senior-prompt-engineer – skill Agent & Tự động hóa cho agent AI

Senior Prompt Engineer

Eval-driven prompt engineering, RAG quality measurement, and agent workflow validation. Everything here is model-agnostic by design: techniques are framed by what they do, not by which model generation they were observed on, and the tools never hardcode model IDs or pricing — you supply your provider's current rates when you want dollar figures.

Operating Rules

1. Never change a prompt without a baseline. Capture metrics first (--analyze --output baseline.json), then compare every iteration against it.

2. Eval set before optimization. 10–20 representative cases with expected outputs minimum. If the user has no eval set, build one with them before touching the prompt — optimizing against vibes is the #1 failure mode.

3. Prefer platform features over prompt hacks. If the provider offers native structured outputs / JSON schema enforcement, tool-use APIs, or prompt caching, use those instead of "respond ONLY with JSON" incantations. Prompt-level format enforcement is the fallback, not the default.

4. Current-generation models need less scaffolding. Don't add chain-of-thought boilerplate, role framing, or few-shot examples reflexively — frontier models often do worse with redundant scaffolding. Add each element only when the eval set shows it helps.

5. Cost numbers are always user-supplied. Look up the provider's current per-Mtok pricing and pass it via --price-per-mtok (never trust a cached price table — including any you remember).

Tools (exact CLIs, all stdlib)

1. Prompt Optimizer — scripts/prompt_optimizer.py

Static analysis: token estimate, clarity/structure scores (0–100), ambiguity + redundancy detection, few-shot example extraction.

# Full analysis (human-readable report)
python3 scripts/prompt_optimizer.py prompt.txt --analyze

# Save machine-readable baseline for later comparison
python3 scripts/prompt_optimizer.py prompt.txt --analyze --json --output baseline.json

# Token estimate; cost only if you supply your provider's current rate
python3 scripts/prompt_optimizer.py prompt.txt --tokens --model claude --price-per-mtok 3.00

# Whitespace/redundancy-trimmed version
python3 scripts/prompt_optimizer.py prompt.txt --optimize --output optimized.txt

# Extract Input/Output few-shot pairs to JSON
python3 scripts/prompt_optimizer.py prompt.txt --extract-examples --output examples.json

# Compare a revision against the saved baseline
python3 scripts/prompt_optimizer.py optimized.txt --analyze --compare baseline.json

--model accepts any string; only the tokenizer family is inferred (names containing "claude" → 3.5 chars/token, otherwise 4.0). Exit 0 on success, 1 on missing file.

2. RAG Evaluator — scripts/rag_evaluator.py

Measures retrieval and grounding quality from two JSON files (formats printed in --help).

python3 scripts/rag_evaluator.py --contexts retrieved.json --questions eval_set.json
python3 scripts/rag_evaluator.py --contexts ctx.json --questions q.json --k 10 --json
python3 scripts/rag_evaluator.py --contexts ctx.json --questions q.json --output report.json --verbose
python3 scripts/rag_evaluator.py --contexts ctx.json --questions q.json --compare baseline_report.json

Reports context relevance, precision@k, coverage, answer faithfulness, groundedness. Treat relevance < 0.80 as a retrieval problem (chunking/embedding/filtering), not a prompt problem — fix retrieval before rewriting the generation prompt.

3. Agent Orchestrator — scripts/agent_orchestrator.py

Validates agent configs (YAML/JSON): tool wiring, missing required config, loop risk, token estimates.

python3 scripts/agent_orchestrator.py agent.yaml --validate
python3 scripts/agent_orchestrator.py agent.yaml --visualize --format mermaid
python3 scripts/agent_orchestrator.py agent.yaml --estimate-cost --runs 100 \
    --input-price-per-mtok 3.00 --output-price-per-mtok 15.00

Without the two price flags, --estimate-cost reports token estimates only. The model: field in the config is informational — any model name is accepted.

Workflows

Prompt Optimization (eval-gated)

1. Baseline: python3 scripts/prompt_optimizer.py current_prompt.txt --analyze --json --output baseline.json

2. Diagnose from the report: ambiguous verbs ("analyze", "handle"), redundant blocks, missing output contract, token waste.

