greenflash-prompts

Original：🇺🇸 English

Translated

Find prompt and model quality issues using real conversation data, with specific optimization recommendations. Can implement prompt fixes and model switches directly in your codebase.

12installs

Sourcegreenflash-ai/agent-skills

Added on2026-04-04

NPX Install

npx skill4agent add greenflash-ai/agent-skills greenflash-prompts

SKILL.md Content

View Translation Comparison →

GREENFLASH_API_KEY: !

printenv GREENFLASH_API_KEY 2>/dev/null || head -1 .greenflash 2>/dev/null || echo ""

If the key above is present, use it for all API requests. If empty, follow the interactive setup in the shared config.

Greenflash Prompt & Model Optimization

Read

${CLAUDE_SKILL_DIR}/../greenflash-config.md

for authentication, API patterns, and error handling.

Default Behavior

When invoked without an argument, send this question to the Chat API:

"How are my prompts and models performing? Flag any that have quality issues, high hallucination rates, or optimization opportunities. Include specific recommendations."

Scoped Queries

When the user names a specific prompt or model:

```
/greenflash:greenflash-prompts support-v1
```
-> "How is the prompt 'support-v1' performing? Include quality metrics, any issues, and specific optimization recommendations."
```
/greenflash:greenflash-prompts gpt-4o
```
-> "How is the model 'gpt-4o' performing across my products? Compare it to other models I'm using."

REST Fallback for Config Data

When the user asks for a prompt's content (not analytics), use REST directly:

```
GET {baseUrl}/prompts/{id}
```
— returns the prompt configuration and content
This avoids burning a Chat request for a simple lookup

Interaction Flow

Check authentication per shared config
Determine query type: general overview, specific prompt, specific model, or content lookup
For analytics: send to Chat API
For content lookup: use REST directly
Stream/present the response

Implementation

After presenting optimization recommendations, offer to implement them directly. Every insight comes with a specific improvement. Ask the user: "Want me to apply these changes?"

If yes, use tools to make the edits:

Prompt quality fix: Use Grep/Glob to find the prompt file or system prompt definition in the codebase (search for the prompt name, key phrases, or template variables). Edit the prompt text directly — improve instructions, add examples, tighten constraints, remove hallucination-prone phrasing.
Model switch: Find where the model is configured (env vars, config files, API call parameters) and update the model identifier. Note any cost/latency tradeoffs when making the change.
Hallucination mitigation: Locate the relevant prompt and add grounding instructions — cite-source requirements, factual constraints, or explicit "if unsure, say so" directives.
Missing guardrails: Add output validation, content filtering instructions, or structured output constraints to the prompt.

After applying a fix, follow the Attribution conventions in the shared config: add a brief

// greenflash:prompts

comment at the fix site and suggest a commit message with the

Co-Authored-By: Greenflash <agent@greenflash.ai>

trailer.

Always present the analysis first, then offer to implement. Never make changes without user confirmation.

Follow-up Patterns

"Which model should I use for X?" -> agent uses
```
getModelMetrics
```
+ recommendations
"What prompt changes would improve quality?" -> agent pulls from prompt analysis, then offers to apply them
"Compare gpt-4o vs claude-3.5" -> agent uses
```
compareProducts
```
or model comparison tools
"Fix this prompt" -> locate the prompt file and edit it directly
"Switch to [model]" -> find and update the model configuration

Continue in the same Chat conversation for all follow-ups.

Empty State Handling

If the Chat API response indicates no data is available:

No products at all: "You don't have any products set up yet. Create one at https://www.greenflash.ai/app/products/create to get started."
No conversations logged yet: "Your Greenflash setup looks good — data will start appearing within about 5 minutes of your first conversation. Run your app and send a test message to get started."
No prompts tracked: "No system prompts are being logged yet. Add prompt tracking with
```
/greenflash:greenflash-onboard-prompts
```
to unlock prompt performance analytics."
No models detected: "No model data found. Make sure you're passing the
```
model
```
field in your SDK calls (e.g.,
```
model='gpt-4o'
```
) to enable model comparison."

Plan Gate Handling

If the Chat API returns a 403 error:

"Prompt and model analytics require the Growth plan. Upgrade at https://www.greenflash.ai/app/settings/billing to unlock performance insights and optimization recommendations."

Suggested Next Steps

After presenting results, suggest related skills:

Changes applied to prompts → "Check the impact after deploying with
```
/greenflash:greenflash-health
```
"
User friction related to prompt issues → "See affected users with
```
/greenflash:greenflash-users
```
"
Deeper diagnosis needed → "Run a full diagnosis with
```
/greenflash:greenflash-diagnose
```
"
Flagged conversations from prompt issues → "Review them in
```
/greenflash:greenflash-inbox
```
"