agentxp: Skill for Claude Code

Skill: /design

Purpose

The design verb pre-registers an experiment. You walk from intent → semantic models → metrics → hypothesis → brief → data plan → sealed brief, dispatching specialists at each step. The SQL safety pipeline runs in mode="design" — Layer 3d rejects any query that references outcome columns (variant, arm, assigned_arm, metric values). There is no --force.

The verb terminates when the brief seals with the three-part integrity lock (design_chain_hash + metric_snapshot + expected_shape). At that point the user invokes /analyze --brief experiments/<id>/brief.sealed.yaml to enter the analyze verb.

When to invoke

Direct: /design [--data PATH] [--exp-id ID]

Plain-English routing:

Phrase	What to do
"I want to test a new checkout button"	/design, then capture intent (step 2)
"Design an experiment against the demo warehouse"	/design --data sample-data/agentxp_demo.duckdb
"Continue exp_a3f9c102"	/design --exp-id exp_a3f9c102
"Show me the lift" (in design mode)	REFUSE — R11 wall; that is the analyze verb's job

Procedure (do these in order)

1. Allocate the experiment directory

from pathlib import Path
from agentxp.workflows.design import allocate_experiment

exp_dir = allocate_experiment(
    project_root=Path.cwd(),
    data_path=Path(args.data) if args.data else None,
    experiment_id=args.exp_id,  # None → ULID-flavored auto-id
)

The helper creates experiments/<id>/, seeds log.md, and stashes the data path if supplied. Print the experiment id so the user can reference it later.

2. Capture intent

Ask the user for their intent in plain English. When you have it, persist:

from agentxp.workflows.design import record_intent

intent_path = record_intent(
    exp_dir,
    intent_text=user_supplied_text,
    captured_by=user_email_or_handle,
)

This writes intent.yaml and appends to log.md. Then render the intent share-tail: