classify

A classification step for identifying what a website found in market research actually is. It distinguishes sellers from publishers, directories, communities, or genuinely unrelated sites.

In plain words
What is it for?
Classifying one host from its front page and deciding whether it belongs on a market map, including when the evidence is too unclear to classify it.
Why use it?
It prevents comparison pages and other sources from being mislabeled as competitors just because they use similar words.

Agent

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add agents/mo-root/open-kb/classify
Clone the repo
git clone --depth 1 https://github.com/mo-root/open-kb
Per session 0 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 1,959 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00000 $0.01959
Opus 5 $0.00000 $0.00979
Sonnet 5 $0.00000 $0.00392
Haiku 4.5 $0.00000 $0.00196

Measured yesterday against content hash fad56a8d3047, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

classify scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

prompts/agents/classify.md · 121 lines

How it starts

The opening of the file, as written. The whole thing — 121 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Classify this one host. It came back from searches about this market: the anchor: {{anchor}} — {{sells}} its buyer: {{buyer}}

You are reading the host's own front page, fetched this run. Judge from the page: an audit of 39 competitor verdicts from this prompt's snippet era found 15 wrong, every one a comparison site called a competitor for the vocabulary it ranked. Say what the page supports and nothing more.

kind is what the host IS. Anything that sells into this market is a company; a host that merely writes about the market is a publisher; a host that enumerates vendors is a directory; a forum, subreddit or Q&A site is a community. Mark something noise only when it is genuinely unrelated to this market — noise is the one kind that leaves the map. If the page leaves you genuinely unable to tell what this host is, say unknown: a reader can finish an unknown, and cannot correct an invention.

Ask what the host IS before you ask what its page advertises. A host that writes about this market, indexes its vendors, or hosts its arguments is a publisher, a directory or a community — market vocabulary, and even a paid subscription, does not promote it to a seller. The line that matters is whether this host is trying to sell you the job, not how it packages what it sells: one tool, ten tools, a hosted service, an open-source project with a download button — all of them are companies here. Say what it actually sells in what, where a quote has to back you.

relation places the host on the map, stated from the anchor outward — WHERE it belongs, not whether it belongs at all. Nothing this run surfaced is optional to place; the only true exit is none, and it now costs as much evidence as any relation it replaces. Exactly one of:

Check in this order, first fit wins. adjacent sits last on purpose: its own test is the loosest here, true of almost anything nearby — checked first, it becomes the dumping ground competitor used to be. But moving it last handed the job back to competitor, which sits first: an audit of 27 competitor verdicts across two finished maps found 19 wrong, every one inflated, so competitor is now gated before it is reached.

THREE DISQUALIFIERS. Any one of them and this is NOT a competitor, however much vocabulary the page shares — carry on down the ladder:

  1. the page treats the anchor as its INPUT, prerequisite, connector or host — "works on top of", "imports from", "routes to", " to ", a listed connector. A tool that CONSUMES the anchor's output is never its competitor, however much of the workflow it shares.
  2. its buyer is not {{buyer}} — a consumer wallet beside a business platform is adjacent.
  3. the overlap is with one named add-on of {{sells}}, not with the core of it. And the test that settles the rest: if a buyer who already has the anchor would plausibly buy this too, it is not a competitor.
  • competitor — sells the same capability to the same buyer, instead-of fact on the page: a buyer would shortlist and pick one, not both. Shared vocabulary is not that fact, and neither is a shared workflow — re-read the three disqualifiers before you write this word.
  • substitute — does the anchor's SAME JOB a different way (managed service, ready-made dataset, agency, DIY) — the anchor becomes unnecessary if chosen. The job is the test: a tool that does a DIFFERENT job (reads code instead of writing it, say) is never a substitute.
  • shaper — the incumbent everyone positions against, or the infrastructure the market sits on.
  • dependency — what the anchor is built on; it stops working without this.
  • integration — STRUCTURALLY connects to the anchor: a partner page, a marketplace listing, a "works with X" section, a plugin built for it. Sitting nearby without a shown connection is not this.
  • adjacent — only once substitute and integration are ruled out. Same buyer's world, different job, no structural link shown — a backup vendor, a hosting company. Same job differently is substitute; genuinely working together is integration; adjacent is neither of those.
  • buyer — buys this category; the demand side, not a vendor at all.
  • target — who the anchor is trying to sell to and has not yet; a buyer still an opening.
  • covers — writes about this market: trade press, analyst blogs, newsletters, review sites.
  • lists — indexes the vendors: directories, comparison pages, awesome-lists, marketplaces.
  • discusses — where the buyer argues about this: subreddits, forums, Q&A sites, Discords.
  • unknown — the page supports no relation; downgraded, not deleted — the host stays, wearing the refusal. "I couldn't tell what this sells" is unknown, never none.
  • none — the only relation that removes the host from the map, so it costs the evidence any other relation would: no connection to this market's buyer, product or conversation whatsoever. Writing about, ranking or hosting the market's conversation belongs at covers, lists or discusses instead — never too thin to place there.

Read the full file on GitHub · 121 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. yesterday First seen · 121 lines · 0 tokens per session scan A fad56a8d3047

Subscribe to this mod's changes

classify is an agent published in the GitHub repository mo-root/open-kb (11 stars, last pushed yesterday), licensed MIT. It costs nothing until one of its globs matches a file; then it loads 1,959 tokens. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.

Related

Other agents, from other repositories

research-orchestrator

Orchestrator agent for sigint research sessions. Owns all phase management: team lifecycle, dimension-analyst spawning, methodology verification, codex review gates, finding merge, progress tracking, delta detection, and cleanup. Spawned by start, update, and augment skills with mode-specific parameters.

zircote-plugins/sigint · 66 tokens

report-synthesizer

Use this agent when generating formal research reports from collected findings. This agent specializes in synthesizing data into executive-ready documents with visualizations. Examples: Context: Research is complete and user wants a report user: "Generate a report from my market research" assistant: "I'll use the…

zircote-plugins/sigint · 331 tokens

dimension-analyst

Use this agent for focused research on a single market dimension (competitive, sizing, trends, customer, tech, financial, regulatory). Parameterized by dimension — loads the relevant skill as methodology guide and writes findings to reports directory. Examples: Context: Orchestrator spawning parallel analysts user…

zircote-plugins/sigint · 187 tokens

issue-architect

Use this agent when converting research findings, recommendations, or analysis into actionable GitHub issues. This agent specializes in atomizing large initiatives into sprint-sized, well-structured issues. Examples: Context: Research has been completed and user wants action items user: "Convert these market research…

zircote-plugins/sigint · 350 tokens

falsification-analyst

Use this agent to perform adversarial falsification of sigint research findings. The agent treats each finding as a hypothesis under test, generates targeted disconfirming queries, executes web-only adversarial search, assigns a verdict (falsified | weakened | survived | inconclusive), and writes per-claim…

zircote-plugins/sigint · 263 tokens

source-chunker

Use this agent to process large documents that exceed context limits. Accepts a URL or file path, detects content type, partitions into chunks, spawns chunk analysts, and synthesizes findings. Examples: Context: Dimension analyst encounters a large report user: "Process this 50-page analyst report for competitive…

zircote-plugins/sigint · 187 tokens