nodecode-hades-coding-workflows
Hades coding workflows: three skills for planning a change, executing it through delegated subagents, and running a design–implement–verify harness. A Nodecode rewrite of singhguri/hades-coding-workflows by singhguri.
The three skills are prose, not code, and they are read where they sit:
| Skill | What it is for |
|---|---|
writing-plans |
A plan as bite-sized tasks of two to five minutes each: exact file paths, complete code, the command that proves each task, the output it should print. DRY, YAGNI, TDD. |
subagent-driven-development |
Execute such a plan by giving each task to a fresh subagent with its own context, then two reviews: does it match the spec, then is the code good. |
dual-model-harness |
Design → plan → implement → verify, with the main agent doing the work and delegation only for genuinely independent parts; and the lesson that delegation overhead usually costs more than it saves. |
Get it
Tell the agent "add the hades coding workflows cell", or run:
nodecode add nc://hades-coding-workflows
Set it up
Nothing to set: no key, no service, no dependency. The folder is a knowledge cell, so Nodecode never loads it — the index reads the SKILL.md files as data.
Use it
Ask the agent to plan a change, or to work through a plan; the index lists the three skills, and the agent reads one whole with (help 'writing-plans), (help 'subagent-driven-development) or (help 'dual-model-harness). Each carries its workflow, its prompt templates, its review gates and the pitfalls that cost the original author a session.
What the skills say about the harness is said in Hermes Agent's words — delegate_task, todo, read_file, patch. In Nodecode the same moves are:
| The skill says | In Nodecode |
|---|---|
delegate_task(goal=..., context=...) |
(fork "task 1" "task 2") for independent work in parallel, or a (sh ...)/eval step when it is serial; the delegation cell adds named model routes for subagents |
todo([...]) |
the board cell, or (context :note ...) for what the turn must not forget |
read_file / patch |
(sh "sed"), (read-file ...), (edit path old new) |
npx tsc --noEmit |
whatever your project's check is, run through (sh ...) — the point stands: the check is the evidence, not your summary of it |
An edit to a SKILL.md here is the skill at the next read, so a lesson learned in a session is kept by editing the file.
What is in it
| Path | What it is |
|---|---|
nodecode-hades-coding-workflows.asd |
The folder as a knowledge cell: :properties ((:knowledge . t)), no component. |
skills/<name>/SKILL.md |
The three skills, unchanged from the original, each with the frontmatter that names it and the one-sentence description the index lists. |
skills/<name>/references/ |
The four reference notes the skills point to, unchanged: GLM provider quirks, context-budget discipline, the gates taxonomy, and the server-action result pattern. |
test/hades-test.lisp |
Reads the folder as the index does: three skills, each found, named, described and read whole. |
License
MIT. The skills are Copyright (c) 2026 Gurpreet Singh: see LICENSE and NOTICE.
