Skip to content

matt-pocock-skills-subagents.md

Latest commit

 

History

History
150 lines (98 loc) · 5.91 KB

matt-pocock-skills-subagents.md

File metadata and controls

150 lines (98 loc) · 5.91 KB
Raw

Matt Pocock Skills + Hermes Sub-Agents for Feature Work

← Browse use cases · Coding & Engineering

At a glance Details
Outcome A planning-to-issues workflow with example artifacts and delegated implementation steps.
Requires Matt Pocock skills; Hermes delegation; a scoped feature and project context.
Available Walkthrough
Validation Sources linked below; workflow not run during this catalog redesign.
Aspect Detail
Category Software Development
Complexity Advanced
Hermes Features Used skills, delegate_task, hermes config set, file-writer skill
Primary Sources Reddit post by Hermes user (setup/guide), linked repos, star counts verified live

Overview

Matt Pocock's public skill repo handles planning and issue breakdown. Hermes's builtin sub-agents handle implementation. Together they give you a planning layer that prevents "AI slop" and an execution layer that parallelizes across vertical slices.

Pocock's repo had 45.6k GitHub stars at verification time (up from 42.5k when the Reddit post was written).

Since Hermes already has built-in sub-agents, the Hermes user skips Sandcastle and runs implementation natively.

Use case: an existing project, one big feature or refactor, multiple surfaces. Backend, frontend, docs, landing page, tests, etc. Not a whole project from scratch.

Who's Doing This

A Hermes user running a mature codebase who wants to land a cross-cutting feature without giving a single agent a vague giant task. They know Matt Pocock's skills from the TypeScript community and want the same rigor for AI-assisted development.

How It Works

Step 1: Ask Hermes for a Local Markdown Plan

Scope, risks, surfaces touched, rough outline. No code yet.

Hermes, write a plan for ~/projectA/big-feature-x.md covering scope,
riskiest assumptions, surfaces touched, and a rough task outline.
Save as local markdown, do NOT create a GitHub issue.

Step 2: Run Matt's grill-me or grill-with-docs Skill on the Plan

Answer the hard questions. If the agent drifts into implementation, tell it to refocus on the feature plan. Let it update the markdown.

/grill-with-docs on ~/projectA/big-feature-x.md

This skill also updates CONTEXT.md and ADRs inline, building a shared domain language that makes later sub-agent instructions shorter and more precise.

Step 3: Run Matt's to-spec Skill

/to-spec on ~/projectA/big-feature-x.md, create local md spec not in github

Step 4: Run Matt's to-tickets Skill

/to-tickets on ~/projectA/big-feature-x-spec.md, create local md tickets not in github

Step 5: Check the Issue Split Before Running Agents

It should be vertical, not horizontal.

Bad split (horizontal):

  • backend
  • frontend
  • docs
  • tests

Better split (vertical):

  • small backend path + one frontend entry point
  • user-facing flow + error state
  • docs/landing update for the behavior
  • verification/tests/cleanup

Ask Hermes to apply more verticalization; it will know what you mean.

Step 6: One Sub-Agent Per Issue, Halt on HITL

Implement the feature split slices in ~/projectA/big-feature-x-issues.md,
use sub-agents for every issue, and parallelize when possible but carefully.
Halt on tasks that are HITL or otherwise require human input.

Still needs human judgment. The user reads the plan, rejects bad issue splits, and stops AI slop before it turns into code.

Configuration Tweaks (Required)

Out-of-the-box Hermes can hit iteration exhaustion on deep sub-agent chains. Bump before starting:

hermes config set delegation.max_iterations 100
hermes config set delegation.child_timeout_seconds 1200

Then restart the Hermes session.

Real-World Constraints

  • Not for greenfield projects. The workflow assumes an existing codebase with tests, types, and documented patterns.
  • Requires domain context. grill-with-docs builds CONTEXT.md over time. First run is slower; subsequent runs in the same repo are faster.
  • Human gatekeeper. The user explicitly intervenes at the plan-to-code boundary. This is not fully autonomous.
  • Skill compatibility. Pocock's skills use Claude Code slash-command conventions. Hermes parses them as skill invocations. The user is effectively running Pocock's prompts inside Hermes's tool-calling system.

Primary Sources

  • mattpocock/skills — 45.6k stars, MIT license. Skills used: grill-me, grill-with-docs, to-spec, to-tickets
  • mattpocock/sandcastle — Pocock's Claude Code runner. Not used here; Hermes sub-agents replace it
  • Reddit post by Hermes user (setup/guide category) describing the combo workflow, April 2026, including delegation-limit tweaks and the HITL stop gate

Related Use Cases

Why It Works

Pocock's skills solve planning and scoping — the most common failure mode of coding agents. Hermes sub-agents solve execution — the most common failure mode of Pocock's Sandcastle (which is Claude-specific).

Combining the two means you get Matt's rigor on the front end and Hermes's flexible tool-calling on the back end, with a human gate at the plan-to-code boundary.