Init Project

SkillDev tools

Internal Harness instruction source for init-project. Route through visible Harness aliases instead of invoking directly.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Init Project skill

What this skill tells your AI

The instructions your AI receives, as published by linzhe001/harness-research in .agents/skills/init-project/SKILL.md and read by ahel’s review.

Use this Skill for init, update, update-from-grill, or deps-changed. It maintains compact guidance, preserves ## Custom, and does not approve research contracts.

Read First

  • ./references/claude-md-template.md
  • ./references/claude-maintenance.md
  • context, language, documentation, and ubiquitous-language rules under ../../../.agents/references/
  • Existing PROJECT_STATE.json, CLAUDE.md, AGENTS.md, README.md, and OPERATOR_CONTEXT.md when present
  • For update-from-grill: docs/05_intake/Research_Intent_Draft.md, docs/05_intake/Grill_Round_Log.md, docs/05_intake/Execution_Readiness_Packet.md, and docs/05_intake/Execution_Readiness.json when present

Modes

init

Collect project name and known environment name. Create the minimal CLAUDE.md from the template. If no runnable environment exists, leave an explicit WF5-owned environment placeholder. Create dynamic-context directories only when the project opts in; do not fabricate evidence. Write OPERATOR_CONTEXT.md only for explicit stable operator preferences.

update

Read PROJECT_STATE.json and source artifacts. Fill only sections whose source is now known: idea, debate/refinement, tech stack, environment, dataset paths, baseline references, project structure, core artifacts, language policy, entry scripts, and current stage. At WF4, sync dataset paths into CLAUDE.md and ensure AGENTS.md points to CLAUDE.md instead of duplicating volatile paths. At WF5, replace environment placeholders with first runnable environment facts.

update-from-grill

Use immediately after grill_draft_ready or an accepted Grill draft.

  1. Read the three Grill docs and docs/05_intake/Execution_Readiness.json when it exists.
  2. Read existing guidance/state files. Missing PROJECT_STATE.json is not a failure.
  3. Create CLAUDE.md from template if missing; otherwise use ./references/claude-maintenance.md for section-safe edits.
  4. Write only candidate-clear Grill context: current intent, accepted draft status, startup artifacts, candidate datasets/baselines and targets, unresolved questions, falsifiers, claim boundaries, and prepare blockers.
  5. Keep candidate dataset paths and baseline clone targets explicitly labeled candidate until prepare / WF4-WF5 verify them.
  6. Ensure AGENTS.md and README.md exist or point to CLAUDE.md plus Grill artifacts; keep them concise and avoid volatile path duplication.
  7. Do not write .workflow_supervisor/** or .evidence/** by hand, do not mark WF1-WF3 complete, and do not promote Grill draft facts into approved contracts. Do not create PROJECT_STATE.json, project_map.json, or iteration_log.json from update-from-grill.
  8. In update-from-grill, absence of PROJECT_STATE.json, project_map.json, iteration_log.json, or docs/05_intake/Execution_Readiness.json is expected unless another tool already produced them.

deps-changed

Refresh only ## Environment, equivalent to $env-setup refresh.

Write Rules

  • Preserve ## Custom, ## Language Policy, and Ubiquitous Language guidance.
  • Use precise section replacement, not whole-file rewrites.
  • Keep CLAUDE.md line-budget discipline.
  • Report Gate ledger for CLAUDE.md, AGENTS.md, README.md, OPERATOR_CONTEXT.md, dynamic-context directory, or state writes.

Durable Docs Render

temporary drafts.

Signals

GitHub stars
86
Forks
9
Last commit
Aug 2026
Advanced
Catalog kind
skill
Gateway key
init-project-linzhe001
Source
github.com/linzhe001/harness-research