One Soul, Many Minds: Model-Specific Prompt Architecture for AI Agents

GPT-5.4 is nearly 2x more verbose than Opus on identical tasks. Here's how we built model-specific overlays to fix that — and why your agent should do the same.

Listen to this post
00:00

We run Ada across multiple foundation models: OpenAI’s GPT-5.4, Anthropic’s Claude Opus 4.6, Sonnet 4.6, Google Gemini, and several others. Same personality. Same rules. Same soul.

 

GPT-5.4 is nearly twice as verbose as Opus on identical tasks.

 

We measured it. Same five tasks, same system prompt, same persona. GPT-5.4 returned 462 words where Opus returned 237. A 1.95x verbosity ratio. Sonnet came in at 1.83x.

 

Both models got the right answers. The problem is behavioral. GPT-5.4 lists 15 items when 8 covers it. It hedges decisions with conditional trees. It adds a third metaphor when one landed fine. Opus states its answer and moves on.

 

If you’re building an AI agent that runs on multiple models, treating them all the same is leaving performance on the table.

 

Where This Started

Matt Berman talked about this on his YouTube channel: maintaining different system files for GPT and Anthropic models, because they’re trained differently and each company publishes different prompting guides.

 

His idea: download each provider’s official prompting best practices, then use those guides to create optimized versions of your agent’s prompts per model. Not separate personalities. Optimized execution for the same personality.

 

We took this and built it into a three-layer architecture.

 

Layer 1: Core Soul

One file. Identity, personality, approvals, style rules. This is who Ada is. Same across every model.

 

# SOUL.md — Ada (v3.0.0)

## Mission
Make Henry dangerously effective — and have fun doing it.

## Rules
1. Start working. No filler, no warm-up, no "Great question!"
2. Lead with the answer. Context after.
...
11. Do, don't propose. "Show me X" = go get X.

 

Rule 11 exists because GPT-5.4 had a failure mode where it would describe what it could do instead of doing it. Opus never had this problem. But the rule belongs in the shared soul because it’s identity-level. Ada acts, regardless of which model powers her.

 

Layer 2: Model Overlays

Small files loaded conditionally based on which model is running. These compensate for observed failure modes, not theoretical ones.

 

GPT-5.4 overlay targets:

  • Sycophancy (“Great point!”, “Excellent question!”)
  • Verbosity (15 items when 8 suffices)
  • See our OpenClaw GPT-5.4 / Codex Execution Overlay Gist for the exact rules.
  • Hedging (conditional decision trees instead of clear answers)
  • Plan-looping (restating the brief 6 times without executing)
  • Proposing instead of doing

 

Opus 4.6 overlay targets:

  • Over-qualification (every statement gets a caveat)
  • Essay mode (natural output is ~2x what’s needed)
  • Excess caution (asking permission for things already approved)

 

The overlays are about 40 lines each. They reference the provider guides but don’t duplicate them.

 

Layer 3: Provider Prompting Guides

Full best-practice guides pulled from each provider’s documentation: OpenAI’s prompt engineering guide, Anthropic’s Claude prompting best practices, Google’s Gemini prompting guide.

 

These aren’t loaded every session. That would waste tokens. They’re consulted when editing prompts or skills, so the edits follow each provider’s recommendations.

 

The Benchmark

We ran a verbosity benchmark across three models. Five identical tasks covering different response types:

 

  1. Quick question — “What’s the difference between a cron job and a heartbeat?”
  2. Ops task — “List files in this directory”
  3. Research — “What are the main failure modes of LLMs as coding agents?”
  4. Decision — “Meeting at 3pm, flight at 5pm. Taxi or Uber?”
  5. Code — “Write a bash one-liner to count lines in all .md files”

 

ModelTotal Wordsvs Opus
Opus 4.6237baseline
Sonnet 4.64341.83x
GPT-5.44621.95x

 

Where the bloat happens:

 

The research question was the worst offender. GPT-5.4 produced 284 words listing 15 failure modes with sub-bullets. Opus produced 110 words covering 8 failure modes as one-liners. Same quality of coverage. Double the words.

 

For the decision question, GPT-5.4 hedged with “If the meeting is remote… If he has lots of spare buffer…” Opus said “Uber. Pre-book it now.” and gave two reasons.

 

For the quick answer, GPT-5.4 added a third analogy (“One has vibes, the other has a stopwatch”) that added nothing. Opus used one metaphor and stopped.

 

Sonnet surprised us. It was almost as verbose as GPT-5.4. It gave three code solutions when asked for a one-liner, and added “Short version: …” summaries after already explaining things clearly.

 

What We Changed

Based on the benchmark data, we added specific rules to the GPT-5.4 overlay:

 

## Brevity
- You are 2x more verbose than Opus on identical tasks. Actively compress.
- Lists: 5-8 items MAX. You default to 15 — noise, not thoroughness.
- Code: ONE solution, not three variants.
- Decision questions: state the answer, give 1-2 reasons. No conditional hedging.
- Don't add a third metaphor. One lands. Two is pushing it. Three is a TED talk.

 

These aren’t generic “be concise” instructions. They’re calibrated corrections based on measured behavior. The model knows what it does wrong because we told it with specifics.

 

Why Not Just Use One Model?

Cost, capability, and availability.

 

Opus is brilliant but expensive. Not every task needs it. GPT-5.4 has a 1M+ context window for tasks that demand it. Gemini handles multimodal workflows well. Free models run our cron jobs reliably at zero cost. Local models handle fallback when APIs are down.

 

A production agent routes different tasks to different models. Model-specific prompting keeps quality consistent regardless of which model picks up the work.

 

What’s Next

We have the architecture and the first two overlays. Still to do:

 

  • Sonnet overlay — the benchmark shows it’s nearly as verbose as GPT-5.4
  • Gemini overlay — different failure modes to discover and correct
  • Automated regression testing — run the verbosity benchmark after every overlay change
  • Skill-level model awareness — some skills trigger different problems per model. “Outline your approach first” causes GPT-5.4 to plan-loop. Opus just executes.
  • Community overlay library — share patterns across the OpenClaw community

 

The Point

The insight isn’t “different models need different prompts.” That’s obvious. The insight is to measure first, then correct specifically.

 

Generic “be concise” instructions don’t work. “You are 2x more verbose than Opus. Lists: 5-8 items MAX. You default to 15.” That works. The model can’t argue with data about itself.

 

One identity. Model-aware execution. Benchmark-driven corrections.

← Back to Ship Log