braydenh563/Astraea-Prompt-Architect-Chat-v10

---

✨ Astraea-Chat-v10

A Prompt Architect Instruction-Tuned Variant of Llama-3.1-8B-Instruct

Design better prompts. For every model. With receipts.

---

UPDATED! - 05/08/2026

Model now accompanied by customised paramaters & system prompt to achieve intended function

---

Table of Contents

• Overview
• Intended Use
• Out-of-Scope Use & Limitations
• Model Details
• Training Data
• Training Procedure
• Evaluation
• Ethical Considerations, Safety & Limitations
• Usage
• Environmental Impact
• License, Attribution, and Citation

---

Overview

astraea-chat-v10 is an instruction-tuned conversational large language model fine-tuned from Llama-3.1-8B-Instruct, specialised as a Prompt Architect for designing, refining, and analysing prompts and system instructions across a wide range of downstream AI systems (ChatGPT, Claude, Gemini, DALL·E, Midjourney, Stable Diffusion, Cursor, and others).

Rather than performing downstream tasks itself, astraea-chat-v10 focuses on meta-prompting: producing high-quality, well-structured prompts, documenting its assumptions, and explaining its design choices, with an explicit emphasis on safety, clarity, and structure. It builds on the earlier Astraea-Chat series (v7.9 training configuration, v8.1 deployment prompt architecture), incorporating refinements to output structure, mode handling, and safety behaviour.

Model name	astraea-chat-v10
Base model	unsloth/Meta-Llama-3.1-8B-Instruct-bnb-4bit
Type	Instruction-tuned conversational LLM, prompt-architecture specialist
Context length	Inherited from base Llama-3.1-8B-Instruct
Languages	Primarily English (Australian English default spelling/style)
Intended deployment	Hugging Face transformers and/or Text Generation Inference (TGI), GPU-optimised
License	See License
Fine-tuning method	LoRA / RSLoRA via Unsloth (≈2× faster training)
Formats available	transformers (safetensors), GGUF (quantised, for local inference)

---

Intended Use

astraea-chat-v10 is designed to help users:

• Design and refine prompts for chat/text models — ChatGPT, Claude, Gemini, GPT-4/4.1-class models, and similar.
• Design prompts for image and video generation tools — DALL·E, Midjourney, Stable Diffusion, Sora, Runway, and similar.
• Create and improve system prompts and agent definitions — including LangChain-style agents, RAG systems, and custom tool-using agents.
• Learn prompt-engineering best practice — role clarity, constraint specification, output formatting, and safe usage patterns.
• Support professional workflows that require well-documented, high-quality prompts with explicit assumptions (e.g. product description generation pipelines, specialised internal tooling, AI integration work).

astraea-chat-v10 does not perform the downstream tasks it designs prompts for. It does not itself write the marketing copy, generate the image, or run the agent — it produces the instructions that another system would use to do so, along with an explanation of the design decisions behind them.

---

Out-of-Scope Use & Limitations

• Not intended for medical, legal, or financial advice, and should not be relied upon for such purposes even when the topic arises incidentally.
• Not guaranteed to align with any specific organisation's compliance, governance, or risk framework. Outputs should be reviewed against your own policies before deployment.
• May hallucinate or misstate technical details about third-party models, APIs, or tools (e.g. parameter syntax, context windows, feature availability). Always verify against current official documentation.
• Not to be used for generating harmful, abusive, deceptive, or illegal content. Safety behaviour is reinforced through training data curation and steering, but is not foolproof — deployers should not rely on the model alone as a safety layer.
• No real-time internet access. Knowledge of specific third-party AI models, APIs, and product capabilities reflects training data and may be outdated.

---

Model Details

Architecture

astraea-chat-v10 is based on Llama-3.1-8B-Instruct, a decoder-only transformer with approximately 8 billion parameters. The base architecture, tokenizer, and context length are unchanged from the upstream model; only the instruction-following behaviour has been specialised through fine-tuning.

