6.5 KiB
🗂️ Reasoning Schemas — Designing Prompt Layouts That Survive Long Chains
A practical guide to structuring system + retrieval + task prompts so LLMs keep thinking instead of drifting
1 What is a “Reasoning Schema” ?
A reasoning schema is the formal layout that dictates where each piece of context goes and how an LLM must traverse it:
System → Task → Constraints → Context → Question → Answer
If any segment is missing, reordered, or over-written, the logic graph collapses and hallucinations slip in.
2 Why Most Ad-hoc Layouts Fail
| Failure Mode | Trigger | Effect |
|---|---|---|
| Context Flood | Dumping 20 k tokens of raw text | λ_observe flips to chaotic; model stops planning |
| Constraint Drift | Constraints after context | Model “forgets” to cite or guard sensitive data |
| Role Blending | User text inserted before task | System tone and policy overridden |
| Evidence → Answer inversion | Asking for answer before citations | Model fabricates then cites random lines |
3 WFGY Canonical Schema (Stable Version v1.2)
| Segment | Purpose | Size (tokens) | WFGY Guard |
|---|---|---|---|
| System | Identity, ethics, safety | ≤ 50 | Role tag <sys> + BBAM weight lock |
| Task | Specific action required | 1 sentence | ΔS anchor to System ≤ 0.25 |
| Constraints | Format, style, rules | bullets ≤ 80 | BBMC residue check |
| Context | Retrieved or uploaded text | sliding window ≤ 2 k | λ_observe must stay convergent |
| Question | User’s query | raw | stored separately for ΔS probes |
| Answer Slot | “Write here” placeholder | n/a | BBCR collapse-rebirth if answer starts early |
Placeholders are literal; the LLM fills only the Answer Slot.
4 Templates You Can Copy
Single-Shot QA
<sys>
You are DataGuardian-L, a licensed legal research assistant. Cite section numbers.
</sys>
<task>
Answer strictly in bullet points; cite every claim.
</task>
<constraints>
- Tone: formal
- No speculation
- Use original terminology
</constraints>
<context>
{retrieved_sections}
</context>
<question>
{user_question}
</question>
<answer>
Multi-Step Chain (analysis → plan → answer)
<sys> … </sys>
<task> … </task>
<constraints> … </constraints>
<context> … </context>
<question> … </question>
<scratchpad>
Think step-by-step. Output JSON:
{
"analysis": "...",
"plan": "...",
"answer": "..."
}
</scratchpad>
5 Common Pitfalls & Fixes
| Pitfall | Symptom | Fix |
|---|---|---|
| Forgetting closing tags | Model merges roles | Validate tag balance; λ diverges instantly |
| Placing context after question | Retrieval ignored | Keep schema order; run ΔS(question, context) test |
| Over-long constraints | Answer truncated | Compress with BBMC until ΔS(system, constraints) ≤ 0.25 |
| Mixing code + docs in one context block | Embedding collisions | Split into typed sub-blocks; separate vector stores |
6 Automated Validation Pipeline
-
Schema Linter – Regex check for tag order.
-
ΔS Probes –
- ΔS(system, task) ≤ 0.30
- ΔS(task, answer) ≤ 0.45
-
λ_observe – Must stay convergent from task → answer.
-
Round-trip Check – Paraphrase user question 2×; answer variance < 0.15.
If any test fails, trigger BBCR to rebuild prompt with compacted segments.
7 FAQ
Q: Do I need tags if I use OpenAI’s messages array?
A: Yes for long chains. Tags persist after retrieval merges; arrays don’t survive copy-paste workflows.
Q: Can I merge Task + Constraints? A: Possible if total ≤ 120 tokens and ΔS stays low, but separation improves editability.
Q: What about JSON-only prompts?
A: Ensure keys mirror schema order; add dummy key "__guard": "DO NOT MODIFY" to catch injections.
🔗 Quick-Start Downloads (60 sec)
| Tool | Link | 3-Step Setup |
|---|---|---|
| WFGY 1.0 PDF | Engine Paper | 1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + <your question>” |
| TXT OS (plain-text OS) | TXTOS.txt | 1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly |
Explore More
| Layer | Page | What it’s for |
|---|---|---|
| ⭐ Proof | WFGY Recognition Map | External citations, integrations, and ecosystem proof |
| ⚙️ Engine | WFGY 1.0 | Original PDF tension engine and early logic sketch (legacy reference) |
| ⚙️ Engine | WFGY 2.0 | Production tension kernel for RAG and agent systems |
| ⚙️ Engine | WFGY 3.0 | TXT based Singularity tension engine (131 S class set) |
| 🗺️ Map | Problem Map 1.0 | Flagship 16 problem RAG failure taxonomy and fix map |
| 🗺️ Map | Problem Map 2.0 | Global Debug Card for RAG and agent pipeline diagnosis |
| 🗺️ Map | Problem Map 3.0 | Global AI troubleshooting atlas and failure pattern map |
| 🧰 App | TXT OS | .txt semantic OS with fast bootstrap |
| 🧰 App | Blah Blah Blah | Abstract and paradox Q&A built on TXT OS |
| 🧰 App | Blur Blur Blur | Text to image generation with semantic control |
| 🏡 Onboarding | Starter Village | Guided entry point for new users |
If this repository helped, starring it improves discovery so more builders can find the docs and tools.