AI Native Workflow

AI-Native Workflow

Routing layer between user intent and the specialized skills that own the rules. Load first on every non-trivial task.

For deeper context: see references/three-layer-model.md, references/conversation-contract.md, references/principles.md, references/examples.md, references/troubleshooting.md.

Critical

• Low confidence means Phase 3 is incomplete. Would-be confidence < 5 → do NOT emit Phase 4; return to Phase 3 and ASK the user. Full thresholds live in one place — "Phase 4 — Confidence-Gate Format" below; other docs link there and never restate them. This rule is the most leveraged in the workflow: it stops plausible-looking plans built on guesses.
• Ask, don't invent. Never guess folder names, file paths, env-var names, enum values, credentials, or message strings. ls, grep, playwright-cli, OpenAPI — or ask.
• Refuse placeholders. Guessed selectors, unverified message strings, made-up enum values, secret-shaped strings — refuse and re-explore. TODO / skeleton / "to fill in later" outputs count as placeholders too — offering a "skeleton page object with TODO locators while we wait for playwright-cli" is the same failure mode as inventing locators outright; don't.
• Verify the user's premise even in Direct Mode. Before applying a one-line fix, confirm the reported defect actually exists (the typo on the cited line, the import the user wants removed, the value the user says is currently set). If the premise doesn't match the file, switch out of Direct Mode and ASK — applying a "fix" to a defect that isn't there invents a change.
• Specialized skills own the rules. This skill never restates rules from api-testing, page-objects, etc. It tells you which skill to load and in what order.
• CLAUDE.md Constitution is the safety floor. MUST/SHOULD/WON'T tables are hard stops; they take precedence over any prose, template, or example.
• Audit-then-edit by default. For any non-trivial change, follow the 8-phase workflow below. Phase 4 (Plan + Confidence) is mandatory before Phase 6 (Apply).
• Confidence gate is required. Every Plan output must include a 1-10 confidence + rationale + unknowns block. See "Phase 4" below.
• Exploration is non-negotiable. UI → playwright-cli only (no IDE browser MCP, no Cursor browser, no playwright codegen). API → OpenAPI/docs first, live HTTP only as fallback.
• One skill at a time. Load skills sequentially per the routing table. Don't stack 5 skills' Critical blocks before starting work.
• After any test edit, run the affected tests. On red, load debugging — never suppress, never bump timeouts.

Main Workflow (8 Phases)

Every non-trivial task runs through these phases in order.

Phase 4 — Confidence-Gate Format (mandatory)

Every Plan output before the human gate uses this shape:

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

Confidence rules (canonical home — conversation-contract.md and the Critical block above link here; never restate these thresholds elsewhere):

• < 5 → do NOT emit Phase 4. Exploration is incomplete. Return to Phase 3 and ASK the user for the missing primary input (URL, OpenAPI source, area folder, field list, etc.). Frame the gap as questions, not as a low-confidence proposal — Phase 4 exists for honest trade-off decisions, not for documenting "I don't have enough data".
• 5-7 → emit Phase 4 with explicit unknowns. Proceed only if the human accepts the trade-offs at this confidence level.
• ≥ 8 → full evidence in hand; proceed normally.
• ≥ 9 only if: explicit OpenAPI / spec match, no {area} ambiguity, no missing enum / env / credential, no untested branch, no in-flight conflicting work.
• Rejection by human → re-enter Phase 3 with stated gap. Do not retry Phase 4 with the same plan.