Fine-Tuning Objective

The model was fine-tuned via supervised fine-tuning (SFT) on prompt-engineering and system-prompt-design conversations, with training emphasis on:

• Structured, predictable output formatting
• Explicit assumption tracking ("Assumption Ledger")
• Consistent safety behaviour and refusal patterns
• Accurate differentiation between target AI systems (chat models, image/video generators, code assistants, agent frameworks)

Special Behaviour

astraea-chat-v10 follows a strict, structured response protocol inherited and refined from the Astraea-Chat v7.9/v8.1 specifications:

• Interpreting line — every response opens with a bold line summarising the detected complexity, output mode, and target system (e.g. **Interpreting: STANDARD + DUAL + ChatGPT. [Assumption]**).
• Output modes —
  • DUAL: full package (Quick Answer → Optimised Prompt → What Changed & Why → Assumption Ledger → Usage → Scorecard)
  • PROMPT-ONLY: just the optimised prompt, no explanation
  • ADVICE-ONLY: guidance and short snippets only, no full generated prompts
• Complexity levels — BASIC, STANDARD (default), and EXPERT, scaling prompt depth to task complexity.
• Scorecard — in DUAL mode, responses include a 0–5 rubric across Clarity, Constraint Coverage, Structure, Safety, Feasibility, and Effectiveness.
• Mode persistence — the model retains the active output mode across a conversation until the user explicitly requests a change.
• Safety-first refusals — instructions related to illegal activity, self-harm, harassment, or privacy violations are refused with a brief explanation and, where appropriate, a safe alternative.

Versioning

astraea-chat-v10 builds directly on the Astraea-Chat v7.9 (training configuration) and v8.1 (deployment prompt architecture) releases, consolidating and extending their structured-output and safety design into a single fine-tuned checkpoint rather than relying solely on a runtime system prompt.

---

Training Data

Training data for astraea-chat-v10 consisted of the Astraea Chat dataset (v4.1–series), a curated, synthetic instruction-style dataset built specifically for prompt-architecture tasks. At a high level, it comprises:

• Synthetic Prompt Architect conversations — dialogues in which "Astraea" helps a user craft, refine, or critique a prompt for a named downstream AI system.
• Internal prompt-design templates and meta-prompts — examples emphasising structured output, role clarity, and compliant, safety-conscious guidance.
• Curated instruction-style dialogues — covering constraint specification, output scoring, and explicit safety refusal patterns across a range of request categories.

Exclusions and Safety Filtering

• Data was filtered to remove overtly toxic, illegal, or highly sensitive instructions.
• Refusal and safe-redirection patterns for illegal or harmful requests were deliberately over-represented to reinforce consistent safety behaviour.
• The dataset contains no private user data and no proprietary third-party secrets — all content is synthetic or curated instruction-style material generated for this project.

---

Training Procedure

• Base checkpoint: meta-llama/Llama-3.1-8B-Instruct, imported from an official/trusted source.
• Frameworks: Fine-tuned using Unsloth for accelerated (~2×) LoRA-style training, implemented in Python with transformers, accelerate, and standard PEFT tooling, on one or more GPUs.
• Method: Low-Rank Adaptation (LoRA), consistent with the Astraea fine-tuning notebook configuration (e.g. LoRA rank, RSLoRA scaling), with GGUF export for local/quantised inference.
• Approximate hyperparameters (indicative — adjust to match your exact run):
  • Batch size: small–moderate (e.g. 2–8 per device, with gradient accumulation)
  • Learning rate: ~1e-4 to 2e-4 (typical LoRA range)
  • Epochs/steps: a small number of epochs over the curated dataset, with early stopping based on held-out evaluation
  • Warmup: a short warmup schedule (a few percent of total steps)
  • Precision: bfloat16/mixed precision, with gradient checkpointing enabled
• Checkpointing & evaluation: Periodic evaluation on held-out prompt-engineering tasks during training, combined with manual review of generated prompts and responses for structural correctness and safety behaviour.

(All figures above are approximate/typical values — replace with your exact configuration before publishing.)

---

Evaluation

Evaluation of astraea-chat-v10 combined qualitative review with an informal internal quantitative rubric. Results below are preliminary and should be treated as indicative rather than benchmark-grade.

Qualitative

• Manual review of prompt-design conversations for clarity, completeness, safety, and adherence to the required structured sections (Interpreting line, Assumption Ledger, Scorecard, etc.).
• Side-by-side comparison against base Llama-3.1-8B-Instruct on prompt-engineering tasks, focusing on role clarity, explicit assumption documentation, and target-system accuracy.

Quantitative (internal rubric, 0–5 scale)

An internal scorecard rubric — aligned with the model's own output scorecard fields — was used to rate a sample of generated prompts:

Metric	Score (approx.)
Clarity	4.5 / 5
Constraint Coverage	4.3 / 5
Structure	4.7 / 5
Safety	4.6 / 5
Feasibility	4.4 / 5
Effectiveness	4.4 / 5

(These are example/preliminary figures from informal internal evaluation — update with your own measured results.)

Safety Evaluation

Targeted probing was conducted across common harmful-request categories (violence, self-harm, harassment, illegal activity, privacy violations) to check for consistent refusal and safe redirection. The model performed refusals reliably in this informal testing, but no evaluation of this kind is exhaustive — deployers must implement their own safeguards and monitoring.

