roast-ai-usage

name	roast-ai-usage
description	Use when the user asks to roast, audit, rate, or improve their AI-agent usage, prompt quality, context management, model choice, feature usage, verification habits, handoffs, or local session history.

Roast AI Usage

Audit how the user actually works with coding agents and give a useful, evidence-based roast. Be direct, practical, and mildly sharp, but do not invent evidence.

Ground Rules

• Prefer read-only inspection of local history, project instructions, skills, memory files, agent config, and conversation metadata.
• Use planning or read-only mode when available to avoid accidental writes.
• Do not upload private history or secrets. Redact sensitive values from examples.
• State what you inspected and what was unavailable.
• Distinguish confirmed findings from plausible patterns and weak evidence.
• Technical correctness matters more than sounding clever.

Best-Practice Criteria

Use current primary docs and the active agent's own conventions as the benchmark. Check these areas:

• Prompt clarity: clear task, outcome, constraints, audience, examples, acceptance criteria, and explicit non-goals when needed.
• Context discipline: relevant files and prior decisions are provided or discoverable, stale context is cleared, long dumps are avoided, and persistent rules live in the right place.
• Feature use: planning mode, repo instructions, skills, MCP/connectors, screenshots, attachments, subagents, task plans, browser tools, tests, and handoffs are used when they fit the job.
• Model and effort fit: simple tasks use fast or cheap settings, difficult reasoning or high-risk work uses stronger models or higher effort, and model limitations are acknowledged.
• Verification habits: the agent is asked to inspect evidence, run tests, check diffs, cite sources, reproduce bugs, and stop when assumptions are unverified.
• Handoffs and resumability: long work is captured in durable files, summaries, plans, tasks, or skills instead of relying on chat memory.
• Reuse opportunities: repeated prompts, repeated corrections, workflow rules, and recurring review checks become skills, templates, or project instructions.

Reference anchors:

• OpenAI prompt engineering guide: https://platform.openai.com/docs/guides/prompt-engineering
• OpenAI Codex docs: https://developers.openai.com/codex/
• OpenAI reasoning best practices: https://developers.openai.com/api/docs/guides/reasoning-best-practices
• Anthropic Claude Code best practices: https://www.anthropic.com/engineering/claude-code-best-practices
• Anthropic building effective agents: https://www.anthropic.com/engineering/building-effective-agents

Workflow

1. Identify the agent harness and available evidence sources.
2. Inventory active features: model, effort mode, planning mode, tools, MCP/connectors, skills, repo instructions, memory, and handoff files.
3. Sample local history. Count conversations, sessions, messages, skills, instruction files, and the observed date range when available.
4. Look for repeated patterns instead of isolated mistakes.
5. Score the workflow.
6. Give the roast, then the fixes.

If local history is unavailable, audit the current conversation, visible project instructions, and any user-provided examples. Mark the score as limited.

Rating System

Score each category from 0 to 5:

• Prompt quality
• Context management
• Feature usage
• Model and effort fit
• Verification discipline
• Handoff quality
• Skill and rule extraction

Use this scale:

• 0: harmful or absent
• 1: mostly absent
• 2: inconsistent
• 3: workable
• 4: strong
• 5: excellent and repeatable

For each score, include one short evidence note and one concrete fix. Do not penalize missing features if the current harness does not support them.

Output Shape

Return:

1. Stats: conversations, sessions, messages, skills, instruction files, date range, and source limits.
2. Overall score: total score and one-sentence diagnosis.
3. Roast: short, specific, and evidence-based.
4. Scorecard: category, score, evidence, fix.
5. Top fixes: three to seven concrete changes in priority order.
6. Skill opportunities: repeated behaviors that should become skills, templates, or project rules.
7. Unverified: anything that could not be checked.

Keep the final result useful enough that the user can change their workflow immediately.