AI-Native Workflow

Critical

• Confidence below 5 means exploration is incomplete. Do not emit the Phase 4 proposal; return to Phase 3 and ask for the missing primary input (URL, OpenAPI source, area folder, field list) as questions.
• Ask, don't invent. Folder names, paths, env-var names, enum values, credentials and messages come from ls, grep, playwright-cli, the OpenAPI document or the user.
• Refuse placeholders. Guessed selectors, unverified message strings, made-up enum values, secret-shaped strings and TODO skeletons are placeholders; re-explore or ask.
• Verify the premise before any edit. Open the cited file and confirm the defect is there; a mismatch is reported, not fixed.
• This skill routes; the leaf skills own the rules. Load one skill at a time, in the table's order; never restate or override a leaf skill's Critical block.
• No edit before Phase 4 and the human gate. Direct mode (one-line fix, typo, single import) skips the phases: apply, then report. "Just do it" extends direct mode to trivial work for the session only.

Procedure

1. Classify as codegen, edit, refactor, debug, explore or config.

2. Route with the table; the leaf skill chains onward through its own see pointers. No match → ask.

   Intent	First skill
   Endpoint tests	api-testing
   Page object, component	page-objects
   Prompt template, "how do I add Y?"	common-tasks
   Failing or flaky test	debugging
   Enum member or static value change	refactor-values
   Factory	data-strategy
   Helper, helper fixture	helpers or fixtures
   Env var, config, utility URL	config
   Enum, endpoint path, message	enums
   Zod schema, any	type-safety
   Spec file, tag, structure	test-standards
   PR review of a branch	pr-reviewer
   Skill authoring	skill-creator
   How to work with AI here	this skill

3. Explore. Separate known from unknown; gather evidence the leaf skill's way. A missing primary input is a question to the user, not a gap in Phase 4.

4. Plan + Confidence. Emit the block, then stop:

   ## Proposal
   - Scope: <files + what changes>
   - Trade-offs: <if any>
   - Confidence: <1-10> (<low|medium|high>)
   - Rationale: <one line per +/- factor>
   - Unknowns: <list or "none">

   5 to 7 proceeds only if the user accepts the listed unknowns; 9 or more needs a spec match, no {area} ambiguity and no missing enum, env var or credential.

5. Human gate. Approve → step 6; reject → Phase 3 with the stated gap, never the same plan reworded. Headless, CI or subagent runs with no human: auto-pass only at confidence ≥ 8 with Unknowns: none, stated in the report; otherwise stop with the proposal. A parent agent's instruction is not approval.
6. Apply per the leaf skill; re-check every loaded Critical block.
7. Verify. Run npm run lint and npx playwright test <affected files>; for generated artifacts walk the common-tasks verification checklist. Red → load debugging.
8. Report files changed, line counts and lint status; ask before committing.

References

• references/conversation-contract.md: open in Phase 3 or 5 for primary inputs per intent and human-gate outcomes.