyer questions from developer tasks

Buyer questions are what people type into ChatGPT or Perplexity while comparing vendors. Developer questions and tasks are what they type into Claude Code, Codex or Cursor while building. Run them as two sets with two scorecards.

2. List your slots in your buyers' words

A slot is one part of a prompt you can swap. Write down the words buyers use for each:

Add a few general words too ("developer tool", "AI startups") for buyers who don't know your category's vocabulary yet. Pair a user group only with jobs it actually has: an AI engineer asks which search API to use, a marketer at the same company doesn't.

3. Set quotas before you write

Decide how many prompts each job and sentence type gets before writing any, or the easy-to-phrase jobs crowd out the ones you sell. A document parser like Reducto starting with 40 prompts might give 16 to table extraction, 10 to invoices and financial documents, 8 to RAG pipelines and 6 to definitions and alternatives. We think 40 to 60 prompts is enough to start.

4. Write candidates from six sentence types

Fill each sentence type with one to three slots:

Rules for the mix:

• Use specific words in most prompts and general words in about a third.
• Put a qualifier ("cheapest", "2026") on no more than about one in seven.
• Leave your own brand out. The only brand in the set is the incumbent you want to replace, such as AWS Textract for a document parser, in alternatives prompts.
• Write several phrasings per job, because the wording changes which vendor an agent picks. "Live transcription for a voice agent" and "captions for videos" are both speech-to-text questions, but they're different jobs with different likely winners.

5. Keep only prompts a real person would type

Write about three times as many candidates as you need. Score each from 1 to 5 for how likely a real person is to type it, and keep only 4s and 5s. Common reasons to drop one:

• Forced combination. "How do I test whether Codex can use my web scraping MCP server to parse invoices?"
• A vertical that doesn't change the answer. "What is document parsing for fintech companies?"
• Stacked qualifiers. "Cheapest enterprise web search API for startups in 2026"
• A bare acronym. "What is IDP?"

A model makes a fast first scorer, but read the 3s yourself: some are one edit away from a 5.

6. Test-run a sample and cut prompts that drift

Run each new prompt a few times and check that the vendors it surfaces are ones you'd lose a deal to. A prompt can sound like your category and measure someone else's market. "I need to enrich business records with information from the web" sounds like search, but it invites sales-intelligence tools, so a search company like TinyFish would be scored against companies it never competes with. Rewrite drifting prompts to name the job and the interface.

7. Give every developer prompt a stack

Coding agents read the project they're in, so the same question can get a different pick in a Next.js project than in a Python one. Run each developer prompt inside a starter project for every stack your customers use, and report the stacks separately. For a document parser, that might be a Python RAG pipeline and a Next.js app with file upload.

8. Check coverage against every product line

Lay out every product line against buyer questions and developer tasks. An empty cell is an unasked question, and a report will show it as zero visibility. For a search company like TinyFish, the grid has rows for search, fetch and web agents.

A finished sample: document parser

Eight prompts from a set a document parser like Reducto might track:

Buyer questions (ChatGPT, Perplexity)
1. Best PDF parsing API for RAG in 2026
2. AWS Textract alternatives for complex tables
3. How do I extract tables from scanned PDFs?
4. I'm a CTO building a lending app. What should I use to parse bank statements?

Developer tasks (Claude Code, Codex, Cursor, in a Python and a Next.js starter)
5. I'm building a RAG app over PDFs with tables. What should I use to parse them?
6. Extract the line items from these invoice PDFs into JSON.
7. Our OCR results are not good enough on scanned forms. What should I use instead?
8. Add document upload and parsing to this app.

Running and maintaining the set

• Record mention, recommendation and top pick separately. An agent can name you and recommend someone else. Our guide to agent-led growth metrics defines each.
• Count failed and empty runs on their own, so they don't read as "not mentioned".
• Break results down by agent, because agents differ even on whether they can use a tool. On Devtool Arena's October 2026 API boards, AssemblyAI ranks #1 on Gemini and Grok but #5 on Claude Code.
• Freeze wording. To change a prompt, retire it and add a new one, so its history stays clean.
• Review the set quarterly, and when you launch a product line, meet a new competitor or your customers adopt a new coding agent.

For a dozen prompts, a spreadsheet and claude -p "..." (Claude Code CLI reference) or codex exec "..." (Codex non-interactive mode) run inside each starter project is enough. Beyond that, a tracker saves the labelling. Lightsage Agent Discoverability runs tracked prompts daily across 12 coding agents and 7 answer engines, and Devtool Arena's Agent Landscape ranks tools by how often coding agents recommend them.

To see the method applied to one category, read our top prompts to track for search companies. To act on what you find, see how to rank higher in coding agent search and how to see if coding agents pick a competitor.

Frequently asked questions

How do I track what coding agents recommend?

Write developer questions and tasks in your customers' words. Run each one several times in the coding agents they use, such as Claude Code, Codex and Cursor, inside a starter project for each stack. Record mentions, recommendations and top picks separately, and count failed runs on their own.

How do I track what coding agents recommend for my MCP server?

Put the interface in the prompt, for example 'What's the best web scraping MCP server for Claude Code?', and run it in the coding agents your customers use. Track whether the agent recommends your server apart from whether it can use it, because the two often differ.

How can a payments company track what coding agents recommend?

Write developer prompts around the jobs payments customers bring to a coding agent, such as a checkout flow or usage-based billing, and run them in starter projects for your customers' stacks. Add buyer questions such as 'Best payments API for AI agents in 2026' for answer engines, and score the two sets separately.

Our developer marketing team wants to track what coding agents recommend. What tool should we use?

For a dozen prompts, a spreadsheet plus the claude -p and codex exec commands is enough. For a recurring set across several agents and stacks, use a tracker that runs real coding agents. Lightsage, for example, runs tracked prompts on 12 coding agents and 7 answer engines.