A skills library for Claude Code that trades interviewing for building.
hordev takes a small amount of input, decides the rest, writes a spec and a test-driven design without asking anyone to approve them, and then throws a horde of cheap fast agents at the problem in parallel. The goal is a running prototype you can judge, quickly, across several directions at once.
It is a descendant of superpowers and keeps its tenets: skills as version-controlled markdown, tests before implementation, artifacts that outlive the session. It diverges on pacing.
They are aimed at different jobs, and hordev is not a strict upgrade.
Use superpowers when structure is what you need: spec-driven development, a framework to work inside, long autonomous runs on well-understood work, and review checkpoints you actually want. That structure is where superpowers earns its keep, and hordev deliberately gives some of it up.
Use hordev when you need to find out fast, and broadly: you are not sure what to build yet, several approaches look plausible, and the cheapest way to learn is to have working versions of them in front of you. Breadth is the point as much as speed — a horde is good at exploring four directions at once, not just at doing one direction faster.
If you are executing a spec you already trust, hordev's speed buys you little and costs you the checkpoints. Use superpowers.
Every hordev run moves through five stages:
| Stage | Skill | Model |
|---|---|---|
| Extract | rapid-spec |
haiku |
| Design | writing-tdds |
haiku |
| Cut | decomposing-for-hordes |
opus |
| Swarm | dispatching-hordes ,reconciling-horde-output |
haiku fan-out, opus merge |
| Verify | horde-qa |
opus |
When the right approach is genuinely unknown, the horde goes wide instead of
deep: racing-prototypes builds two to four competing versions at once, under
identical budgets and pre-committed judging criteria, and picks a winner on
evidence from running code.
Supporting skills: prototype-first, racing-prototypes, assumption-ledger,
debugging-in-a-horde, isolating-horde-workspaces, improving-hordev,
writing-hordev-skills. using-hordev is the entrypoint, injected at session
start by a hook.
No approval gate. hordev never asks you to sign off on a spec or a TDD. It decides, writes them, and starts building. This is the whole speed bet, and it is why QA is the most important skill in the library — with nobody approving the design, verification is the only thing between a wrong spec and a wrong prototype.
Every question you did not get asked is written down. Not asking is only
honest if the decision is visible, so each one becomes an entry in
.hordev/assumptions.md with its blast radius and how it will be falsified.
You see the open ones at the end of a run. That ledger is the contract that
makes skipping the interview defensible instead of reckless.
Cheap agents do the volume, strong agents do the judgment. Specs, TDDs, and implementation run on Haiku; orchestration, decomposition, reconciliation, and QA run on Opus. Reaching for the strongest model everywhere defeats the purpose of a horde.
Runs append to .hordev/run-log.md. A Stop hook notices when the log has
grown and asks whether anything should change a skill. The bar is deliberately
high: amend on a repeated failure, never on a one-off. improving-hordev maps
symptoms back to the skill that probably needs the edit — seam bugs usually
mean the decomposition rules are too loose, agents returning plans instead of
code mean the dispatch prompt contract is weak.
The skills drop the occasional orcish line — "Zug zug" taking the job, "For the Horde!" when a wave dispatches, "Work complete!" when the tree comes back together. It is a running joke with rules: conversational output only, never two messages in a row, and never on top of bad news. Specs, TDDs, commit messages, and QA findings stay clean, because those outlive the joke.
In Claude Code, add this repository as a plugin marketplace, then install from it:
/plugin marketplace add heffrey/hordev
/plugin install hordev@hordev
The entrypoint skill loads itself at session start, so there is nothing else to configure. Restart the session after installing.
Requires bash. Nothing else.
Early. Version 0.2.0.
The skills are written. Both hooks run correctly when executed directly —
session-start.sh emits the entrypoint as valid JSON, and reflect.sh blocks
once when the run log grows, honors the stop_hook_active loop guard, and
stays silent afterward. They have not yet been exercised inside a live Claude
Code session, which is a different thing and the next thing to verify.
The library has not been run in anger across enough projects for its own self-improvement loop to have taught it anything. Expect rough edges, and expect these rules to change as real runs falsify them.
MIT. See LICENSE.
Derived in part from superpowers by Jesse Vincent, used under the MIT License.