What Skillsmith writes and how to run it

Skillsmith emits one markdown file with frontmatter, the same shape any single skill in the hub uses. The body carries three parts: a PREREQUISITE block naming the source skills, a SEQUENCE of numbered steps, and an OUTPUT contract that tells the agent how to report what ran.

There is no new field, no schema extension and no runtime. An agent that can read a skill from the hub can read a build.

---
name: smart-money-entry-check
description: >-
  Reads a smart money buy event, refuses to act until the contract passes an audit and the pool is deep enough, then swaps inside your wallet limit.
  Trigger on: smart money just bought, should I follow this whale, copy trade this signal.
---

# Smart money entry check

## PREREQUISITE

Install and authenticate these skills first:

- `trading-signal`
- `query-token-audit`
- `query-token-info`
- `binance-agentic-wallet`

If any of them is missing, say which one and stop. Do not improvise a replacement.

## SEQUENCE

### Step 1. `trading-signal`

Pull buy events from tracked smart money wallets on BSC for the last 60 minutes.

**Continue only if** max gain since trigger is under 40% and exit rate is under 25%.

**Otherwise:** Report the signal as late and stop.

### Step 2. `query-token-audit`

Audit the contract address carried by the signal.

**Continue only if** no honeypot flag and token risk level is below 3.

**Otherwise:** Report the audit findings verbatim and stop. Do not swap.

### Step 3. `query-token-info`

Read liquidity, holder count and 24h volume for the same address.

**Continue only if** liquidity is above the floor set in the user config.

**Otherwise:** Report thin liquidity and stop.

### Step 4. `binance-agentic-wallet`

Quote the swap, then execute it at or below the configured slippage and daily limit.

## RULES

- Every gate is a stop, not a warning. The agent does not swap on a failed audit.
- Position size stays with the wallet skill. This build never overrides a daily limit.

## OUTPUT

Report each step that ran, the gate result, and the step that stopped the sequence if one did.
Never report a step as done when its skill returned an error.

A gate is the sentence between two calls. The SKILL.md spec has no branching field, so the condition is prose the agent evaluates when it gets there. Skillsmith writes every gate the same way, because consistency is what makes an agent obey it.

The pattern is already in use. Skills published by Binance use PREREQUISITE lines and point at other skills when a dependency is missing, so a build is written in a convention the ecosystem already reads.

A build installs the same way a single skill does. Downloading the file and pointing the installer at it works offline, and installing from the library pulls it by name.

The source skills a build calls are not bundled with it. Install and authenticate each one separately, then the build orchestrates them.

A build references its source skills by name, so a source skill that changes upstream can drift from what the build assumed. This release does not watch for that.

Rebuild when a skill you depend on ships a change. Open the build in the composer, recompile against the current index, and compare the sequence before replacing the file.

Matching reads the published text of each skill in the index: its name, its domain, its declared trigger phrases and its summary. A brief is scored against all of them, the survivors are ordered by the stage they belong to, and the gates are attached between the steps.

Ordering by stage is why an audit always lands before a swap. A skill that acts never runs before a skill that verifies, whatever order you described it in.

Skillsmith writes instructions on top of skills published by other people. It does not host them, modify them, review them or vouch for them. Responsibility for a source skill stays with its author and with the review process it passed.

Read the sequence before you install a build, and read the source skills before you give them credentials.