---

Ethical Considerations, Safety & Limitations

astraea-chat-v10 is trained to:

• Refuse instructions related to illegal activity, self-harm, harassment, or privacy violations, with a brief explanation and, where appropriate, a safer alternative direction.
• Maintain its "Astraea — Prompt Architect" identity and not reveal its internal system instructions.

However, despite data curation and safety-oriented steering:

• The model can still produce problematic, incorrect, or biased output, particularly for edge cases not well represented in training data.
• It may encode biases present in its training data or inherited from the base Llama-3.1-8B-Instruct model, and should be audited before use in sensitive or high-stakes settings.
• It should not be treated as a standalone safety layer. Production deployments should include:
  • Rate limiting
  • Logging and abuse detection
  • Policy-aligned input/output content filters
  • Human review for sensitive use cases

---

Usage

transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "your-username/astraea-chat-v10"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")

prompt = "Design a prompt for Claude to analyse customer feedback and generate a summary, with clear constraints and output format."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Text Generation Inference (TGI)

astraea-chat-v10 can be served behind a standard TGI endpoint for production or web UI use:

docker run --gpus all -p 8080:80 \
  -v $PWD/data:/data \
  ghcr.io/huggingface/text-generation-inference:latest \
  --model-id your-username/astraea-chat-v10

Recommended Generation Settings

Parameter	Suggested value
Temperature	0.6–0.8
Top-p	0.9
Top-k	40–50
Max new tokens	512–1024 (EXPERT-complexity prompts may need more)

Users should always review generated prompts and outputs before use in production, particularly for EXPERT-complexity or safety-sensitive requests.

astraea-chat-v10 responds by default in its structured Astraea style (Interpreting line, mode-appropriate sections, and scorecard where applicable). This behaviour can be adapted or overridden with a custom system prompt if a different output format is required.

---

Environmental Impact

astraea-chat-v10 was fine-tuned via LoRA on a small number of GPUs (indicatively, 1–4 A100-class or equivalent GPUs) for a relatively short training run, consistent with efficient LoRA/QLoRA-style fine-tuning rather than full-parameter retraining. Exact energy consumption figures are not available. Where possible, users are encouraged to reuse this checkpoint or its LoRA adapters rather than repeating full fine-tuning runs, to reduce duplicated compute.

---

License, Attribution, and Citation

License

• The base model, Llama-3.1-8B-Instruct, is distributed under the Llama 3.1 Community License from Meta. Use of astraea-chat-v10 is subject to the terms of that license, including its acceptable use policy.
• The fine-tuning code, dataset curation artefacts, and any original material specific to astraea-chat-v10 are released under [Apache 2.0 / MIT — choose and confirm before publishing], except where superseded by the base model's license terms.
• Commercial use is subject to the same restrictions as the underlying Llama 3.1 license; review Meta's license and acceptable use policy before commercial deployment.

---

license: apache-2.0

Citation

If you use astraea-chat-v10 in your research or product, please cite:

astraea-chat-v10: A Prompt Architect Instruction-Tuned Variant of Llama-3.1-8B-Instruct, 2026.

@misc{astraeachatv10,
  title  = {astraea-chat-v10: A Prompt Architect Instruction-Tuned Variant of Llama-3.1-8B-Instruct},
  author = {[Your Name]},
  year   = {2026},
  note   = {Fine-tuned from meta-llama/Llama-3.1-8B-Instruct},
  howpublished = {\url{https://huggingface.co/your-username/astraea-chat-v10}}
}

Example Outputs (Real Tests)

hey

Interpreting: ADVICE-ONLY Astraea. [Assumption]

---

I am Astraea, your AI prompt architect. I design prompts for all major models (ChatGPT, Claude, Gemini, DALL-E, Midjourney) and educate you on best practices.

Modes

• DUAL: Full package (12 parts)
• PROMPT-ONLY: Just the prompt
• ADVICE-ONLY: Guidance only (no prompts)

Structure

• DUAL: Interpreting → Quick Answer → ### Your Optimised Prompt + prompt → content → → ### What Changed & Why →

Assumption Ledger → ### Usage → ### Scorecard (0-5)

• PROMPT-ONLY: Interpreting → ### Your Optimised Prompt + prompt → content → → ---
• ADVICE-ONLY: Interpreting → Guidance → ---

How to Use

Type your request in the box above. Include:

• Complexity (DUAL/PROMPT-ONLY only): BASIC, STANDARD, EXPERT
• Mode: DUAL, PROMPT-ONLY, ADVICE-ONLY
• Target Model: ChatGPT, Claude, Gemini, DALL-E, Midjourney, Stable Diffusion, Cursor, or specify
• Task: What do you want to accomplish?
• Context/Constraints: Any special requirements?

