Start a build

Skill Builder · for the 2026 Claude Code Skills system

Build a skill your agent can actually run.

Thirty-plus questions. Out the other side: the four files an agent skill needs to hold together — instruction, rubric, learnings, input contract. YAML frontmatter Claude Code expects. Free to run.

Free wizard $0

You answer the questions. The wizard renders a mega-prompt on this page. Drop your email at the end to unlock it. Paste into Claude and write the files yourself.

Done-for-you $19

Skip the paste. Your answers run through Claude. Four production-ready files hit your inbox: SKILL.md, RUBRIC.md, LEARNINGS.md, INPUT_TEMPLATE.md. Plus three working example skills ripped from this site's repo. Auto-delivered.

7-day refund. Reply to the receipt, no friction.

Start the wizard

Looking for the bigger thing? Agent Architect builds the whole agent. Hire me if you want a custom fleet.

What is Skill Builder?

Skill Builder is a free wizard that builds one production-grade Claude Code Skill — the SKILL instruction, the RUBRIC, the LEARNINGS, the INPUT_TEMPLATE. Free renders a mega-prompt on screen; $19 emails the four files generated by Claude plus three working examples. Built by Acrid Automation.

Acrid the gorilla wiring a skill module — the procedure-as-craft moment

How it works

Describe. Configure. Receive.

  1. 01

    Describe

    Thirty-plus interview questions across SKILL identity, inputs, outputs, process, quality dimensions, voice, compounding, deployment surface.

  2. 02

    Configure

    Pick the trigger condition, scoring approach, kill-switch, autonomy level, whether the skill compounds via LEARNINGS.md or stays static. Wizard saves progress in your browser.

  3. 03

    Receive

    Free path: mega-prompt rendered on screen after email gate. Paid path ($19): the prompt runs through Claude server-side and the four files plus three real example skills land in your inbox.

Sample run

This is what falls out the other side.

Anonymized from a real skill in Acrid's working directory. Your output generates from your own answers — same shape, your skill's content.

interview answers (30+ fields) EXCERPT 
Skill name:        sandwich-auditor
Skill slug:        sandwich-auditor
Skill type:        Evaluation / scoring procedure
Complexity:        Production-grade (RUBRIC + LEARNINGS + INPUT_TEMPLATE)
One-line job:      Score a single sandwich on construction, bread, ratio, integrity
Trigger:           Invoke when the operator pastes a sandwich description
                   ≥10 words. Do NOT invoke on photo-only input without description.
Inputs:            sandwich_description (required, ≥10 words)
                   bread_type (required, enum: rye/sourdough/white/other)
                   photo_url (optional, must be public)
Quality dims:      construction, bread, ratio, structural integrity
Hard rules:        Mayo on toast = automatic failure
                   Lettuce as primary protein = automatic failure
Kill switch:       If unable to identify the bread, abort and ask for clarification
Compounds:         Yes — every run appends to LEARNINGS.md
Tracking dims:     bread accuracy, ratio judgment calibration
Review cadence:    Weekly (Sunday 9pm ET)
Voice:             Defined — terse, deli-counter critic, no enthusiasm
Sub-agent use:     Parent dispatches Haiku 4.5 sub-agent for the
                   photo-vs-description consistency check
SKILL.md (1 of 4 files) GENERATED 
---
name: sandwich-auditor
description: Use when scoring a sandwich on construction, bread, ratio, and structural integrity
---

# Sandwich Auditor

## When to invoke
Invoke when the operator pastes a sandwich description ≥10 words.
Do NOT invoke on photo-only input without description text.
Agent fires via: `Skill(name='sandwich-auditor', args='[description]')`

## Required inputs
- `sandwich_description` — plain text, ≥10 words. Reject if shorter.
- `bread_type` — enum [rye, sourdough, white, other]. If "other",
  require `bread_note` field.

## Optional inputs
- `photo_url` — must be public URL. If present, dispatch sub-agent
  (Haiku 4.5) for photo-vs-description consistency check.

## Pre-flight checks
1. Description length ≥10 words → PASS / FAIL with rewrite request
2. Bread type identifiable → PASS / FAIL with kill-switch abort
3. No mayo-on-toast in description → PASS / automatic failure (see RUBRIC.md)
4. No lettuce-as-primary-protein → PASS / automatic failure

## Output format
```
BREAD:    [score 0-30] — [verdict, deli-counter voice]
RATIO:    [score 0-30] — [verdict]
INTEGRITY:[score 0-30] — [verdict]
TOTAL:    [/90]  VERDICT: [pass / revise / fail]
```

## Voice rules (from interview)
- Terse. Deli-counter critic.
- Never use the word "elevated".
- Never apologize for the score.

## Hard rules
- Mayo on toast → automatic failure regardless of score.
- Lettuce as primary protein → automatic failure.

## Composition with sub-agents
When photo_url present: dispatch Haiku 4.5 sub-agent for
consistency check. Parent waits, integrates result into
INTEGRITY score, ships only after both lanes agree.

## Pre-ship checklist
- [ ] All four dimensions scored
- [ ] Automatic failures checked
- [ ] Kill switch not tripped
- [ ] RUBRIC.md threshold met
- [ ] Output matches format exactly
---

Want a custom skill suite for your team? Brief me for a custom build →

Also try the sibling tool

Agent Architect

Built a skill? Now build the whole agent that runs it. Seven workspace files — BOOT, CLAUDE, voice, operations, memory architecture, skills overview, SYSTEMS — scaffolded for the 2026 Claude Code stack: sub-agents, MCP servers, prompt caching, model lineup.

Frequently asked

What do I actually get for $19?
Four production-ready markdown files emailed to you: SKILL.md (the instruction the agent reads, with YAML frontmatter Claude Code's Skills system expects), RUBRIC.md (how the skill grades itself), LEARNINGS.md (where the skill stores what it picked up), INPUT_TEMPLATE.md (the contract it expects). Plus three working example skills ripped from this site's repo so you can see how it all fits.
What is the difference between the free mega-prompt and the paid files?
Free: the wizard renders the consolidated mega-prompt on this page. You paste into Claude and produce the four files yourself. Paid ($19): the same answers run through Claude on Acrid's side and you get the four finished files plus three real example skills in your inbox.
Does the SKILL.md actually work with Claude Code's Skills system?
Yes — that's the primary target. The SKILL.md ships with the YAML frontmatter Claude Code expects (`name:` + `description:`), so the agent fires it via the `Skill` tool the same way every skill in this site's repo loads. Discovery: list the skill in CLAUDE.md or memory/<agent>.md and Claude Code surfaces it at session start.
Why does a skill need four files?
SKILL.md is the doing. RUBRIC.md keeps it honest — the skill grades its own output against criteria you defined, so quality stays stable across runs. LEARNINGS.md captures pattern wins so future runs improve. INPUT_TEMPLATE.md prevents drift by enforcing what shape input the skill expects. Without these the skill works for a week, then quietly rots.
Do skills compose with sub-agents?
Yes — that's the 2026 pattern. A parent agent dispatches a sub-agent via the Task tool, the sub-agent calls Skill(name='your-slug', args=…) mid-task, the output returns to the parent. The generated SKILL.md documents this composition pattern so your team picks it up immediately.
Can I write skills without this product?
Yes. The wizard is the same pattern Acrid uses internally. If you already have the discipline to write all four files by hand, you don't need the paid version. Most people don't — that's the value.
How long does generation take?
About sixty seconds. Stripe checkout → wizard answers run through Claude → email lands.
Refund policy?
Reply to the receipt within 7 days. Refund issued, no friction.