Be Framework Skills

Agent skills for building applications with the Be Framework.

Skills

• be — Development workflow: from user stories to ALPS profiles to implementation
• be-semantic — Story → ALPS → Fake → Agreement → Schema → Be workflow
• semantic-ex — Semantic Exercise: AI-driven data generation → constraint discovery → JSON Schema

Demo

Try this with your agent:

Build a Be app from this story using the be-semantic skill.

As a user, I want to log my weight every day.
Because I want to see if I'm getting closer to my goal weight.

ユーザーとして、毎日の体重を記録したい。
なぜなら、目標体重に近づいているか確認したいから。

Now change just the "Because" / 「なぜなら」 line and run it again:

... Because I want to share the data with my doctor at checkups.
... Because I want it to become a lasting habit.

... なぜなら、医師に診察時に共有したいから。
... なぜなら、三日坊主にならず習慣として続けたいから。

Same noun, three different apps:

Because	Resulting design
I want to see if I'm approaching my goal weight	Goal entity + reached/not-reached Branching
I want to share the data with my doctor	Date-range export + normal-range Branching
I want it to become a lasting habit	Streak counter + continued/broken Branching

The "Because" clause reshapes the domain — that's why be-semantic starts from a story, not a schema.

Behind the prompt, the agent walks Story → ALPS → Fake → Schema → Be. The Fake step is the pivot: the agent generates 50 realistic records from your story, you skim them together, and constraints (maxLength: 80, optional fields, value ranges) emerge from observation — not from defaults. Those 50 records become a shared image of the domain, a common language between you and the agent. Artifacts land under design/.

Installation

Claude Code Plugin

claude plugin marketplace add be-framework/be-skills
claude plugin install be-framework-skills

All skills are automatically available after installation.

Manual

Point your agent at the relevant SKILL.md to teach it how to build Be applications.

Using with Claude Opus 4.7

Recommended settings follow Anthropic's Best practices for using Claude Opus 4.7 with Claude Code.

Runtime settings

• effort: xhigh — the new default in Claude Code for 4.7. This is where the skills work best for coding and agentic tasks.
• max_tokens: aim for 64k or higher. 4.7 uses 0–35% more tokens for the same text than 4.6, so leave headroom for compaction and final reporting.
• thinking: adaptive only. Manual budget_tokens returns an error on 4.7. If you want more thinking, steer with a prompt like "Think carefully and step-by-step before responding; this problem is harder than it looks."
• Don't pin effort: you can toggle mid-task with /effort. Use xhigh for schema design; drop to medium for a typo fix.

How to work with Claude Code on 4.7

4.7 works best when you treat it as a capable engineer you're delegating to, not a pair programmer you guide line by line. The skills in this repo were designed that way from the start:

• Front-load the first turn. be-semantic Step 1 requires "story + because + entity enumeration" precisely so intent, constraints, and acceptance criteria land in turn 1. Vague prompts drip-fed across many turns hurt both token efficiency and quality on 4.7.
• Use auto mode (Y) when you can. The Y/n gate at the start of be-semantic runs ALPS → Fake → Schema → Be implementation end-to-end. In Claude Code Max you can also toggle auto mode with Shift+Tab.
• Minimize user interrupts. Take agreement only at the steps where the domain can bend — Step 2 (ALPS HTML review) and Step 3 (Fake 50-item preview) — not at every turn.

Why the skills were tuned for 4.7

Compared to 4.6, Claude Opus 4.7:

• Follows instructions more literally (vague adjectives like "appropriate" are taken at face value).
• Spawns fewer subagents by default (state the fan-out trigger if you want parallelism).
• Calls tools less often and reasons more (if a tool execution is the completion gate, say so).
• Reports progress on its own (scaffolding like "summarize every N steps" is unnecessary).

The edits in this PR concretize vague phrasing and make subagent triggers explicit to match these behaviors. The skills still work as-is on 4.6.

Philosophy

Be's development flow is a process of raising resolution for AI-driven development:

User Story        ← domain language (ambiguous)
  → ALPS Profile  ← formalized state transitions & semantics
  → Semantic-Ex   ← data generation → observation → constraint discovery
  → JSON Schema   ← unambiguous goal (no room for interpretation)
  → Be code       ← implementation converges to schema

Natural language specs produce 100 different implementations. Schemas produce one.