Competitors

Competitors

Manage the competitive landscape for the founder's project. This skill handles both working with existing competitor data and orchestrating new discovery when needed.

Before you start

Read startup/core.md to load project context (name from frontmatter, seed description, and all fields under ## Core — audience, problem, solution, ICP, geography, etc.). If this file does not exist, this usually means the project has not been started yet, and the idea has not yet benn properly discussed with the user. Although this skill can technically be used without it, a well-thought description of what the user whant to build or discover is paramount for targeted and insightful research. Make the user know the project does not seem to be initilised and propose to do so via /whats-next (they can use this command, or you can refer to this skill yourself).

So, if the file does exist or user insists on going forward without, it, continue with the instructions that follow.

Check if startup/competitors/ exists and contains any .md files.

Scaffold the folder if it doesn't exist yet:

mkdir -p startup/competitors

When competitors already exist

Load and understand them for context.

Before inferring intent — check for thin files.

A file is thin if any of these are true:

• The ## Description body contains *To be filled in.* or is a single short sentence with no real substance
• The file is missing a ## Core Features section entirely
• The file is missing a ## Notes section AND the description is thin

If any thin files are found, surface this before doing anything else:

"I see {N} competitor(s) that haven't been fully researched yet: {names}. Want me to fill those out before we continue?"

• Yes → dispatch a single web-researcher call covering all thin files (same focused prompt as discovery.md Path A). The web-researcher agent is generic, so include the Competitor output format spec from references/discovery.md in the dispatch prompt (the per-competitor fields + direct/indirect + maturity), asking for description, core features, and differentiation notes for each. Save raw output to startup/research/{YYYY-MM-DD}-{slug}-research.md for each, or one combined file if multiple. Write the filled content back to each competitor file. Then proceed with whatever the founder originally asked.
• No → proceed directly with whatever the founder originally asked.

If all existing files are properly filled out, skip this check silently and proceed.

Infer intent from the conversation — don't mechanically ask "what do you want to do?" If the founder is:

• Asking about a specific competitor — load that file, discuss, help update it
• Adding a new competitor — help create a new file following the conventions below
• Updating an existing competitor — read the file, propose changes, get confirmation, write it back
• Asking about the landscape broadly — summarize what exists, grouped by type (direct/indirect), with key differentiators
• Wanting deeper research on a specific competitor — dispatch the web-researcher agent with a focused prompt about that competitor (the agent is generic, so include the Competitor output format spec from references/discovery.md in the prompt), update the file with findings, and save the full web-researcher output to startup/research/{YYYY-MM-DD}-{competitor-slug}-research.md with frontmatter date, topic, and source_skill: competitors
• Wanting to know what users think of a competitor — when the founder asks specifically about user feedback / reviews / what people love or complain about (for one competitor or the whole set), load references/user-feedback.md and follow it. It mines review sites and communities and writes a ## What Users Say section into each competitor file.
• Checking on / refreshing / monitoring existing competitors — when the founder wants to re-check competitors they already have ("check on my competitors", "what's changed since we looked?", "are these still accurate?", "refresh the landscape"), this is a watch pass, not first-time discovery. Load references/watch.md and follow it. It re-checks each existing competitor for changes (features, pricing, pivots, shutdown signals), scans for new entrants, refreshes files in place with a ## Change Log, appends to the rolling digest startup/competitor-watch.md, and re-syncs the landscape map. Requires at least one non-archived competitor — if none exist, route to references/discovery.md instead.
• Archiving — when a competitor is no longer relevant (e.g., after a pivot to a different market), set status: archived and add archived_reason to frontmatter with a one-line explanation. Archived competitors stay in place and can be restored by removing the status or setting it to active

When adding or updating competitor files, follow the file conventions:

• YAML frontmatter with type (direct or indirect), url (competitor's main website), and optionally status (active or archived — defaults to active when absent)
• Optional maturity — one of incumbent (established/large), scaleup (growing, Series A+), startup (early/small), or unknown. Classified during discovery from funding/age/size signals; omit or use unknown when unclear
• Optional last_checked — ISO date (YYYY-MM-DD) of the last competitor-watch pass over this file. Set/bumped by the watch workflow (references/watch.md); absent on files that have never been watched
• H1 heading: the competitor's name
• ## Description — what the company does and who it targets (2-3 sentences)
• ## Core Features — bullet list of notable capabilities
• ## Notes — comparison to the project, differentiation angles, founder comments
• Optional ## What Users Say — machine-generated by the user-feedback workflow (references/user-feedback.md). H3 subsections ### What Users Love, ### Complaints, ### Unmet Needs, ### Misc; only non-empty buckets appear. Not authored by hand
• Optional ## Change Log — machine-generated by the watch workflow (references/watch.md). Append-only dated bullets (- YYYY-MM-DD: shipped X / raised pricing / pivoted / archived — site dead), only added when a watch pass detects a change. Not authored by hand

Slug convention: lowercase the name, replace spaces and non-alphanumeric characters with hyphens, collapse multiple hyphens. "Notion AI" -> notion-ai.

Read before writing, and if you need to make some updates and it might not be oobvious from the context, it's always better running them by the user.

Research guidance

Ad-hoc web search vs. dispatching web-researcher

Not every web question needs a subagent:

• Inline WebSearch / WebFetch (you, the main agent): single-fact lookups ("does Notion have a free tier?", "is ACME still operating?"), quick verification of a claim, one data point asked about in flow. Stays in conversation, no persistence needed.
• Dispatch web-researcher: multi-source passes that benefit from an isolated context (deep dive into a specific competitor's features and positioning, full landscape scan, cross-source verification of a competitive claim). Output is structured and gets saved to startup/research/ for later reference.

Rough rule: one fact in flow → inline. Multi-source or results-should-persist → dispatch.

Model choice when dispatching

When dispatching the web-researcher agent for competitor research, use a fast model unless the founder explicitly asks for a more thorough or higher-quality search. Competitor research is token-intensive due to web fetching, and a lighter model handles it well.

When no competitors exist (or a large-scale competitor discovery or landscape reassessment needed)

If startup/competitors/ is empty or the situation suggests that a profound research is needed, load the reference file:

.claude/skills/competitors/references/discovery.md

The reference file's instructions take over from this point.