bmtrnavskyA coding agent for non-engineers and vibe coders
A lean skill pack + persona for any coding agent (built for Hermes;
portable to any agent that supports skill files). For someone who wants to
build real software — small fixes through multi-week projects — without
being a software engineer, with lead-dev discipline explained in plain
language at every step.
This isn't a shortcut around learning. It's scaffolding: the same rigor a
senior engineer would insist on (define before building, test before
shipping, verify before claiming done), applied consistently so you can lean
on it instead of having to already know it.
Sibling project:
spec-driven-data-analysis
applies the same discipline to data work — lock the spec before touching
the data, verify every run.
The loop
Define (intake -> brainstorm) -> Plan (research <-> plan) -> Teardown -> Build (test-first -> execute <-> debug) -> Verify (verify -> version)
Full pipeline with gates and artifacts: plugins/disciplined-vibe-coding/SKILL.md
(Intake → Define → Economy gate → Plan → Teardown → Build → Verify → Commit, Wayfinder
riding along). Using another agent? See READERS.md.
One loop per small task, or per milestone of a larger project. Same four
words every time — the repetition is deliberate. You learn the shape of good
engineering practice by watching it applied the same way, over and over,
not by reading a lecture about it once.
What it actually stops you from doing
- Building before anyone (you or the agent) has said out loud what "done"
looks like. - Shipping a destructive change with no rollback plan.
- Claiming something works because a command didn't error, rather than
because someone actually checked the output. - Storing payment card data instead of using a provider-hosted checkout.
- Running something autonomous (a bot, a scheduler) with no kill-switch.
- A near-miss on a metric getting quietly rounded up to "good enough."
- Moving past a failed gate without a logged bypass (who approved, what
was red, why it was waived). - Committing secrets: a secret scan runs before every commit.
Stakes-aware, not one-size-fits-all
Every task gets a tier: TOY (learning, nothing at risk), WORKFLOW
(your own systems could break), or MONEY (payments, live trading,
anything public-facing). The tier changes what gets asked, what gets a
rollback plan, and whether a stronger model reviews the work before it's
called finished.
Layout
SOUL.md the persona — how it talks, what it never compromises on
USER.md.template fill in your own background, environment, and preferences
UPDATE.md update contract: shipped paths read-only to you, user edits in user/
scripts/update.sh updater for manual checkouts only (catalog installs: `hermes plugins update disciplined-vibe-coding`)
user/ your overlay — customizations live here, never touched by updates
scripts/selfcheck.sh acceptance checks for the pipeline's hard rules
HOOKS.md optional shell hook for research-freshness nudges
hooks/ the one hook this pack ships
skills/ the ten skills that make up the loop
plugins/disciplined-vibe-coding/
v0.2.0 pipeline overlay (stages, gates, rung rules),
economy ladder, to-spec/plan/wayfinder templates,
changelog
READERS.md use this pack from any agent with a skill loader
(or none — paste-per-stage fallback included)
ATTRIBUTIONS.md what this pack draws from, and where
SOURCE-LICENSES.md full license text and copyright lines for every source
CATALOG.md draft entry for the Hermes plugin catalog
Install
Hermes: SOUL.md → your profile's SOUL.md; skills/ → your profile's
skills directory (or a shared skills path referenced via
skills.external_dirs); USER.md.template → fill in and save as USER.md;
hooks/ → optional, wire into config.yaml per HOOKS.md if you want the
research-freshness nudge.
Pipeline orchestrator: plugins/disciplined-vibe-coding/SKILL.md is the
entry skill that runs the whole loop; the root __init__.py registers it as
disciplined-vibe-coding:disciplined-vibe-coding alongside the ten stage
skills. Load this repo as a Hermes plugin (or point your loader at that
SKILL.md) so the loop has a driver — the stage skills alone have none.
Elsewhere: SOUL.md → your system prompt; the SKILL.md files → wherever
your tool loads skills from.
License
MIT for everything original to this pack (SOUL.md, USER.md.template,
HOOKS.md, this README). The ten skills derive from several MIT-licensed
sources, and the pipeline's elephant/goldfish framing is adapted from
Google's Elephant-Goldfish Model (Apache-2.0). See ATTRIBUTIONS.md for
lineage and SOURCE-LICENSES.md for the full license text and copyright
line of every upstream project. Those notices travel with this pack per
their license terms; keep them if you fork or redistribute.
README from the reviewed commit ac4de52; it updates when the author re-pins.
