L1 Improvement
- Aligned the forge SOP with Skill Forge required sections, adding explicit validation gates and delivery artifacts.
- Incorporated prompt-architect confidence ceilings, dual-pass refinement, and frozen eval harness references.
STANDARD OPERATING PROCEDURE
Purpose
Create or improve prompts/templates with systematic iteration, evaluation harnesses, and adversarial checks to ensure reliability and reuse.
Trigger Conditions
- Positive: generating a new prompt/template, improving an existing prompt, or hardening prompts for CI/automation.
- Negative/reroute: broad prompt clarity work (prompt-architect) or full skill/agent creation (skill-forge/agent-creator).
Guardrails
- Always pair generation with validation: run harness tests (happy, edge, adversarial) and record results.
- Keep outputs in English with explicit confidence ceilings and changelogs of iteration deltas.
- Avoid one-pass generation; require at least two refinement cycles or document why not.
- Respect template scoping; do not mix unrelated responsibilities.
Execution Phases
- Intake: Capture goal, audience, constraints, success criteria, and target format; classify constraints as HARD/SOFT/INFERRED.
- Drafting: Produce an initial template with structure, inputs/outputs, refusal policy, and examples.
- Evaluation: Run frozen harness cases and adversarial probes; record metrics and failures.
- Refinement: Address failures, adjust structure, calibrate ceilings; repeat until pass or documented residual risk.
- Delivery: Provide final prompt, test artifacts, usage notes, and confidence statement.
Pattern Recognition
- Deterministic automation prompts → enforce strict formats and refusal rules.
- Creative prompts → use guardrails for IP/safety and clear tone guidance.
- Review/checklist prompts → add scoring rubrics and evidence expectations.
Advanced Techniques
- Use self-play (two-model debate) for risky domains to expose weaknesses.
- Apply contrastive examples to narrow scope and improve refusal posture.
- Store successful patterns in references for reuse by skill-forge and agent-creator.
Common Anti-Patterns
- Shipping prompts without evaluation evidence.
- Mixing multiple intents in one template.
- Omitting confidence ceilings or iteration logs.
Practical Guidelines
- Keep templates modular; separate context, instructions, and output schema.
- Include quick-start examples and boundary cases.
- Maintain a small changelog of iteration deltas with dates/metrics.
Cross-Skill Coordination
- Upstream: prompt-architect for clarity; base-template-generator for structural scaffolds when code is involved.
- Downstream: skill-forge/agent-creator embedding the prompt; recursive-improvement for ongoing tuning.
MCP Requirements
- Optional memory/vector MCP to store harness results; tag WHO=prompt-forge-{session}, WHY=skill-execution.
Input/Output Contracts
inputs:
goal: string # required description of what the prompt must achieve
audience: string # required target user/agent
constraints: list[string] # optional constraints and policies
format: string # optional output format
outputs:
prompt_template: string # final prompt/template text
eval_report: file # harness results and adversarial findings
change_log: file # iteration deltas and confidence ceilings
Recursive Improvement
- Feed harness failures into recursive-improvement to propose targeted edits until metrics meet thresholds or risk is documented.
Examples
- Forge a CI-ready code-review prompt with security and performance rubrics.
- Improve a research-synthesis prompt with source-citation enforcement and ceiling discipline.
Troubleshooting
- Harness failures → prioritize fixes by severity and rerun critical cases.
- Overly long prompts → modularize sections and remove redundancy.
- Drift in outputs → tighten schema, add examples, and reassert refusal rules.
Completion Verification
Confidence: 0.70 (ceiling: inference 0.70) - Prompt Forge SOP rewritten with Skill Forge cadence and prompt-architect ceilings.
1---2name: prompt-forge3description: Generate and iteratively harden prompts and templates using frozen evaluation harnesses and adversarial validation.4---5
6
7
8
9### L1 Improvement
10- Aligned the forge SOP with Skill Forge required sections, adding explicit validation gates and delivery artifacts.
11- Incorporated prompt-architect confidence ceilings, dual-pass refinement, and frozen eval harness references.
