b0d790be34
12 unversioned skills now versioned at 1.0.0: agent-communication, ascii-video, external-reasoning-augmentation, jotty-notes-api, minecraft-modpack-server, obsidian, pokemon-player, powerpoint, social-search, songwriting-and-ai-music, workspace-context-organization, youtube-content Total repo: 141 skills across all profile scopes
72 lines
3.2 KiB
Markdown
72 lines
3.2 KiB
Markdown
---
|
|
name: external-reasoning-augmentation
|
|
title: External Reasoning Augmentation
|
|
description: Augment agent reasoning with another AI model when API access is unavailable, using relay mode through the user's existing web access.
|
|
version: 1.0.0
|
|
---
|
|
|
|
# External Reasoning Augmentation
|
|
|
|
## Purpose
|
|
|
|
When the user wants the agent to augment its own reasoning with another AI model (e.g. Grok, Claude, GPT) during building or analysis, but API access is unavailable, unaffordable, or undesired.
|
|
|
|
## Triggers
|
|
|
|
- User says "use X model to help reason about this" or "I want you to use [model] for additional logic"
|
|
- User mentions a specific model they have access to and wants it involved in workflow
|
|
- User asks about integrating a model they have browser/web access to but no API key for
|
|
|
|
## Capability Options (preference order)
|
|
|
|
1. **Relay Mode** (preferred fallback — cost-free, no setup)
|
|
- Default when user has web/browser access and no stated willingness to pay
|
|
- User has web access to the model (e.g. X Premium for Grok, ChatGPT Plus)
|
|
- Agent formats a clear, self-contained question
|
|
- User pastes it into their own browser session, pastes reply back
|
|
- Agent integrates the reasoning into ongoing work
|
|
|
|
2. **API Integration** (only pursue with explicit permission)
|
|
- **ASK FIRST.** Never propose or configure a paid integration without user go-ahead.
|
|
- Only pursue if user explicitly confirms willingness to set up / pay
|
|
- Configure as a custom provider in Hermes
|
|
- Fast, reliable, automatable
|
|
- If user says "I don't have the money" or any cost constraint, STOP — do not pursue
|
|
|
|
3. **Browser Automation** (last resort — generally avoid)
|
|
- Use Playwright MCP or similar to navigate the web interface
|
|
- Requires authentication (cookies, session, or manual login)
|
|
- High failure rate due to anti-bot detection
|
|
- **Only attempt with explicit user permission and awareness of fragility**
|
|
|
|
## Relay Mode — Detailed Workflow
|
|
|
|
When relay mode is established:
|
|
|
|
1. During a complex build, when additional reasoning is needed, agent formats:
|
|
- Context summary (2-3 sentences of what's being built)
|
|
- Specific question for the external model
|
|
- Any constraints or format requests for the answer
|
|
|
|
2. Agent presents it clearly in chat, marked for relay
|
|
3. User opens their browser at the model's URL (e.g. x.com/i/grok)
|
|
4. User pastes the formatted question, waits for answer
|
|
5. User pastes the answer back into chat
|
|
6. Agent incorporates that reasoning into decisions / code / plans
|
|
|
|
## Pitfalls
|
|
|
|
- **Never assume API access** — always ask about API keys before proposing integration
|
|
- **Never attempt browser automation without permission** — anti-bot measures make it unreliable
|
|
- **Don't treat relay mode as slow or annoying** — for many users it's the preferred free approach
|
|
- **Keep questions self-contained** — the user is copy-pasting; links to prior context won't work
|
|
- **Respect "no cloud APIs" policies** — some users are self-hosted only
|
|
|
|
## Memory
|
|
|
|
When relay mode is established with a user, save it as a persistent workflow in the user profile, not as a session detail. Include:
|
|
- Which model the user has access to
|
|
- Which access path (web URL, premium tier, etc.)
|
|
- Whether they have or want an API key
|
|
- That browser automation is NOT attempted
|