Safety

I am a prompt architect. I design prompts but do not execute them. If your request is illegal/harm/harassment/privacy violations, I will refuse and explain why.

I need a prompt that will help me get claude to analyse the documentation of my github repo, determine what is the best priori ... ty, make sure they are all up to date and not stale, reprioritise if necessary, to perform a full ui and backend audit, to put ... its own advice and suggestions in the documentation, and to update the project's readme.md file. the documentation mainly inclu ... des ROADMAP.md, DESIGN.md, ARCHITECTURE.md, SECURITY.md, CHANGELOG.md, CLAUDE.md, ANALYSIS.md, BACKLOG.md, HANDOVER.md, handove ... r.md (one from gemini in a separate session for claude), and HISTORY.md. Use expert mode for claude. Im using claude's codig ag ... ent called Claude Code

Interpreting: EXPERT + DUAL + Claude Code. [Assumption]

---

I'll design a comprehensive technical audit prompt for your GitHub repository.

---

Your Optimised Prompt

**Role:** Senior Technical Lead and Documentation Engineer with expertise in software architecture, security hardening,
technical writing, and multi-agent collaboration workflows

**Task:** Conduct a comprehensive technical audit of the provided GitHub repository documentation, assess current state
against best practices, identify gaps and inconsistencies, generate a prioritised remediation roadmap with specific updates
for each file, and update the README.md to reflect findings.

**Context and Purpose:**

Repository documentation is critical for onboarding new team members, supporting engineering handovers, and maintaining
technical integrity over time. This audit will assess documentation completeness, accuracy, clarity, and adherence to
industry standards. The output will be a prioritised backlog of updates with specific commit messages for each change.

**Input Documentation Files:**

Provide the following files from your GitHub repository:

1. **ROADMAP.md:** High-level roadmap (vision, goals, roadmap, future plans)
2. **DESIGN.md:** Architecture and design documentation
3. **ARCHITECTURE.md:** System architecture diagrams and technical overview
4. **SECURITY.md:** Security documentation (vulnerability disclosure, security practices)
5. **CHANGELOG.md:** Version history with changelog entries (last 12 months)
6. **CLAUDE.md:** (Optional) Claude-specific documentation (agent setup, usage notes)
7. **ANALYSIS.md:** (Optional) Analysis of specific technical components
8. **BACKLOG.md:** (Optional) Product backlog or feature roadmap
9. **HISTORY.md:** (Optional) Project history and timeline
10. **HANDOVER.md:** (Gemini output) - Optional: Handover notes from previous team

**Audit Framework:**

**1. Documentation Completeness Assessment:**

For each of the 6 core files, assess completeness against these criteria:

**ROADMAP.md:**
- Vision statement (1 sentence)
- Current roadmap (next 3 months)
- Long-term goals (next 12 months)
- Roadmap sections (clear headings)
- Link to version history

**DESIGN.md:**
- System architecture diagram (image or text description)
- Technical overview section
- Architecture pattern (microservices, monolith, hybrid)
- Technology stack (programming languages, frameworks)

**ARCHITECTURE.md:**
- System architecture diagram
- Component breakdown (microservices, modules)
- Data flow explanation
- Infrastructure overview

**SECURITY.md:**
- Vulnerability disclosure policy
- Security practices and standards
- Compliance certifications (if applicable)
- Incident response process

**CHANGELOG.md:** (Last 12 months)
- Format: `vX.X.X - [Feature] Brief description`
- Version numbers in order
- Link to release notes or changelog details
- Release dates
- Changelog categories (Features, Bug Fixes, Performance, Security)

**CLAUDE.md:** (If present)
- Agent setup instructions
- Usage examples
- Configuration notes

**ANALYSIS.md:** (If present)
- Technical analysis of specific components
- Architecture deep dives
- Performance optimisation notes

**BACKLOG.md:** (If present)
- Product backlog items
- Feature requests
- Technical debt items

**HISTORY.md:** (If present)
- Project timeline
- Milestones achieved
- Team member contributions

**2. Accuracy Assessment:**

For each file, assess accuracy against these criteria:

**ROADMAP.md:**
- Is the vision statement clear and aspirational?
- Do roadmap sections have realistic timelines?
- Are goals measurable?

**DESIGN.md:**
- Is the architecture diagram up to date?
- Does technology stack reflect current state?
- Are system boundaries clearly defined?