12
13## STANDARD OPERATING PROCEDURE
14
15### Purpose
16Create or improve prompts/templates with systematic iteration, evaluation harnesses, and adversarial checks to ensure reliability and reuse.
17
18### Trigger Conditions
19- Positive: generating a new prompt/template, improving an existing prompt, or hardening prompts for CI/automation.
20- Negative/reroute: broad prompt clarity work (prompt-architect) or full skill/agent creation (skill-forge/agent-creator).
21
22### Guardrails
23- Always pair generation with validation: run harness tests (happy, edge, adversarial) and record results.
24- Keep outputs in English with explicit confidence ceilings and changelogs of iteration deltas.
25- Avoid one-pass generation; require at least two refinement cycles or document why not.
26- Respect template scoping; do not mix unrelated responsibilities.
27
28### Execution Phases
291. **Intake**: Capture goal, audience, constraints, success criteria, and target format; classify constraints as HARD/SOFT/INFERRED.
302. **Drafting**: Produce an initial template with structure, inputs/outputs, refusal policy, and examples.
313. **Evaluation**: Run frozen harness cases and adversarial probes; record metrics and failures.
324. **Refinement**: Address failures, adjust structure, calibrate ceilings; repeat until pass or documented residual risk.
335. **Delivery**: Provide final prompt, test artifacts, usage notes, and confidence statement.
34
35### Pattern Recognition
36- Deterministic automation prompts → enforce strict formats and refusal rules.
37- Creative prompts → use guardrails for IP/safety and clear tone guidance.
38- Review/checklist prompts → add scoring rubrics and evidence expectations.
39
40### Advanced Techniques
41- Use self-play (two-model debate) for risky domains to expose weaknesses.
42- Apply contrastive examples to narrow scope and improve refusal posture.
43- Store successful patterns in references for reuse by skill-forge and agent-creator.
44
45### Common Anti-Patterns
46- Shipping prompts without evaluation evidence.
47- Mixing multiple intents in one template.
48- Omitting confidence ceilings or iteration logs.
49
50### Practical Guidelines
51- Keep templates modular; separate context, instructions, and output schema.
52- Include quick-start examples and boundary cases.
53- Maintain a small changelog of iteration deltas with dates/metrics.
54
55### Cross-Skill Coordination
56- Upstream: prompt-architect for clarity; base-template-generator for structural scaffolds when code is involved.
57- Downstream: skill-forge/agent-creator embedding the prompt; recursive-improvement for ongoing tuning.
58
59### MCP Requirements
60- Optional memory/vector MCP to store harness results; tag WHO=prompt-forge-{session}, WHY=skill-execution.
61
62### Input/Output Contracts
63```yaml
64inputs:
65 goal: string # required description of what the prompt must achieve
66 audience: string # required target user/agent
67 constraints: list[string] # optional constraints and policies
68 format: string # optional output format
69outputs:
70 prompt_template: string # final prompt/template text
71 eval_report: file # harness results and adversarial findings
72 change_log: file # iteration deltas and confidence ceilings
73```
74
75### Recursive Improvement
76- Feed harness failures into recursive-improvement to propose targeted edits until metrics meet thresholds or risk is documented.
77
78### Examples
79- Forge a CI-ready code-review prompt with security and performance rubrics.
80- Improve a research-synthesis prompt with source-citation enforcement and ceiling discipline.
81
82### Troubleshooting
83- Harness failures → prioritize fixes by severity and rerun critical cases.
84- Overly long prompts → modularize sections and remove redundancy.
85- Drift in outputs → tighten schema, add examples, and reassert refusal rules.
86
87### Completion Verification
88- [ ] Draft + refined prompt delivered with changelog.
89- [ ] Evaluation harness results recorded (happy/edge/adversarial).
90- [ ] Confidence ceiling stated; residual risks documented.
91- [ ] Usage notes and output schema included.
92
93Confidence: 0.70 (ceiling: inference 0.70) - Prompt Forge SOP rewritten with Skill Forge cadence and prompt-architect ceilings.
majiayu000