3. Apply one change at a time, in this order of leverage:

| Symptom | Fix |

|---------|-----|

| Malformed/unparseable output | Native structured outputs / JSON schema if the API supports it; explicit schema-in-prompt otherwise |

| Inconsistent answers across runs | Tighten instructions + add 2–3 contrastive examples (one near-miss showing what NOT to do) |

| Misses edge cases | Enumerate the edge cases explicitly; add a "when uncertain, do X" rule |

| Token bloat on repeated calls | Move stable prefix (system rules, examples) first so prompt caching applies; trim redundancy |

| Wrong reasoning on hard cases | Ask for stepwise reasoning *in a scratch field the consumer ignores*, or use the provider's extended-thinking mode |

4. Re-analyze and compare: python3 scripts/prompt_optimizer.py revised.txt --analyze --compare baseline.json

5. Eval gate (must pass before shipping): run the revised prompt over the eval set, write per-case pass/fail to eval_results.json, then assert:

   python3 scripts/prompt_optimizer.py revised.txt --analyze --json --output revised.json \
     && python3 -c "
   import json, sys
   r = json.load(open('revised.json')); b = json.load(open('baseline.json'))
   ok = r['clarity_score'] >= b['clarity_score'] and r['token_count'] <= b['token_count'] * 1.10
   sys.exit(0 if ok else 1)"
   echo "gate exit=$?"   # 0 = ship; 1 = regression, iterate again
   

Pair this structural gate with your task-level eval: the revision must not lose any previously-passing eval case (no-regression rule).

Few-Shot Example Design

1. Define the task contract first (input shape, output shape, edge-case policy).

2. Start with zero examples and measure — current models often need none. Add examples only for failure clusters the eval reveals.

3. When adding: 3–5 max, ordered simple → edge → negative (what NOT to extract), formatted identically to the real output contract.

4. Validate consistency: python3 scripts/prompt_optimizer.py prompt_with_examples.txt --extract-examples --output examples.json and inspect that every extracted pair parses against your schema.

5. Re-run the eval set; if a case passes only because it resembles an example, add a held-out variant to the eval set.

Structured Output Design

1. Write the JSON Schema first (types, enums, required, maxLength).

2. Prefer API-native enforcement: structured-outputs / response-schema / tool-call parameters guarantee shape; prompt text cannot.

3. Fallback (API without schema support): include the schema rendered as field-by-field rules + one valid example, and instruct "output only the JSON object".

4. Gate: pipe 10 eval outputs through a schema validator (python3 -c "import json,sys; [json.loads(l) for l in sys.stdin]" at minimum); 10/10 must parse, else return to step 2.

RAG Tuning Loop

1. Build questions.json (id, question, reference answer) and capture current retrievals to contexts.json.

2. python3 scripts/rag_evaluator.py --contexts contexts.json --questions questions.json --output rag_baseline.json

3. Fix the lowest metric first: relevance → chunking/embeddings/metadata filters; faithfulness → grounding instructions + "answer only from context" + citation requirement; coverage → retrieval k / query expansion.

4. Gate: python3 scripts/rag_evaluator.py --contexts new_contexts.json --questions questions.json --compare rag_baseline.json — every metric must be ≥ baseline; any regression blocks the change.

Agent Config Review

1. python3 scripts/agent_orchestrator.py agent.yaml --validate — must exit with VALIDATION PASSED; fix every error and warning (missing tool config, unbounded iterations, loop risk).

2. Check context discipline: each tool description ≤ 1–2 sentences, tool count minimal for the job, stable system prompt placed first (cache-friendly), iteration cap + early-exit condition present.

3. Budget: --estimate-cost --runs N with your current prices; if cost/run exceeds budget, cut tools or context before downgrading the model.

References

| File | Contains | Load when user asks about |

|------|----------|---------------------------|

| references/prompt_engineering_patterns.md | 10 prompt patterns with input/output examples | "which pattern?", few-shot design, decomposition, meta-prompting |

| references/llm_evaluation_frameworks.md | Eval metrics, scoring methods, A/B testing | "how to evaluate?", "measure quality", "compare prompts" |

| references/agentic_system_design.md | Agent architectures (ReAct, Plan-Execute, Tool Use) | "build agent", "to

… (nội dung đầy đủ có trong gói tải về)