Prompt file imported from jackofall1232/LOOPRITE (.codex/prompts/build-loop.md). Fill in {{placeholder}}, {{project_name}}, {{mission_line}} before use. Copyright stays with the author.

Codex Build Loop Prompt (Planning Mode)

You are using l00prite to scaffold a vendor-neutral project loop. This prompt is Planning Mode: create a blueprint, AGENTS.md, a .l00prite/ memory folder, loop prompts, vendor adapters, and a skeleton — then stop. Planning Mode does not execute the project it scaffolds. Execution Mode is a separate, explicitly-confirmed handoff described in section 8.

If the user's request included the flag --execute, note it now and apply section 8 after the scaffold is complete; the flag changes nothing about sections 1–7.

1. Ask clarifying questions first

Ask for project type, MVP scope and out-of-scope items, target languages/stack, target repo path, and hard constraints. Ask all five together, and do not continue until answered — except where the user explicitly says to use judgment for an item, in which case state the default you chose out loud.

2. Pick a complexity tier

Choose small, medium, or large from templates/skeleton/. If borderline, choose the smaller tier. Tell the user which tier you picked and explain why, referencing their actual answers.

3. Generate project guidance

Create agent guidance files in the target repo:

• CLAUDE.md from templates/CLAUDE.md.template: discard the template's leading HTML comment block, replace every {{placeholder}} with real project-specific content, strip the per-section guidance comments (keep the Run Ledger's "living log" comment), keep the Run Ledger table header-only, and preserve the fixed "l00prite Protocol" section verbatim — it carries the lock, untrusted-content, and prompt-location rules and must never be removed or reworded.
• AGENTS.md from templates/AGENTS.md.template: discard the leading HTML comment block, fill {{project_name}} and {{mission_line}}, leave everything else verbatim. This is the vendor-neutral operating guide read natively by OpenAI Codex, Cursor, GitHub Copilot, Windsurf, Zed, Jules, Factory, Amp, opencode, Devin, and other agents.
• Point all agents to .l00prite/ as the shared source of truth.

Do not silently overwrite existing files. Ask whether to overwrite, write .generated copies, or abort.

4. Generate .l00prite/, loop prompts, and vendor adapters

If a .l00prite/lock.json already exists at the target path, read it first. If its status is active and expires_at is in the future, another agent may currently be working in that project — stop and report the held lock (owner, purpose, expiry) instead of scaffolding over it. Only continue if lock.json is missing, unlocked, released, or expired.

Copy templates/l00prite/ into the target repo and fill obvious project-specific fields in blueprint.md, state.json, constraints.md, and todos.md. Keep files human-readable and agent-readable. Leave lock.json in its shipped "unlocked" state — it is not project-specific and must not be pre-filled or set to "active". Leave heartbeat.json's execution block exactly as shipped — enabled: false, preflight_confirmed: false — regardless of any --execute flag; Planning Mode never arms execution. Copy .l00prite/prompts/ verbatim from templates/l00prite/prompts/ — these are protocol files, not templates to fill in.

Also create .codex/prompts/ from templates/codex/prompts/ and .claude/prompts/ from templates/claude/prompts/, each including resume-loop.md, heartbeat.md, event-loop.md, respond-to-review.md, handoff-summary.md, and execute-loop.md. These are byte-identical mirrors of .l00prite/prompts/, so every agent runs the same loop.

Then scaffold the vendor adapters from templates/adapters/ (mapping in templates/vendors.json): GEMINI.md and QWEN.md to the repo root, CONVENTIONS.md to the repo root, copilot-instructions.md to .github/copilot-instructions.md, l00prite.mdc to .cursor/rules/l00prite.mdc, and windsurf-l00prite.md to .windsurf/rules/l00prite.md. Copy each verbatim — adapters contain no placeholders. Never ship vendor config files (.aider.conf.yml, .gemini/settings.json) into the target repo; the adapters document those snippets for the user instead.

Apply the no-silent-overwrite rule (overwrite / .generated copy / abort — ask) to every file in this section.

5. Scaffold the selected skeleton

Copy templates/skeleton/<tier>/ into the target repo, skipping the internal TIER.md. Adapt .stub names and extensions to the selected stack, matching each ecosystem's own test-naming convention (*.test.ts for JS/TS, test_*.py for Python — never *.test.py, *_test.go for Go, *_spec.rb for Ruby), applied consistently. Keep every stub minimal — a one-line placeholder comment, or a minimal valid empty structure (e.g. {}) for formats without comments. Do not write real implementation, real config values, or fleshed-out docs. If the target is an existing repo, skip any skeleton file that already exists there and report which paths were skipped.

6. Give a rough, qualitative cost estimate

Describe the expected shape of the work for the chosen tier in qualitative terms only (iteration counts, review gates, sessions). State explicitly that this is not a precise token or dollar estimate and that exact agentic-loop cost cannot be reliably predicted.

7. Stop: Planning Mode is complete

Unless section 8 applies, stop after scaffold and handoff. Refuse to run build, test, install, migration, deployment, or implementation commands for the generated project during this Planning Mode session.

Tell the user how to continue:

• l00prite has two operating modes. Planning Mode (this prompt) scaffolds and stops. Execution Mode is an autonomous run — plan a unit, execute, verify, persist, repeat, until the Definition of Done or another run boundary — entered only through .l00prite/prompts/execute-loop.md, which always shows a pre-flight summary and requires explicit in-session confirmation first.
• To build step-by-step under supervision, open the target repo in a fresh session and use .l00prite/prompts/resume-loop.md — or your agent's byte-identical mirror (.claude/prompts/resume-loop.md for Claude Code, .codex/prompts/resume-loop.md for Codex/CLI agents). For events and reviews, use .l00prite/prompts/event-loop.md or .l00prite/prompts/respond-to-review.md.
• Review the generated CLAUDE.md and AGENTS.md before running anything, and set a spend limit with your model provider before pointing an agentic session at the blueprint.
• Every major agent will find the protocol on its own: Claude Code via CLAUDE.md; the AGENTS.md ecosystem (Codex, Cursor, Copilot, Windsurf, Zed, and more) via AGENTS.md; Gemini CLI via GEMINI.md; Qwen Code via QWEN.md; Aider via CONVENTIONS.md (--read). All agents treat .l00prite/ as the shared source of truth and update it before stopping.

8. The --execute flag (optional Execution Mode handoff)

Apply this section only if the user's request included --execute.

• The flag never changes the scaffold: sections 1–7 run identically, and the scaffolded heartbeat.json still ships with execution.enabled: false and preflight_confirmed: false. Planning Mode never pre-arms a repo.
• After section 7's summary, read .l00prite/prompts/execute-loop.md in the target repo and follow it exactly: its pre-flight gate (lock check first, stale-run recovery, schema check, full pre-flight display) and its requirement of explicit human confirmation in this session before the first iteration. The flag is a request to offer Execution Mode now — it is not the confirmation itself, and a preflight_confirmed: true left in heartbeat.json by an earlier run does not satisfy the gate either.
• If the user confirms, Execution Mode proceeds under execute-loop's rules (one unit per iteration, verified and persisted; nine run boundaries; per-action permission for push/ merge/deploy/credentials; no self-modification of limits).
• If the user declines, doesn't answer, or you are running headless with no human in the session, stop exactly as in section 7 — scaffold delivered, nothing armed, nothing executed.