Matt Pocock Skills + Hermes Sub-Agents for Feature Work

How to use

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

Use case: an existing project with one big feature or refactor touching multiple surfaces — backend, frontend, docs, landing page, tests — not a whole project from scratch.

The issue split should be vertical, not horizontal.

A better vertical split looks like:

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

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

This step still needs human judgment: the user reads the plan, rejects bad issue splits, and stops AI slop before it turns into code.

Skill compatibility: Pocock's skills use Claude Code slash-command conventions, which Hermes parses as skill invocations — the user is effectively running Pocock's prompts inside Hermes's tool-calling system.

Step 1: Ask Hermes for a local markdown plan covering scope, risks, surfaces touched, and a 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, and let it update the markdown.

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

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 6: Run one sub-agent per issue, halting on HITL tasks.

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.

Configuration tweaks (required).

Out-of-the-box Hermes can hit iteration exhaustion on deep sub-agent chains, so bump the settings before starting.

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

Then restart the Hermes session.

Limitations and notes:

• 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, so the first run is slower and 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.

Links

• Self-Evolution
• Zouroboros Swarm Bridge
• Paperclip Managed Employee