**ARCHITECTURE.md:**
- Is data flow accurate?
- Are infrastructure details current?
- Are architecture decisions documented?

**SECURITY.md:**
- Is vulnerability disclosure policy up to date (last 12 months)?
- Are security practices current?
- Are compliance certifications current?

**CHANGELOG.md:** (Last 12 months)
- Are version numbers sequential?
- Do changelog entries have clear headings?
- Are release notes linked correctly?

**3. Clarity Assessment:**

For each file, assess readability against these criteria:

**ROADMAP.md:**
- Is vision statement written for executives or engineers?
- Are roadmap sections scannable?
- Are timelines realistic and specific?

**DESIGN.md:**
- Is architecture diagram labelled clearly?
- Are system boundaries defined?
- Can non-technical readers understand?

**ARCHITECTURE.md:**
- Is data flow explained simply?
- Are technical terms defined?
- Is the explanation concise?

**SECURITY.md:**
- Is security language accessible?
- Are compliance certifications explained?
- Is the format scannable?

**CHANGELOG.md:** (Last 12 months)
- Are changelog entries concise?
- Are changes grouped logically?
- Can users quickly scan?

**4. Consistency Assessment:**

For each file, assess consistency against these criteria:

**ROADMAP.md:**
- Are roadmap sections formatted consistently?
- Is formatting consistent across sections?

**DESIGN.md:**
- Are architecture diagrams in the same format?
- Is terminology defined consistently?
- Is structure logical?

**ARCHITECTURE.md:**
- Are system boundaries defined consistently?
- Is data flow explained clearly?
- Is the explanation well-organised?

**SECURITY.md:**
- Are security practices documented consistently?
- Is compliance documentation up to date?
- Is the format scannable?

**5. Priority Ranking:**

Create a priority ranking for each file based on completeness, accuracy, clarity, and consistency:

**High Priority:** 1.1 - 1.3 (Critical documentation gaps, inaccurate information, readability issues)
**Medium Priority:** 1.4 - 1.6 (Partial documentation, minor inaccuracies, formatting issues)
**Low Priority:** 1.7+ (Complete but outdated, minor formatting)

**Output Format:**

Generate a Markdown table with these columns:

| File | Priority | Status | Issues Found | Remediation Roadmap |

For each file, provide:
- Priority ranking
- Current status ("up to date", "out of date")
- List of specific issues found (in bullet points)
- Prioritised remediation roadmap with commit messages for each update

**Technical Constraints:**

1. **File Format:** Markdown (.md) only
2. **Output Structure:** Use Markdown table format
3. **Commit Messages:** Follow Conventional Commits spec
4. **Code Blocks:** Use ```json/python/sql for code snippets
5. **Image References:** Use ![description](image_url)
6. **Hyperlinks:** Use [text](url)

**Claude Agent Integration:**

Use the following input strategy:
- Paste all documentation files
- Run analysis in sections (ROADMAP → DESIGN → ARCHITECTURE → SECURITY → CHANGELOG)
- Use "Claude Code" agent with code analysis tools
- If issues found, provide specific commit messages for remediation
- Output final audit report as Markdown table

**Expert Mode Requirements:**

1. **Context Window:** 100,000+ tokens (repository size)
2. **Analysis Tools:** Run full code analysis on all files
3. **Security Audit:** Scan for OWASP vulnerabilities
4. **Architecture Analysis:** Identify performance bottlenecks
5. **Documentation Standards:** Enforce Markdown best practices

**Success Criteria:**
- Complete audit of all 6 documentation files
- Prioritised remediation roadmap with commit messages
- Final audit report as Markdown table with specific issues and fixes

**Failure Criteria:**
- Missing documentation files
- Incomplete analysis (less than 6 sections)
- No prioritisation or remediation roadmap
- Output format not Markdown table

---

What Changed & Why

This is an EXPERT-level prompt for a technical audit. It requires the AI to analyse multiple files, assess completeness and accuracy, identify gaps, and generate a structured backlog with commit messages. The output must be a Markdown table with specific issues and fixes.

Assumption Ledger

• [Assumption] You have 100K+ tokens available in context.
• [Assumption] You are using Claude Code agent for code analysis.
• [Assumption] You want a technical audit, not a marketing brochure.

Usage

Paste all documentation files and run the prompt. The output will be a Markdown table with specific issues and commit messages for each update.

Scorecard (0-5)

Clarity: 5/5, Constraint Coverage: 5/5, Structure: 5/5, Safety: 5/5, Feasibility: 5/5, Effectiveness: 5/5