Build Status

Inspect the repo and the chain → record what is true today → write .claude/context/build.md. Every claim gets a source. This file feeds pitch-deck, hackathon and idea-sprint, so a wrong number here ends up on a slide.

Format and merge rules (the contract every producer follows): build-md-format.md.

Context handoff

• At start: read .claude/context/build.md if present and merge into it; read .claude/context/idea.md if present for the product name and the core flow.
• On completion: write .claude/context/build.md. Never delete a section another skill or tool wrote.

Workflow

1. Gather from the repo, not the founder

Ask only for what the repo cannot tell you. Look first:

• Program IDs: declare_id!, Anchor.toml [programs.*], and the files ai-kit's /deploy records (.program-id-devnet, .program-id-mainnet, deployment-mainnet.json with deployer, upgrade authority and commit).
• Tests: run the suite once and record pass/fail with the date. Don't record "passing" from memory or from a CI badge you did not check.
• Audit state: the latest report in the repo, or an /audit-solana findings file. Count open findings by severity.
• Milestones: git log tags and merge commits, dated.

2. Confirm against the chain

For each program ID on mainnet (and devnet, labelled as such):

• solana program show <ID> -u <cluster> → exists, upgrade authority, last deploy slot.
• Verified build: https://verify.osec.io/status/<ID> returns is_verified, the repo and the commit. "Built with solana-verify" is not the same as verified; only the status counts.
• Upgrade authority is a trust fact, not a detail: deployer key, a Squads vault, or none (immutable). Write which. For a vault, take the threshold from the Squads multisig account, not from solana program show, which prints only the address.

3. What works today

List only flows a stranger could run right now, each with its cluster and a proof — a transaction signature, an explorer link or a live URL. Re-check the list on every run and drop anything that no longer works; roadmap never goes here. Optionally note known-broken flows underneath so nobody demos them by accident.

4. Traction, with sources

Mainnet only. For each number: value, window, as-of date, and the query or system it came from. A number without a source is not written. Methods per metric, and the Solana-specific ways these numbers get inflated: traction-sources.md.

Pre-launch is fine — write the waitlist or pilot counts with their source, or write "none yet". Don't let a deck invent what this file says is zero.

5. Write and show the diff

Merge per the format rules, set Updated, and show the user what changed against the previous version. Flag any number that moved more than 2x since the last run and ask why before keeping it.

Hand-offs

• Deploy, verification, audits: ai-kit's /deploy, the verify-PDA upload and solana-verify remote submit-job in its deployment skill (/deploy only verifies locally; the verify.osec.io status flips after the submit-job), /build-program and /audit-solana. This skill records their results; it doesn't redo them.
• Next: pitch-deck or hackathon, which pre-fill traction and "what works" from this file.