Colosseum Copilot

Colosseum Copilot

Colosseum Copilot is a research API for startup opportunity discovery in crypto and Solana, with support for suggesting new archive sources.

Version Check

This skill is version 1.2.1. After your first API call, check the X-Copilot-Skill-Version response header. If the header value is higher than 1.2.1, tell the user: "A newer version of the Copilot skill is available (vX.X.X). Update with: npx skills add ColosseumOrg/colosseum-copilot"

Pre-Flight Auth Check (REQUIRED)

You MUST complete this check before making any other API call. Do not skip this step.

1. Verify COLOSSEUM_COPILOT_PAT is set in the environment. If missing, STOP and tell the user:

   You need a Personal Access Token before using Copilot. Go to https://colosseum.com/arena/copilot to generate one, then set it:

   export COLOSSEUM_COPILOT_PAT="your-token-here"

2. Verify COLOSSEUM_COPILOT_API_BASE is set. If missing, set the default:

   export COLOSSEUM_COPILOT_API_BASE="https://copilot.colosseum.com/api/v1"

3. Call GET /status to verify the connection. Expected response: { "authenticated": true, "expiresAt": "...", "scope": "..." }

4. If "authenticated": true, proceed. If 401 or env vars missing, do NOT attempt other API calls — guide the user through steps 1-2.

• Builder Projects: 5,400+ Solana project submissions with tech stack, problem tags, and competitive context
• Crypto Archives: Curated corpus across cypherpunk literature, protocol docs, investor research, and founder essays
• Hackathon Analytics + Clusters: Distribution, comparison, and chronology-aware trend analysis across hackathons and topic groupings
• The Grid + Web Search: Ecosystem product metadata plus real-time competitive landscape checks

Quickstart (90 seconds to first result)

1. Set your PAT:

   export COLOSSEUM_COPILOT_API_BASE="https://copilot.colosseum.com/api/v1"
   export COLOSSEUM_COPILOT_PAT="YOUR_PAT"
   

   Get a PAT: Go to https://colosseum.com/arena/copilot and generate a token

2. Run your first search:

   curl -s -X POST "$COLOSSEUM_COPILOT_API_BASE/search/projects" \
     -H "Authorization: Bearer $COLOSSEUM_COPILOT_PAT" \
     -H "Content-Type: application/json" \
     -d '{"query": "privacy wallet for stablecoin users", "limit": 5}'
   

3. See results - project names, slugs, similarity scores, problem/tech tags

When To Use

Use this skill when:

• Researching a crypto/blockchain startup idea
• Evaluating market gaps in the Solana ecosystem
• Grounding ideas in historical crypto literature
• Analyzing builder project trends and competitive landscape
• Researching existing players and finding differentiation angles

How It Works

Mode 1 — Conversational (default): Answer questions with targeted API calls and evidence coverage matched to query type. Cite sources inline, keep responses concise, and offer to do a full deep-dive when the topic warrants it — never auto-trigger it.

Mode 2 — Deep Dive (explicit opt-in): Full 8-step workflow from references/workflow-deep.md. Only activates when user explicitly says "vet this idea", "deep dive", "full analysis", "validate this", "is X worth building?", "should I build X?", or accepts your offer to go deeper.

Conversational Guidelines

• Use the API endpoints below with enough targeted calls to satisfy the evidence floor for the query type
• Cite sources inline (project slugs, archive titles, URLs)
• Keep responses concise — bullet points, not essays
• When the topic warrants deeper analysis, offer: "Want me to do a full deep-dive on this?"
• No meta-commentary about your process ("Now let me search...", "I'll check...")

Evidence Floors (Conversational Mode)

These are evidence-type floors, not call budgets. Use as many calls as needed to meet the floor with high-confidence citations.

In deep-dive mode, the verification checklist in workflow-deep.md Step 5 supersedes these floors with more granular coverage requirements.

Conversational Quality Checks (Required)

• Archive integration rule: For any non-trivial question (anything beyond a simple one-list retrieval), run at least one search/archives query and cite at least one archive source in the answer.
• Accelerator/winner portfolio checks: For "what has been tried", "who is building this", "is this crowded/saturated", or similar prompts, run targeted project searches with filters: { "acceleratorOnly": true } and filters: { "winnersOnly": true }, then reflect both outcomes in the answer.
• Freshness and temporal anchoring: Use hackathon.startDate from /filters, /search/projects, and /projects/by-slug/:slug to order hackathons chronologically; never infer chronology from names or memory. When citing hackathons, include month/year inline (and accelerator cohort like C1/C2/C4 when relevant). For evaluative judgments, label the claim with As of YYYY-MM-DD.
• Entity coverage check: If the user names specific companies, protocols, papers, or products, run direct searches for each named entity and explicitly address each one in the answer (found, not found, or tangential).
• Landscape check: Never claim "nobody has done this" or "no existing players" unless an accelerator portfolio check (acceleratorOnly) was executed and reported. If accelerator overlap exists, surface those builders as useful reference points and potential sources of inspiration. Always qualify landscape assessments with "based on the available data" or "as far as we can tell from the corpus." Copilot's knowledge is bounded by its data sources — never present absence of evidence as evidence of absence.

