Zero-Shot Harness — spec-first, AI-native project building

Three skills for building software spec-first. Give it a one-line idea; walk away with a working, tested, phased project — on whatever stack the requirements pick. The skills are repo-independent: install them once and they operate on whatever project repo you point them at. No default stack, no boilerplate to clone — the spec derives the stack, designs the layout, and a scaffold gate proves the skeleton runs before any feature work.

The three skills

The skills are disable-model-invocation: true — trigger them explicitly:

/build 
/fix 
/sync [scope]

The Spirit

1. Spec is the source of truth. Written before code; when they disagree, the spec wins (/sync).
2. No default stack. Your stated preferences are binding; everything else is derived from the requirements and recorded with rationale. The scaffold gate proves the skeleton runs before any feature work.
3. Smallest first-time-right win, phase by phase. Each phase ships the smallest increment a human can test, and it must work the first time.
4. A human gates every phase. Autonomous within a phase; stops at each boundary for you to test the increment.
5. AI-native by design. Every spec includes spec/agent.md — the idea evaluated against the agentic-patterns catalogue. “No AI capability needed” is a legitimate written conclusion; an unexamined design is not.
6. Real services or it doesn’t count. Gates run against the real external services the spec names — the real LLM when the design has AI capability, the production database engine always. A stubbed pass is not a pass.

Install

The skills/ tree follows the open Agent Skills standard, so it runs on Claude Code (as a plugin), Hermes (as skills), and any other Agent-Skills-compatible tool. The support/ tree is the shared doctrine; only the thin platform wrappers differ.

Claude Code

/plugin marketplace add smallTechOrg/zero-shot-harness
/plugin install zero-shot-harness@zero-shot-harness

The three skills, four sub-agents, and the shared support/ material all come with the plugin. Confirm with /plugin (it should list zero-shot-harness as enabled). Then do the one-time Claude Code autonomy setup for each project you build in — without it, every build command raises a permission prompt and the “autonomous within a phase” promise breaks.

To uninstall:

/plugin uninstall zero-shot-harness --scope project
/plugin marketplace remove smallTechOrg/zero-shot-harness

Use --scope user if you installed globally.

Hermes

hermes skills tap add smallTechOrg/zero-shot-harness
/reload-skills        # or restart Hermes

Full instructions (per-skill install, support/ placement, autonomy model) in docs/hermes-setup.md.

To uninstall:

hermes skills tap remove smallTechOrg/zero-shot-harness   # if tapped
# or, if per-skill installed:
hermes skills uninstall build fix sync
rm -rf ~/.hermes/support

Layout

skills/
  zero-shot-build/   SKILL.md + references/hermes-pitfalls.md
  zero-shot-fix/     SKILL.md
  zero-shot-sync/    SKILL.md
support/             shared doctrine the skills reference (as support/…)
  agents/            spec-writer, code-generator, qa-auditor (dual-mode)
                     + project-builder (Claude-only orchestrator — never run on Hermes)
  patterns/          spec-driven, phases, agentic-ai, test-driven, ui-ux, …
  rules/             ai-agents, git, secret-hygiene

.claude-plugin/      Claude Code wrapper: plugin.json + marketplace.json
agents/              Claude-native sub-agent pointers into support/agents/
CLAUDE.md            Claude entry point   (AGENTS.md / .hermes.md = Hermes)
docs/                claude-code-autonomy.md · hermes-setup.md

The skills register as slash commands themselves — there is no separate commands layer, so the skill file is the single source of truth for each flow.

Each SKILL.md is the process — follow it exactly; never improvise your own build/fix flow.

FAQ

What if I already have a stack in mind? State it in the idea: /build [idea] — use Python + FastAPI + PostgreSQL. Stack choices are binding. With no preference stated, the spec-writer derives the best fit from your requirements and records the rationale in spec/architecture.md.

Does every project get AI bolted on? No. spec/agent.md is always written, but “no AI capability needed” is a first-class conclusion — deterministic logic that fully serves the requirements beats an LLM every time.

What if something breaks? /fix [what's broken] — the qa-auditor role classifies SPEC vs CODE, the generator fixes, qa-auditor re-gates, the root commits + pushes.

What if spec and code drift? /sync — qa-auditor audits, generators fix, spec wins.

Why didn’t my build land on main? By design. Builds live on feature branches whose PRs target the branch they were cut from — you review and merge; the build never merges its own PR.