| name | love20-prompts |
| description | Frame LOVE20 requests into precise prompts and response contracts for protocol explanation, contract interaction, state inspection, troubleshooting, selector decoding, code review, and docs or skill updates. Use when a LOVE20 ask is underspecified, when you want a reusable prompt for another AI agent, or when a broad task should be split into smaller prompts with the right evidence requirements and companion skills. |
LOVE20 Prompts
Use this skill to turn a vague LOVE20 request into a prompt that is specific enough for another agent to execute correctly.
Path Convention
- Cross-repo references use canonical GitHub repo names:
docs, core, periphery, script, interface, extension, extension-lp, extension-group, group, group-chat.
- If local checkout names differ, map local aliases to these canonical names before following any path.
Workflow
- Read
references/prompt-selection.md.
- Choose one prompt family:
references/protocol-and-architecture-prompts.md
references/interaction-state-and-debug-prompts.md
references/review-and-doc-sync-prompts.md
- Read
references/junior-agent-acceptance-cases.md when you need reusable eval or smoke-test questions for another agent.
- Use
assets/junior-agent-eval-handbook.md and assets/junior-agent-scorecard.csv when you want a ready-to-run manual evaluation pack for other models.
- Copy the closest template and replace placeholders such as token, round, action, account, network, file, or symptom.
- Keep only the companion LOVE20 skills that the task actually needs.
- If one ask mixes explanation, execution, and debugging, split it into multiple prompts.
Working Rules
- Every prompt should specify:
- objective
- entity anchors such as token, action, round, account, network
- source-of-truth requirements
- expected output shape
- conflict handling between docs and code
- Ask for file paths, contract names, function names, or script paths instead of generic descriptions.
- Prefer prompts that force the agent to separate:
- protocol intent from docs
- actual behavior from deployed contract repos (
core, extension, extension-lp, extension-group, group, group-chat)
- convenience wrappers from core rules
- current state from indexed history
- For prompts aimed at junior or generic agents, force a triage step that classifies:
- base vs extension or group
- business round vs contract-local round
- current state vs history
- protocol truth vs adapter behavior
- If the prompt uses unqualified
action or 行动, explicitly say that the term includes extension-backed and group-backed actions unless the task is intentionally scoped to base/core only.
- When validating another agent after a LOVE20 docs or skill update, reuse the acceptance cases in
references/junior-agent-acceptance-cases.md instead of inventing new eval questions every time.
- When the task could touch immutable on-chain behavior, require the prompt to treat
periphery, script, and interface as adapters, examples, or read models rather than the final behavior authority.
Guardrails
- Do not ask one prompt to both teach the protocol and produce exact write calls unless the user explicitly wants both.
- Do not request debugging help without providing the best available symptom:
revert string, selector, topic, Chinese UI message, tx hash, or screenshot path.
- Do not request state inspection without naming at least one anchor if it is known:
token, symbol, actionId, round, account, or network.
- For review prompts, require findings first and force the agent to cite files and lines.
- For docs or skill update prompts, require the agent to validate and mention what was regenerated.
References
references/prompt-selection.md
references/protocol-and-architecture-prompts.md
references/interaction-state-and-debug-prompts.md
references/review-and-doc-sync-prompts.md
references/junior-agent-acceptance-cases.md