Agent skills for the non-program half of shipping a product on Solana — the things a small team hits around the code.
The all-in-one support skill pack for startups building on Solana. Program work — Anchor patterns, CU optimisation, account layout — is covered well by solanabr/ai-kit and the packs it pins. This repo owns the startup-operations layer: positioning and pricing, launch, fundraising, go-to-market and growth, support and operations, and compliance-adjacent operations (pointing at solanabr/crypto-legal-skill for the legal lane rather than restating it).
Full charter, the gap analysis behind it, and the design constraints: issue #1.
Two constraints from that charter shape everything here:
- Skills stay off the always-on path. A top-level skill's
descriptionis listed in every session and every subagent, so fan-outs of 100+ subagents multiply that cost. These load on demand, with detail behind links inreferences/. - One
SKILL.mdper skill directory, routing-only descriptions, progressive disclosure.
| Skill | Use it when |
|---|---|
| idea-sprint | Deciding what to build, or stress-testing an idea before any code. Interview → crypto-necessity gate → three candidates → score /15 → go/no-go. Output is a decision, not a brainstorm. |
| pitch-deck | You need slides — demo day, a VC meeting, a grant application, an accelerator form, a hackathon final. Detects the audience, checks the title spine ("but" or "therefore" between slides, a bridge from wedge to vision), writes speaking notes, then drills the hostile questions. |
| pitch-review | The deck, script or pitch video is about to go out. An answer key written first, an anonymized copy, three readers with no context (VC, buyer, judge), a rank against decks that raised, and the objections that repeat. Writes .claude/context/review.md. |
| hackathon | A submission is due. Track choice by crowdedness, a description a judge can skim in 90 seconds, scripts for the two videos Colosseum asks for (a 2–3-minute pitch and a demo of at most 3 minutes), and the grant follow-on when the track doesn't land. |
| build-status | Before a deck, submission, grant or investor update. Records program IDs, upgrade authority, verified-build and audit state, what works today, and traction with the query behind each number into .claude/context/build.md. |
| positioning | Before the deck, site, docs intro or launch post. Writes the one-liner, ICP with disqualifiers, job to be done, three named alternatives with the dimension you beat each on, sourced proof points and words to avoid into .claude/context/positioning.md. |
| pricing | Deciding what an onchain product charges: fee surface, who pays, the floor-cost math (fees, rent, RPC, oracles), comparables pulled from public fee data, and the test that would prove the number wrong. Writes .claude/context/pricing.md. |
| market-sizing | Before the market slide, or when choosing where to enter. Ranks entry wedges by problem quality × how much each builds the vision, sizes the first one bottom-up with a source on every number, checks it against a named top-down figure, and keeps money already spent apart from money at risk. Writes .claude/context/market.md. |
| fundraising | After the deck lands: a crypto data room index (program IDs, authorities, audits vs deployed commit, onchain metrics with their queries), a diligence question bank with honest answers, monthly investor updates from the same metrics, and token-vs-equity framing routed to crypto-legal-skill. Writes .claude/context/data-room.md. |
| launch | Mainnet goes live on a date and the announcement has to wait for a transaction. Writes a runbook gated on the onchain event: legal gates via crypto-legal-skill, slip and abort rules, channel order, the Superteam and partner layer, and pre-staged copy including the "it slipped" message. Also runs as /launch. |
| incident-comms | Something is broken and users are asking. Severity set by onchain state, the three pre-written updates, an exploit track routed to crypto-legal-skill, a static HTML status page that stays up when the app is down, a user-facing postmortem, and a known-issue entry community-moderation can dedupe against. |
| ecosystem-bd | You want Jupiter, the wallets, the explorers, DefiLlama or a Superteam chapter to list, route or surface you. A dated map of what each surface requires, the readiness gaps to close in ai-kit first, diligence in both directions (inbound via counterparty-gate), the partner ask and the charter. |
They chain through files in the project's .claude/context/. Each skill writes one file, in a
format its own references/ defines, and reads the others when they exist:
| File | Written by | Read by |
|---|---|---|
idea.md |
idea-sprint |
build-status, positioning, pricing, market-sizing, pitch-deck, pitch-review, hackathon, fundraising, launch, ecosystem-bd |
build.md |
build-status |
idea-sprint, positioning, pricing, market-sizing, pitch-deck, pitch-review, hackathon, fundraising |
positioning.md |
positioning |
idea-sprint, pricing, market-sizing, pitch-deck, pitch-review, hackathon, fundraising, launch, incident-comms |
pricing.md |
pricing |
market-sizing, pitch-deck, pitch-review, fundraising |
market.md |
market-sizing |
pitch-deck, fundraising |
data-room.md |
fundraising |
— |
review.md |
pitch-review |
— |
launch.md |
launch |
incident-comms, ecosystem-bd |
incidents.md |
incident-comms |
— |
partners.md |
ecosystem-bd |
— |
Run the producers first and the later skills get shorter: they pre-fill instead of
re-interviewing you, and a deck, a data room and a launch post all quote the same numbers.
build.md's format is a contract other tools can write to as well:
build-md-format.md.
Rendered decks and any graphic or marketing asset come out as HTML — one file, your own CSS. It renders anywhere, diffs in git, and the agent can design it directly.
Copy the skills into your project's skills directory. Copy the whole set: the skills link to
each other's format files (pitch-deck to build-status's and pricing's, for example), so a
skill copied on its own keeps working but has dead links where it hands off.
git clone https://github.com/solanabr/startup-builder-skill
cp -r startup-builder-skill/skills/* .claude/skills/
Use .agents/skills/ instead of .claude/skills/ for Codex, Cursor, Copilot and other
Agent Skills clients.
The layout is skills/<name>/SKILL.md at the repository root, which is what an
upstream-folder fetch expects — so a kit can pin individual skills from here by name without
vendoring the whole repo as a submodule.
The interview, scoring and validation rubrics, the pitch structure, frameworks, templates and
deck shell, and the hackathon submission guide and video scripts are vendored under each
skill's references/, rewritten from sendaifun/solana-new
with attribution. The hackathon judging criteria and winners index are written from
Colosseum's own FAQ and announcement posts, linked inline.
What remains upstream is small and optional: the Superteam ideas dataset (one pinned JSON
file, 240 entries), ColosseumOrg/colosseum-copilot
(5,400+ past submissions, for crowdedness checks) and Anthropic's
frontend-design. Links are
pinned or point at stable pages, so nothing has to be installed for the skills to work.
Each skill's references/upstream-packs.md gives the details.
Text adapted from solana-new, beingsmit/technical-product-gtm and jonathimer/devmarketing-skills keeps its MIT notice in THIRD_PARTY_NOTICES.md.
python3 .github/scripts/validate.py (standard library only) checks what review used to:
- skill frontmatter
- relative links and anchors
- GitHub links pinned to a commit
- the charter's hazard rules (#1): only
.md,.htmlandLICENSE, no.mcp.jsonor settings, no download piped into a shell, no telemetry - attribution against
THIRD_PARTY_NOTICES.md - README coverage of every skill and context file
CI runs it on every pull request, and a weekly job reports dead external links. A file other than Markdown or HTML needs an entry with a reason in the script's ALLOWED_OTHER_FILES.