Competitor Intel Agent
It researches a company and its rivals, and publishes a one-page brief where every finding cites the page it came from.
Example output: ryanhennebry.xyz/competitor-intel-agent
What the brief contains
- Quick take: three findings the company doesn't already know, each one sourced.
- Recommendations: what to do next quarter, each with a next step.
- What rivals changed: what each rival changed recently, or what it has stayed quiet about.
- Comparison table: terms, pricing and positioning side by side, with an "unknown" wherever a rival's own site leaves the point open.
- Open questions: what the evidence couldn't settle.
Each claim carries a source chip. Hover or tap it to see the evidence and the page it came from. A page that blocked the fetch is labelled as blocked, never guessed at.
How it works
A run is a graph of 69 nodes: 30 Claude Code subagents and 39 deterministic code steps. tools/run.py owns order, dispatch and the gates between nodes. Agents write typed data. Code checks it, and code renders the page. No agent writes HTML.
- Frame: identify the company, build a model of the reader, and set the questions the brief must answer.
- Discover and select: map the market, pick the rivals that matter most, and rank them.
- Extract: fetch each company's own pages (
tools/fetch.py, with a headless-Chrome fallback) and log every claim to a ledger with its URL, a verbatim quote and a state: verified, company-claimed, blocked or absent.
- Lock and shape: commit to a thesis, then allocate the page's word budget to the findings that carry it.
- Author: write the brief as JSON spans, each one citing ledger rows.
- Critique: reader screeners plus red-team, truth and voice critics. An arbiter applies their fixes.
- Verify: re-check every span against its cited rows. The verdict gate passes only when every span passes and every URL on the page appears in a retained snapshot.
- Render and publish:
build/web.py renders a self-contained page (subset fonts, inlined logos, its own share card). tools/publish.py deploys it and probes the live URL.
The rules that govern a run live in AGENTS.md and references/: the runbook, one file per stage, the provenance rules, the page contract and JSON schemas for every artefact. quality-bar/ holds the reader model, the voice card and the banned patterns that the critics enforce.
Run it
This is a snapshot of a working personal tool, not a packaged product. It runs on macOS with Claude Code.
Requirements:
- Claude Code, signed in. Each agent node runs as a
claude subprocess.
- Python 3.13 (the version it was built on) and Google Chrome or Chromium.
python3 -m pip install -r requirements.txt
python3 -m playwright install chromium
bin/intel run https://example.com/a-page-naming-the-company --dry-run
--dry-run prints the node graph. Drop it to run. Runs land in ~/competitor-intel-runs/<company>-<date>/ (override with INTEL_RUNS_DIR), and the rendered page is brief/web/index.html inside that run. bin/intel status shows progress and bin/intel cost shows spend.
Notes:
- Cost: a full run makes many Opus calls. Expect tens of dollars of usage per brief.
- Publishing: the
publish node is wired to the author's site and stops without its checkout. Everything before it runs, so open the local index.html.
- MKT1: the
mkt1 node calls a paid marketing-research service through MCP. It's optional and off the critical path, and the run continues without it.
- Tests: each tool has a
--selftest. Their fixtures come from private runs and aren't published here.
Licence
MIT. See LICENSE.
