Ask-More

Consult multiple LLMs in parallel, then merge answers into a structured diff report: consensus, unique insights, conflicts, and actionable next steps.

Configuration

Config file: INLINECODE0

If missing, copy from skills/ask-more/config.example.yaml and guide user through setup (see § First-Use Setup).

Required

• - models — list of model IDs (≥2), must be providers configured in the gateway

Optional

• - presets — named model groups (see § Presets)
• INLINECODE4 — separate model for diff merge (must NOT be in the models pool)
• INLINECODE5 — log run results to INLINECODE6

Validation

Run bash skills/ask-more/scripts/load-config.sh skills/ask-more to validate config. It uses proper YAML parsing and checks: model count, preset validity, synthesis_model conflicts.

First-Use Setup

When /ask is triggered but no models configured:

1. 1. List providers already in the gateway
2. Suggest 2-3 diverse models from different providers
3. Offer to write config.yaml for them
4. Do NOT block — get them running in one interaction

Commands

Command	Description
INLINECODE9	Consult all configured models
INLINECODE10

Consult on the most recent conversation topic | | /ask --preset <name> [question] | Use a named preset model group | | /ask --compare [question] | Raw compare mode: show responses side-by-side, no synthesis |

Presets

If presets is configured in config.yaml, users can select a named group:

CODEBLOCK0

When --preset <name> is used, override the default models with the preset's list for that run. If the preset doesn't exist, list available presets and abort.

Workflow

1. Parse trigger

• - /ask <question> → use the provided question
• INLINECODE17 (no args) → use the most recent user question/topic from conversation. If the recent topic is ambiguous, ask the user to clarify instead of guessing.
• INLINECODE18 → use preset model group
• INLINECODE19 → raw compare mode (skip step 6, go directly to step 7b)

Pre-flight check: If the question is trivially simple or empty (e.g., /ask hi, /ask ok), warn the user that multi-model consultation may not add value for this type of question, and confirm they want to proceed.

2. Privacy check (first use only)

Read config. If privacy_acknowledged is false:

⚠️ 隐私提示： ask-more 会将问题摘要发送给您配置的多个模型 provider（如 OpenAI、Google、Anthropic 等）。请勿在涉及高度敏感信息的场景使用。
继续使用即表示您已知悉。

Set privacy_acknowledged: true in config and proceed.

3. Model availability pre-check

Before context packing, verify that configured models are likely reachable:

• - Check that each model ID exists in the gateway's provider config
• If any model is not found, warn the user and suggest removing it
• If fewer than min_models are available, abort early with a clear message

This prevents users from waiting through the full flow only to discover models are misconfigured.

4. Context Packing

Rewrite the user's question into a self-contained description:

• - Summarize relevant background (what was discussed, constraints, requirements)
• Replace all pronouns/references ("this", "that plan") with concrete descriptions
• State the core question explicitly
• Write neutrally — do NOT inject the primary model's own opinion or framing
• Explicitly note any assumptions you introduced that weren't in the original conversation

Summary mode (default, context_mode: summary):

• - Distill to 500-1000 words

Full mode (context_mode: full):

• - Include the last N turns of raw conversation verbatim (N = full_mode_max_turns)
• Prepend a 2-3 sentence summary of what the conversation is about
• Append the explicit question at the end

5. Confirmation (editable)

Present to user:

CODEBLOCK1

Wait for user response:

• - Confirm → proceed
• User sends modified text → use their version, re-confirm
• Cancel → abort

Dedup check: Hash the packed question. If an identical question was asked in the last 5 minutes, warn user and confirm they want to re-run (saves cost on accidental double-triggers).

6. Parallel consultation

For each model, spawn a subagent:

CODEBLOCK2

Progress feedback: After spawning, show status. As each model completes, update:
CODEBLOCK3

State tracking per model:

• - pending → spawned, waiting
• INLINECODE30 → response received
• INLINECODE31 → API error or malformed response
• INLINECODE32 → exceeded timeout_seconds

After all models complete or time out, call sessions_yield() to collect results.

Failure handling:

• - If a model returns an error → mark as failed, log reason
• If a model returns an extremely short response (<50 words) or a refusal → mark as degenerate, include in report metadata but exclude from synthesis
• If fewer than min_models succeeded (non-degenerate) → inform user, offer to show whatever raw responses are available, abort synthesis

7a. Diff merge (normal mode)

Read the merge prompt from references/prompt-templates.md § "Diff Merge Prompt".

If synthesis_model is configured, use it for this step. Verify synthesis_model is NOT in the consultation pool — if it is, warn and fall back to primary model.

Degenerate response handling in merge prompt:

• - If a model refused to answer, note it as "[Model] declined to answer (safety/policy)"
• If a model gave an extremely short/generic response, note it as "[Model] provided limited input"
• These should appear in the report metadata, not pollute the consensus/diff analysis

Unanimous agreement shortcut:
If all models give substantially the same answer with no meaningful unique insights or conflicts, use a shorter output:
CODEBLOCK4

Normal output structure:
CODEBLOCK5

Chat interface adaptation: In Telegram/Discord/WhatsApp, deliver a 3-5 bullet summary first, then the full report. Avoid wall-of-text.

7b. Raw compare mode (--compare)

Skip synthesis entirely. Present each model's response in sequence:

CODEBLOCK6

This mode is for users who want to judge themselves without synthesis bias.

8. Graceful degradation

When things go wrong, degrade gracefully instead of failing silently:

Scenario	Action
Synthesis model fails	Fall back to primary model for merge. If that also fails, return raw responses (like --compare mode) with a note.
Only min_models barely respond

Proceed but mark report as "⚠️ 低置信度：仅 N 个模型回复" |
| Only 1 model responds | Abort synthesis, show the single response with note: "仅收到 1 个回复，无法进行多模型对比" |
| All models time out | Abort, suggest user check model config and try again |
| Model returns refusal | Exclude from synthesis, note in report as policy difference |
| Model returns gibberish/extremely short | Exclude from synthesis, note as degenerate response |

9. Logging

If enable_logging: true in config, after each run, log to skills/ask-more/logs/runs.jsonl:

CODEBLOCK7

Use bash skills/ask-more/scripts/log-run.sh <skill-dir> '<json>' to append.

Prompt Engineering Notes

Sub-model prompt design

The subagent prompt (in references/prompt-templates.md) explicitly:

• - Structures output as: assumptions → analysis → risks → recommendations
• Encourages challenging the question's premises — don't just answer, check if the question itself has flawed assumptions
• Uses soft word limit (~800 words) without hard truncation

Context packing discipline

• - Write neutrally — no opinion injection
• Explicitly list any assumptions introduced by packing
• User can edit the packed question — this is a core feature, not an afterthought

Merge prompt discipline

• - Sections 1-3 (consensus/diff/conflicts): faithful to model outputs, no added analysis
• Section 4 (synthesis): clearly labeled as synthesis, not truth
• Distinguish true conflicts from mere phrasing differences
• Include uncertainty labels: 高一致性 / 假设敏感 / 证据薄弱 / 价值判断分歧
• Cite which model supports each claim
• Handle degenerate/refusal responses explicitly
• For high-stakes conflicts: escalate with warning instead of resolving

Extensibility

Adding new modes

The architecture supports pluggable consultation modes beyond normal and compare:

• - Debate mode — models critique each other's responses in rounds
• Red-team mode — one model defends, others attack
• Domain-specific merge — different prompt templates for coding / product / legal / research questions

To add a mode: create a new prompt template in references/, add a command flag, and branch in step 7.

Model metadata (future)

For smarter model selection, config could include per-model metadata: CODEBLOCK8