cricd-prompt-standard

name	cricd-prompt-standard
description	The CRICD framework for structuring sub-agent task prompts — Context, Relevance, Instruction, Constraints, Demonstration. Produces dramatically better output from any LLM sub-agent. Use when spawning sub-agents, writing task prompts, or structuring complex instructions for AI models.
metadata	{"emoji":"📐","category":"model-behavior","platform":"any"}

CRICD Prompt Standard

A framework for structuring sub-agent task prompts that consistently produces better output than unstructured instructions. Adopted from EgoAlpha prompt-in-context-learning research + Eric (@outsource_) cognitive framework.

The Framework

Every sub-agent task prompt should include:

Component	What it is	Example
Context	Background, what this is, why it matters	"We're migrating from Postgres to Supabase. 200 tables, 3 services depend on it."
Relevance	Specific files, data, prior decisions to reference	"Schema at db/schema.sql. Migration plan in docs/migration.md."
Instruction	Exact steps, what to do, in order	"1. Audit schema for Supabase incompatibilities. 2. Write migration script. 3. Test against staging."
Constraints	What NOT to do, limits, safety rails	"Do NOT drop tables. Do NOT modify the auth schema. Max 500 lines per migration file."
Demonstration	One compact example of 10/10 output	"Example migration file: ..."

Minimum Viable Prompt

Even simple tasks benefit from C + I minimum:

Context: The gateway crashed after a config change.
Instruction: Check ~/.openclaw/openclaw.json for syntax errors, fix them, restart gateway.

Full Example

## Context
We're writing a weekly cost report script for OpenClaw. It needs to parse JSONL session
logs, aggregate per-model costs, and produce a formatted summary.

## Relevance
- Session logs at: ~/.openclaw/agents/main/sessions/*.jsonl
- Each line has: {"message": {"model": "...", "usage": {"cost": {"total": 0.001}}}}
- Prior version: scripts/old_report.py (broken, don't reuse — uses deprecated API)

## Instruction
1. Scan all .jsonl files for entries in the last 7 days
2. Aggregate cost by model and by day
3. Print a formatted report with totals, per-model breakdown, and daily trend
4. Handle missing/malformed entries gracefully (skip, don't crash)

## Constraints
- Python stdlib only (no pandas, no external deps)
- Must complete in <5 seconds on 500MB of logs
- Do NOT read entire files into memory — stream line by line
- Output must be WhatsApp-safe (no markdown tables, use plain text alignment)

## Demonstration
Good output looks like:
📊 Weekly API Cost Report
Period: last 7 days

Total: $12.4521 across 8,432 turns

By model:
  claude-sonnet-4-5: $8.2103 (65.9%)
  claude-haiku-4-5:  $3.1204 (25.1%)
  ...

Output Quality Injections

Always inject these into sub-agent tasks alongside CRICD:

1. Voice Persona

Set the tone: "pragmatic startup CTO", "senior engineer", "concise analyst". Never "helpful assistant".

2. Output Format Rules

• Tables for comparisons and risks: Risk | Likelihood | Mitigation
• Word budget matched to complexity — one-liner for simple, detailed for complex
• Code > description. Diff > explanation. Numbers > vibes.

3. End Plans with a Key Principle

One sentence capturing the core insight of the plan.

4. Domain Knowledge

Inject specific mechanics. Don't make the model guess at your stack, conventions, or constraints.

5. Anti-Patterns (never)

• "Great question!" / "I'd be happy to help!" / "Certainly!"
• Listing options without recommending one
• Explaining what you're about to do instead of doing it
• Summarizing what you just did after doing it
• Asking permission for things you can safely try yourself

When to Use the Full Framework

Complexity	Use
Typo fix, simple edit	C + I only
Multi-file feature	Full CRICD
Strategy, roadmap, architecture	Full CRICD + Demonstration mandatory
Money involved, public publishing	Full CRICD + critique loop (see critique-loop-protocol)

Research Background

• CRICD structure draws from EgoAlpha's prompt-in-context-learning research
• The Demonstration component is the highest-leverage addition — showing one example of excellent output anchors model behavior more effectively than pages of instructions
• Anti-patterns list derived from production observation of failure modes across Claude, GPT-4, and Gemini