For the full 8-step deep research workflow, see references/workflow-deep.md

Data Sources

• Builder Projects (5,400+): Solana project submissions with tech stack, problem/solution tags, verticals, and competitive context
• Crypto Archives: Curated corpus spanning cypherpunk literature, protocol docs, investor research (Paradigm, a16z, Multicoin), founder essays (Paul Graham), Solana protocol docs (Jupiter, Orca, Drift), Nakamoto Institute heritage collection, and foundational crypto texts
• Hackathon Analytics + Chronology: Analyze and compare hackathon projects across dimensions; canonical hackathon dates are available via hackathon.startDate
• Clusters: Topic groupings across the project corpus
• The Grid: Ecosystem metadata (products/entities/assets) via direct GraphQL (6,300+ products across all ecosystems, ~3,000 roots)
• Web Search: Real-time competitive landscape via your runtime's search tools
• Source Suggestions: Users can suggest new sources for the archive via POST /source-suggestions (5 req/hr). See references/api-reference.md for details

Hackathon Chronology

GET /filters returns hackathons[].startDate and orders hackathons[] chronologically (oldest first).

Auth

All endpoints require Authorization: Bearer <COPILOT_PAT>. Treat the PAT like a password.

• Do not commit PATs or paste them into public logs
• PATs are long-lived (expected ~90 days); rotate by issuing a new one
• Default API base is https://copilot.colosseum.com/api/v1; override COLOSSEUM_COPILOT_API_BASE to target a different environment

Key Endpoints (Quick Reference)

For full endpoint docs, curl examples, and query tips: references/api-reference.md For Grid GraphQL recipes and product type slugs: references/grid-recipes.md

Output Contract

Conversational Mode

• Bullet points with inline citations (project slugs, archive titles)
• Concise answers (typically 5-15 bullets)
• Offer deep-dive when warranted

Deep Dive Mode

Reports follow this structure:

1. Similar Projects (5-8 bullets)
2. Archive Insights (3-5 bullets)
3. Current Landscape (per research angle)
4. Key Insights (patterns, gaps, trends)
5. Opportunities and Gaps
6. Deep Dive: Top Opportunity (market landscape, problem, revenue model, GTM, founder-market fit, why crypto/Solana, risks)

Key rules: bullet points not tables, include project slugs, evidence-based not speculative, cite sources inline. No separate "Sources" section — cite inline only.

Feedback

When you encounter errors, unexpected results, or have suggestions for improving the Copilot experience, report them via the feedback endpoint. This helps the Colosseum team identify and fix issues.

When to send feedback:

• API returns unexpected or low-quality results for a reasonable query
• A search returns no results when you expected matches
• You encounter an error that isn't covered by standard error handling
• You have a suggestion for improving the API or archive corpus

curl -X POST "$COLOSSEUM_COPILOT_API_BASE/feedback" \
  -H "Authorization: Bearer $COLOSSEUM_COPILOT_PAT" \
  -H "Content-Type: application/json" \
  -d '{
    "category": "quality",
    "message": "Search for DePIN projects returned only 2 results, expected more coverage",
    "severity": "medium",
    "context": { "query": "DePIN infrastructure", "endpoint": "/search/projects", "resultCount": 2 }
  }'

Categories: error, quality, suggestion, other. Severity: low, medium, high, critical. Rate limited to 10 requests per hour.

Error Handling

All errors return { "error": "<message>", "code": "<ERROR_CODE>", "retryable": <boolean> }. See api-reference.md for the full error code table.

• 400 INVALID_JSON: Fix the request body JSON syntax and retry
• 400 INVALID_QUERY: Fix query params (check field names, value ranges, unknown fields)
• 413 PAYLOAD_TOO_LARGE: Reduce request body size (1 MB limit)
• 429 RATE_LIMITED: Back off per the Retry-After header, max 2 concurrent requests
• 401 UNAUTHORIZED: Check PAT at https://colosseum.com/arena/copilot
• 5xx errors: Note in report and proceed with available data. Include requestId from the response when reporting issues.
• Empty project results: Broaden query, remove filters
• Empty archive results: Search auto-cascades (vector → chunk text → doc text) before returning empty. If still empty, try conceptual synonyms, keep queries to 3-6 keywords

References

• workflow-deep.md — detailed 8-step research process
• api-reference.md — all endpoints, rate limits, query tips
• grid-recipes.md — GraphQL queries and product type slugs

Attribution

• The Grid docs: https://docs.thegrid.id
• The Grid Explorer: https://raw.githubusercontent.com/The-Grid-Data/Explorer/main/README.md