Competitor research — the full pipeline

A multi-skill: it orchestrates the whole pipeline and uses three sub-skills on their layers:

• parallel-research — deep-research tasks (briefs, submit, background polling);
• ads-parser — ads from LinkedIn, Google and Meta libraries (and the LinkedIn traps);
• ahrefs — SEO data (quick lookups for a domain).

Keys needed: PARALLEL_API_KEY, SEARCHAPI_KEY, AHREFS_MCP_KEY (see connections/). The pipeline still works without Ahrefs — the SEO layer becomes a gap in the report.

Typical cost of one run: 8–9 Parallel ultra tasks, a few dozen SearchApi calls, a few hundred to a few thousand Ahrefs units, and ~15–20 subagents of model time. Tell the user the expected cost before starting.

Order of work

Phases, agent prompt templates and the folder structure are in references/pipeline.md — read it before starting. In short:

0. Foundation. Check memory and past reports — don't rebuild what is already known. Get the sitemap and cluster the URLs. Try /llms.txt on the docs subdomain.
1. Deep research. Submit 8–9 ultra tasks from references/parallel-briefs.md (placeholder {ANCHOR}) before the collection phase — they run 30–60 minutes on Parallel's side. Start one background poller.
2. Collection. ~15–17 subagents in parallel: site ×6, community ×7, ads, SEO, team ×2. Each writes one file and returns {file, key_findings, gaps}.
3. Synthesis. Three agents: positioning + ICP + sales points; company dossier; battlecard only if explicitly asked.
4. Report. A main report.html plus satellite pages (launches, ads, PR, tutorials, community) linked with card buttons. Paste the style contract from references/report-rules.md verbatim into every writing agent's prompt.
5. Memory. Save where everything lives and the main facts, so the next run starts from them.

Running many agents

• If workflows are available in your Claude Code, run phase 2 as one Workflow with all agents in parallel() and a schema {file, key_findings, gaps}. If it stops on a usage limit, resume it with the same run id — finished agents come back from cache.
• Otherwise launch the subagents with the Agent tool, 5–6 per message, and wait for each batch. Same prompts, same output contract.

Execution rules

• Checkpoints with the user. After collection, show the key findings. The report is iterated on feedback; the first version is never final.
• Raw is separate from analysis. raw/ keeps JSON, HTML and screenshots; every fact in the report is checked against raw.
• The report is about the competitor only. No comparisons with the user's own company and no "what this means for us" sections unless explicitly asked.
• Never log in anywhere. Public pages, public libraries and search results only.
• Publishing. Pages are standalone HTML files. If you publish them as Artifacts, replace relative links between pages with the published URLs.

Adapt it

• Replace the agent list with the layers that matter in your market. A B2C app needs app-store reviews and TikTok more than GitHub and Hacker News.
• Your own report feedback is the most valuable part: after every review, add the corrections to references/report-rules.md. That file is how the next report comes out right the first time.