nova-prompt-author

name

nova-prompt-author

description

Edit the lead agent's system prompt with discipline. Use when the user says "the agent is being too X", "tell the agent to Y", "the prompt needs a rule about Z", or before adding any new rule to the system prompt. The system prompt is at `backend/packages/harness/deerflow/agents/lead_agent/prompt.py` (a single function that returns a `ChatPromptTemplate`). Use this skill BEFORE making the change; the skill walks through the change-and-validate cycle. Triggers on phrases like "edit the agent prompt", "add a rule to the lead agent", "system prompt", "AGENT_MANIFEST", or before merging any change touching `agents/lead_agent/prompt.py`.

nova-prompt-author

The change-and-validate cycle

1. Read the current prompt: wc -l backend/packages/harness/deerflow/agents/lead_agent/prompt.py (should be 500+ lines). Identify the section you're changing (e.g., <skill_persistence>, <command_arsenal>, <self_verify>, etc.).

2. Find or create the test: backend/tests/test_lead_agent_prompt.py (exists per the audit's test inventory). Add a regression assertion: e.g., "the prompt contains <rule>no ENOENT hunts</rule>" so the rule is enforced by the test, not just the prompt.

3. Update the prompt. Constraints (per FORK_V5.md hard rules):

   • Use literal paths (no {...} braces) — braces break .format() (the prompt is .format()-templated).
   • Keep <system-reminder> blocks separate from the static prompt — they're injected at runtime, not at template time.
   • Don't reference deprecated/removed tools (e.g., cowork, the v6 namespace).
   • Don't duplicate the <agent_manifest> content — the manifest is injected separately.

4. Re-run replay E2E: cd backend && PYTHONPATH=. uv run pytest tests/test_replay_golden.py -v. The fixture should still match — if it doesn't, your prompt change drifted the agent graph semantics.

5. Update the test for the new rule so it actually exercises it (not just asserts the rule exists).

Anti-patterns

• ❌ Don't add import statements to the prompt file — it's a function returning a ChatPromptTemplate, not a module.
• ❌ Don't add a rule without a corresponding test. The rule lives in the prompt; the test lives in the code. Both must move together.
• ❌ Don't make the prompt > 2000 lines. Larger prompts slow the agent and dilute the most important rules.
• ❌ Don't reference tool names that are loaded via assemble_deferred_tools — those are HIDDEN from the model until promoted. If you mention tool_search in the prompt, the agent won't find the tool by its deferred name; it finds it via the <available-deferred-tools> block.
• ❌ Don't reword rules without re-running the replay E2E. The fixture is the contract.