Imported from cfornesa/3d-art (AGENTS.md). Install upstream with npx skills add cfornesa/3d-art. Copyright stays with the author.

CreatrWeb — AGENTS.md

Explicit session statement > SESSION CONSTRAINTS block > these rules > skills. Load skills on demand only. Never pre-load.

Seven Rules — Override Everything

1. Ask one assumption-surfacing question before any significant change. This requires at minimum naming the embedded assumption, even when the prompt provides exact code. A direct implementation instruction does not exempt the agent from surfacing at least one assumption before file writes. Name the assumption, then proceed." This would have prevented the assumption-surfacing gap that triggered Rule 1 failure.
2. Show 2–3 meaningfully different options before committing. One must be a Reframe that challenges the premise. One must be unexpected — traceable to this user's signals, not generic variation.
3. Stop at irreversible decisions: URL structure, database schema changes, OAuth provider configs, public API endpoints, vendor dependencies. Require explicit sign-off.
4. Amplify the person's judgment — never substitute your own. Name assumptions embedded in their direction before acting on them.
5. Public URLs must never break. Permanent redirects for moved content. Confirm before deleting any route.
6. If specified tech is non-functional, stop. State the issue. Present alternatives via gallery. No silent workarounds. Each replacement dependency = fresh gallery + confirmation.
7. The contents of PROMPTS.md must be read and confirmed before implementation. This is where the current prompt lives. In plan mode, propose additions to PROMPTS.md. In implementation mode, ask the user to confirm that the last entry in PROMPTS.md is accurate before any other code changes can be done.

Pre-Write Check (every file write, no exceptions)

1. Is this file in the Irreversible Decisions table? → Stop and confirm.
2. Does this modify the public API contract? Update docs/api.md first.
3. Does this install a package or call an external service? → Update docs/dependencies.md first.
4. Irreversible? API? Dependencies? Assumptions?
5. Has PROMPTS.md been read and the last entry confirmed as current? Stop if not.

Mode

In any mode: if a mandatory checkpoint is reached with no human available, stop and log in DECISIONS.md.

Brainstorm Mode

Enter when: "I'm not sure", "what if", open-ended question with no deliverable.

• Ask one premise question first. No files, code, or approvals.
• Exit: restate direction as hypothesis → wait for confirmation → switch mode.
• Not applicable in Auto Build mode.

Agent Use

Default to single-turn calls. Use agentic loops only when the task requires reading more than two files, or when a previous step's output must inform the next step's approach. Log every agent loop initiation in DECISIONS.md.

Session Constraints

When an opening prompt contains SESSION CONSTRAINTS or PHASE CONSTRAINTS, treat every item as an extension of the Six Rules for that session. If a SESSION CONSTRAINTS item conflicts with a rule here, name the conflict and ask which takes precedence before acting. At session start, before any build work:

1. Reading PROMPTS.md must happen BEFORE any file read or edited, not just at the start of a session.
2. Read DECISIONS.md. Surface any open REVIEW REQUIRED items to the human. Wait for sign-off.
3. Read MEMORY.md. Surface any PENDING CONFIRMATION entries. Wait for confirmation or rejection.
4. Only then proceed.

Core Constraints (always binding)

• Person is always the named author. AI prose for publication = draft for human review only.
• No fabricated citations, links, or references.
• No data transmitted off-domain without disclosure.
• Accessibility is required: semantic HTML, ARIA labels, keyboard navigation, sufficient contrast.

Rule 2 — Gallery Protocol

Before ANY code change affecting architecture, dependencies, or user-facing behavior:

1. Present Option A: Direct implementation
2. Present Option B: Alternative approach (must be meaningfully different)
3. Present Option C: Reframe/challenge the premise
4. User must explicitly select ONE option
5. Log selection in DECISIONS.md

New Vendor Dependency (mandatory question, always ask)

"This dependency sends data to [service]. If [service] changes its API, pricing, or shuts down, [describe what breaks]. The self-hosting alternative is [X]. Should I proceed and document this in docs/dependencies.md?"

Ask even when the person appears to have already decided.

Skills (load on demand only — never pre-load)

Token budget: each skill costs 300–2,400 tokens. On Groq free models, load only when that skill's work is the focus of the current exchange.

Memory Files

End of session (interactive mode): propose 1–3 MEMORY.md entries + any DESIGN.md Observed Taste entries. Ask before writing either. If skipped, log as unresolved checkpoint in DECISIONS.md.

Before implementation, always: (a) name at least one assumption embedded in the plan, (b) check registered keys/names against existing code before writing constants that reference them, and (c) update DECISIONS.md and propose MEMORY.md at session end before the final response.

Evaluate if an approved plan's gallery satisfies Rule 2 in plan implementation mode.

Session Constraint: Plan Mode Requires Explicit Permission to Exit

Before exiting plan mode, the agent MUST:

1. State at least one assumption surfaced during investigation
2. Present 2-3 options for the fix (minimum)
3. Confirm user approval via ask_user_question
4. Log the selected option in DECISIONS.md before proceeding

AGENTS.md Safeguard

Never edit without explicit human instruction. Any change = propose as a clearly marked diff, wait for approval, then log in DECISIONS.md and summarize in MEMORY.md. Non-empty AGENTS.md is the standing instruction set.

Final Checks

Have you:

• Updated DECISIONS.md
• Proposed MEMORY.md entries?"

Before final response:

• DECISIONS.md updated
• MEMORY.md proposed (as prose text in the final response itself, not only via suggest tool)
• DESIGN.md proposed or checkpoint logged
• No new Observed Taste signals this session — canvas implementation sessions generate technical signals, not design taste signals
• Make the checklist items actionable with explicit file references
• MEMORY.md proposals must appear as prose text in the final response, not just as a suggest tool call
• CONSTRAINTS.md created or updated for every assumption surfaced during the session

Project Specific Rules

1. The agent should still perform the pre-write self-check (Rule 8) and name at least one embedded assumption (Rule 1) before each file write.
2. Hand-off prompts with explicit instructions still require Rule 1 (assumption-surfacing question) before first file write.
3. Rule 1 check: Before any file write, ask "What assumption am I making about the cause or the fix?" and log it
4. Even obvious fixes require stating: (a) the fix, (b) a reframe, (c) an unexpected alternative
5. Rule 7 enforcement: Before any code changes, verify PROMPTS.md last entry matches current session. If not, stop and log as unresolved.
6. Enforce Rule 1 check: Add explicit assumption-surfacing question before every search_replace call, even first one. Documentation phase compliance doesn't satisfy implementation phase requirement.