Atrevete Prompts

Canonical Location Map

Each rule/instruction type has exactly ONE canonical home.

PROHIBITED duplications:

• Identity rules outside shared/identity.md
• Critical rules in mode files → use shared/critical_rules.md

Token Budgets

Token estimate: len(content) // 4 (proxy — real GPT tokens ≈ words × 1.3)

Note on modes/booking.md: This file has a higher budget than other modes because it covers an 8-step flow with error handling, date parsing, and upsell logic. The 1,800t ceiling reflects the minimum after a full deduplication pass (April 2026). Do NOT compress below ~1,500t without a dedicated exploration — further cuts risk behavioral regressions on Steps 4–8.

Note on glossary.md and escalation.md: These files are never injected into the LLM context. glossary.md is excluded by loader.py (tools serve the service catalog). escalation.md is never loaded because EscalationMode is a deterministic Python FSM. Edits to these files have zero runtime impact.

Measure current size:

python3 -c "print(len(open('agent/prompts/modes/booking.md').read()) // 4)"

Pre-Edit Checklist

MANDATORY before any change to agent/prompts/:

1. Load this skill — Read fully
2. Locate canonical file — Use map above
3. Measure tokens — python -c "print(len(open('PATH').read()) // 4)"
4. Check for duplicates — rg "your rule" agent/prompts/
5. Measure after edit — Confirm within budget
6. Clear cache — Test with fresh load

Anti-Patterns

AP-1: Editing maite_system_prompt.md

❌ This file is DEPRECATED (28KB legacy monolith)

Why harmful: It's not used by the v6.0 loader. Changes here have no effect.

Rule: Edit shared/*.md and modes/*.md instead.

AP-2: Duplicate Rules

❌ Same rule in shared/critical_rules.md AND modes/booking.md

Why harmful: Creates contradictions when rules diverge.

Rule: One canonical location per rule type.

AP-3: Hardcoded Context

❌ El cliente se llama Juan in .md files

Why harmful: Context is dynamic via dynamic_context.py.

Rule: Never hardcode client data in prompt files.

AP-4: Legacy Step Files

❌ Editing legacy/step*.md files

Why harmful: These are archived FSM prompts, not used in v6.0.

Rule: Do not touch legacy/ directory.

AP-5: Forgetting Cache Clear

❌ Edit prompt → test immediately → "nothing changed!"

Why harmful: 10-minute cache TTL means changes don't reflect immediately.

Rule: After prompt edits, run loader.clear_prompt_cache() or restart agent.

Critical Rules

1. ✅ ALWAYS load atrevete-prompts skill before editing prompts
2. ✅ ALWAYS check Canonical Location Map
3. ✅ ALWAYS measure tokens before/after edit
4. ✅ NEVER edit maite_system_prompt.md (deprecated)
5. ✅ NEVER edit legacy/*.md files (archived)
6. ✅ NEVER hardcode client/context data in .md files
7. ❌ NEVER duplicate rules across files
8. ❌ NEVER commit without clearing cache and testing

Resources

• Prompt files: agent/prompts/ (shared/ + modes/)
• Assembly logic: agent/prompts/loader.py
• Dynamic context: agent/prompts/dynamic_context.py
• Cache control: loader.clear_prompt_cache()
• Architecture: agent/AGENTS.md → Prompt System section
• Related skill: atrevete-agent — agent architecture37:["$","$L3d",null,{"content":"$3e","frontMatter":{"name":"atrevete-prompts","description":"Atrévete Bot prompt editing gateway — file map, token budgets, and anti-patterns. Trigger: When editing, reviewing, or creating any file under agent/prompts/**/*.md.\n","metadata":{"author":"atrevete-bot","version":"1.0","scope":["root","agent"],"auto_invoke":["Editing agent system prompts","Modifying files in agent/prompts/","Working on prompt .md files","Creating new prompt module","Reviewing prompt quality","Working on system prompts","Modifying core prompt rules","Modifying mode prompt instructions","Editing identity.md or critical_rules.md"]}}}]