| name | prompt-reliability-engine |
| description | 9-mode prompt reliability framework: diagnose, repair, stress-test, constrain, lock, architect, audit, prune, and self-improve prompts, skills, and workflows. Treats prompting as engineering.
|
| risk | low |
| tags | ["process","quality","testing","meta","skills"] |
| created | "2026-03-25T00:00:00.000Z" |
| updated | "2026-03-25T00:00:00.000Z" |
Prompt Reliability Engine
Use this skill to make prompts, skills, and workflows production-grade through diagnosis, repair, stress testing, constraints, output locking, workflow design, audit, pruning, and self-improvement.
Invocation Triggers
- Weak, failing, inconsistent, or under-specified prompt.
- New skill, command, workflow, or reusable instruction set.
- Prompt outputs drift in format, tone, length, naming, or structure.
- User asks whether a prompt is production-ready.
- User wants stricter wording, test cases, edge-case coverage, or a feedback loop.
Do not use for simple one-shot questions or code-only tasks with no prompt/instruction design component.
Core Workflow
- Identify the target: prompt, skill, command, workflow, or output contract.
- Select one or more modes from the routing table.
- Diagnose before repairing.
- Convert vague goals into observable constraints.
- