ralph（Ralph 助手）

ralph — Specification-First Development + Persistent Completion

Stop prompting. Start specifying.

ralph is the portable method contract for the Ouroboros loop:

Interview → Seed → Execute → Evaluate → Evolve
                     ↓
                 ooo ralph
             (persist until verified)

Keep ralph focused on that contract. Runtime-specific execution belongs to sibling skills:

• omc — Claude-native orchestration
• omx — Codex-native orchestration
• ohmg — Gemini / Antigravity harness projection
• jeo — integrated delivery ledger and review loop
• ralphmode — approvals / sandbox / trust posture

When to use this skill

• The user has a vague request and needs a Socratic interview before code is written.
• Requirements should become an immutable seed/spec before implementation starts.
• The task needs a verify-before-done loop instead of a one-shot answer.
• The user wants to keep going until completion is actually verified.
• You need to measure drift against the original contract.
• Repeated failures mean you need a structured unstuck step instead of more blind retries.

Do not use this skill when

• The task is mainly about Claude / Codex / Gemini runtime setup or platform commands → route to omc, omx, or ohmg.
• The main problem is permission posture, approvals, trust folders, or YOLO mode → route to ralphmode.
• The user needs integrated project ledgers, plan review, browser verification, and cleanup workflow → route to jeo.
• The task is only pre-implementation landscape research → route to survey.

Instructions

Follow the phases in order unless the user explicitly asks for a narrower command such as ooo interview, ooo evaluate, or ooo unstuck. Keep ralph as the method anchor, then route platform-native runtime details outward when the request turns into Claude / Codex / Gemini-specific setup or execution policy.

Core contract

1. Interview

Ask questions until ambiguity is low enough to act.

Use when the user is unclear, underspecified, or mixing goals, constraints, and success criteria.

ooo interview "build a task management CLI"

Ambiguity gate

• Greenfield: Goal 40% + Constraints 30% + Success 30%
• Brownfield: Goal 35% + Constraints 25% + Success 25% + Context 15%
• Do not move to seed until Ambiguity ≤ 0.2.

2. Seed

Freeze the clarified request into an immutable spec.

ooo seed

The seed should lock:

• goal
• constraints
• acceptance criteria
• ontology / core object model

Once generated, the seed becomes the baseline for drift checks.

3. Execute

Run the work against the seed instead of improvising from chat drift.

ooo run [seed.yaml]

Default execution shape:

1. Discover
2. Define
3. Design
4. Deliver

4. Evaluate

Verify before claiming success.

ooo evaluate <session_id>

Use the three-stage gate:

• Mechanical — lint, tests, build, typecheck, coverage
• Semantic — acceptance criteria and goal alignment
• Consensus — optional multi-model or reviewer agreement when needed

Drift thresholds

• 0.00–0.15 excellent
• 0.15–0.30 acceptable but watch closely
• 0.30+ correct course before continuing

5. Evolve

Use when the ontology or solution shape is still changing.

ooo evolve "topic"
ooo evolve "topic" --no-execute

Stop when repeated generations converge strongly enough to count as stable.

6. Persistent completion (ooo ralph)

Use when the user wants a loop that should continue until completion is verified.

ooo ralph "fix all failing tests"
ooo ralph "implement the payment module"

The loop contract is:

1. execute
2. verify
3. record failure evidence
4. adjust
5. continue until verified or capped

Keep the stopping rule explicit. A completion promise such as <promise>DONE</promise> is a useful runtime-level signal, but the real rule is verified completion, not merely emitting the token.

Unstuck mode

When repeated attempts are failing, route to a deliberate mindset instead of retrying blindly.

ooo unstuck
ooo unstuck simplifier
ooo unstuck hacker
ooo unstuck contrarian
ooo unstuck researcher
ooo unstuck architect

Quick routing guide:

• repeated similar failures → contrarian
• too many options / paralysis → simplifier
• missing evidence → researcher
• need momentum → hacker
• wrong foundation → architect

Operating rules

1. Clarify before coding when ambiguity is still high.
2. Freeze the seed before implementation; do not rewrite the contract mid-run.
3. Measure drift against the original seed, not against the latest rationale.
4. Verify before done — tests, checks, and acceptance criteria matter more than confidence.
5. Treat failure as data — every failed loop should feed the next attempt.
6. Keep runtime ownership separate — platform hooks, approvals, trust settings, and shell-specific commands belong in runtime docs and sibling skills.
7. Prefer the live local skill path in setup examples: .agent-skills/ralph/....

Platform boundary

ralph is runtime-agnostic on purpose. It should describe what the loop needs, not every platform's syntax.

• Claude Code can supply rich native hooks and orchestration through omc.
• Codex now has native experimental hooks plus .rules, sandbox, and approval controls, but concrete Codex runtime guidance still belongs in omx and the platform references.
• Gemini supports hook-based continuation and trust-folder controls; concrete setup belongs in ohmg and the platform references.
• ralphmode owns safety posture, not ralph.

See references/runtime-boundaries.md and references/platform-setup.md.

Examples

Example 1: clarify first

ooo interview "build a task management CLI"

Example 2: persistent verified loop

ooo ralph "fix all failing tests" --max-iterations=10

Example 3: local helper setup

bash .agent-skills/ralph/scripts/setup-codex-hook.sh
bash .agent-skills/ralph/scripts/setup-gemini-hook.sh

Available scripts

Best practices

1. Keep ralph as the method anchor, not a catch-all platform runtime wrapper.
2. Pair long-running runs with explicit repo-local verification commands.
3. Route runtime shells and orchestration depth outward instead of duplicating them here.
4. Update compact/catalog/doc surfaces whenever the trigger surface changes materially.
5. If a runtime doc drifts from current official docs, correct the runtime reference instead of bloating this front door.

References

• references/ouroboros-commands.md — command syntax, parameters, and state concepts
• references/nine-minds.md — the nine persona roles and unstuck routing
• references/platform-setup.md — platform-specific local setup and current fallback guidance
• references/runtime-boundaries.md — settings / rules / hooks split and sibling-skill ownership

Source: Q00/ouroboros — MIT License