Small models do the work.
Frontier models direct it.

Agent skills that make your frontier model a senior engineer and a cheaper model the engineer who does the work. Half the cost, same hidden tests.

curl -fsSL https://hstack.shellhaki.dev/install | shView the source

Benchmark 2 · production waitlist backend · Claude Code

Opus directing Haiku: half the cost, same hidden tests.

Cost to build the same task
Frontier alone
Claude Opus 5.5
$0.00
hstack
Opus directs · Haiku builds
$0.00
−53%
cost · $0.75 vs $1.59
100%
quality · 20/20 hidden tests, both

Same spec, separate Claude Code sessions. Opus 5.5 wrote 90% fewer output tokens. Costs are Claude Code's own per-model list-price figures, sub-agent included, judged by a 20-test suite neither side saw.

Plugs into your harness

hstack is a set of skills and sub-agents. It runs inside the agent you already use, on the subscription you already pay for.

Claude Code
/hstack-activate
Codex CLI
coming soon
OpenCode
coming soon
Cursor
coming soon
Uses your existing loginNo API keysNothing changes until you run /hstack-activate

The frontier decides why and what. The worker works out how.

A real run from the benchmark: one brief, one approval, one correction, and no code from the frontier.

waitlist backend
20/20 hidden tests
  1. Opus · frontier · 0:20
    Brief. Build the waitlist backend in BENCH-SPEC.md. Done when tsc and vitest pass. I'll review the rate limiter, CSV formula guard, cursor and signup check order.
  2. Haiku worker · 1:05
    Plan in PLAN.md: SQLite store, pure handlers, tests first for every spec example and error code. ASK: approve
  3. Opus · frontier · 1:12
    APPROVE: build it
  4. Haiku working · 199 tests written
  5. Haiku worker · 5:40
    Done. tsc clean · 199/199 passing.
  6. Opus · frontier · 6:10
    Ran the checks. One bug: re-signup with a pending email replaces the unsubscribe token, so the first email's link breaks. Keep the token; add a regression test.
  7. Haiku worker · 7:20
    Fixed, with a regression test. All tests pass.
  8. Opus · frontier · 7:50
    APPROVE
  1. Frontierstep 1

    Writes the brief

    About 15 lines: why, acceptance criteria, constraints, the risks it will check, and the done commands. No code.

  2. Workerstep 2

    Does the engineering

    A sub-agent on a smaller model plans, writes tests first, builds, and reports. /hstack-swarm splits the work across parallel workers and a reviewer.

  3. Frontierstep 3

    Reviews and directs

    Approves the plan, runs the checks itself, spot-checks its risks, and sends at most one short correction. It never writes the code.

hstack workflow: the frontier writes the brief, the worker plans and builds, optionally with parallel workers and an independent reviewer, and the frontier approves at the plan and done checkpoints.

Four skills today. A skill bank next.

Plus two sub-agents, hstack-worker and hstack-reviewer, that the skills start on your worker model.

hstack-activatefrontier

Turns hstack on and keeps asking for a worker model until you name a valid, smaller one. Loose names like "haiku 5.5" work.

hstackfrontier

The senior engineer: size gate, brief, plan review, build review, one correction round, token discipline.

hstack-swarmfrontier

Parallel workers on independent parts, then an independent reviewer that tries to break the result.

hstack-practiceworker

Engineering habits small models lack: exact specs, state tables, tests first, real-environment facts, security, honest reports.

Skill bank coming next

Curated skills that make the worker better on its own: testing, security, frameworks, data and UI quality. The more it knows, the less the frontier has to correct.

Don't trust the numbers? Run them.

$ curl -fsSL https://hstack.shellhaki.dev/install | sh
✓ linked 4 skills and 2 sub-agents into ~/.claude
$ claude --model opus
> /hstack-activate
Which smaller model should do the work? haiku 5.5
hstack active · opus directs, haiku builds
>
  • Every benchmark is in the repo. Spec, hidden tests, runner and raw numbers. Run it yourself.
  • Results are published either way. Wins and losses, from +68% to −53%.
  • Plain Markdown skills. Read them, fork them, change the brief template.
go install github.com/shellhaki/hstack/cmd/hstack@latest && hstack

Your frontier model shouldn't be typing.

Let it direct. Let a cheaper model build.

Read the docs