project-init
Project Init
Run this workflow only for greenfield initialization.
1. Mode Guard
Before interviewing, check:
- If both
docs/architecture.mdanddocs/plans.mdexist, stop and switch toproject-update. - If docs are missing but code roots/manifests exist (
src/,app/,apps/*/src,package.json, etc.), stop and switch toproject-convert. - Continue only when docs are missing and codebase signals are absent.
2. Interactive Interview
Collect requirements before generating any docs:
- Ask project name and core goals.
- Run one market snapshot (similar products, patterns to reuse/avoid).
- Draft project intro + end-to-end core flow and confirm with the user.
- Propose 3-4 architecture + visual directions; get user choice.
- Run clarifying rounds (user journeys, components, stack, deployment, constraints).
- Assign complexity tier (Standard by default; Complex only when justified) and confirm.
- Provide final synthesis; require explicit user approval before writing files.
3. Generate Documentation
After user confirmation, create/update:
docs/architecture.mddocs/plans.mddocs/implement.mddocs/secrets.mddocs/documentation.mddocs/design.mddocs/decisions.md(for Standard/Complex tiers)tasks/todo.mdtasks/lessons.mdCLAUDE.mdAGENTS.md
Rules:
- Keep
docs/architecture.mdas the stable product/technical source of truth. - Keep
docs/plans.mdmilestone-driven and ensure Production Readiness Gate is the final milestone. - Keep docs consistent with each other; never leave unresolved contradictions.
- In
docs/implement.md, add mandatory sections:First-Principles ExecutionAtomic Task ExecutionRewrite-First PolicyNo-Compatibility Rule
- In
docs/plans.md, every task must includeCommit Boundary: exactly one atomic commit. - In
tasks/todo.md, every task must includeTask IDandCommit Status(pendingordone).
3.5. Implementation Discipline (Mandatory)
Apply these rules to every implementation task generated by this workflow:
- Use first-principles execution: solve root causes from requirements and architecture, not superficial patches.
- Treat
taskandsub-taskas the same execution unit. - Enforce
one task -> one atomic commit. - Finish the full task scope before commit: all affected code, tests, configs, scripts, and docs.
- Follow
code is cheap: if a module becomes hard to fix cleanly, delete and rewrite the module. - Do not add compatibility code (
adapter,shim, dual-path logic, deprecated aliases, temporary compatibility flags). - If a rewrite changes interfaces, fix all impacted callers in the same task.
- Do not leave breakage for a follow-up task just to preserve compatibility.
4. Review Protocol
Run all review gates before handoff:
- Tier review:
- Standard: Agent 2 + Agent 3
- Complex: Agent 1 + Agent 2 + Agent 3
- Mandatory cross-model review via
mcp__codex__codexondocs/architecture.md+docs/plans.md. - Apply valid findings and re-check consistency.
- Ask user for one annotation pass on
docs/architecture.mdanddocs/plans.md, resolve notes, and wait for explicit "ready".
If mcp__codex__codex is unavailable, do not skip silently. Ask user to choose manual checklist path or alternate-model path and record which was used.
5. Hooks Installation (Mandatory)
Before final messaging, auto-install hooks:
- Detect package manager:
bun.lock-> bunpnpm-lock.yaml-> pnpmyarn.lock-> yarn- otherwise npm
- Resolve installer path in this order:
scripts/setup-hooks.sh../full-project-skill/scripts/setup-hooks.sh
- Run:
bash <resolved-setup-hooks-path> --pm <detected-pm> --project-dir <project-dir> --platform both
- If no installer path exists, ask the user where their hook installer lives and stop.
- If installation fails, show the error and stop for user resolution.
6. Final Handoff
Tell the user:
- Docs are ready under
docs/. - Hooks are installed to
.claude/and.codex/. - They should review
docs/architecture.md+docs/plans.md. - Execution should follow
docs/implement.md. - Work is complete only after passing the Production Readiness Gate milestone.
More from phlegonlabs/skills
project-long-task
>
14harness-engineer-cli
>
10full-project-skill
>
4project-update
>
3project-convert
>
3harnass-engineer-plan
Research an existing repository, align with the user, and generate a reviewed machine-readable execution plan. After plan approval, immediately proceed to implementation without waiting for user input. Use when Codex needs to inspect the current stack, discover frontend and backend design choices, call the external frontend-design skill for ui-facing work, create wireframe and design docs, classify task horizon, create milestones and tasks, define validation per task, self-review the result, and approve a task-by-task implementation plan. When the target repo lacks the `harnass-os/...` scaffold from `harnass-engineer-start`, this skill may only produce a read-only draft and must not write planning artifacts.
1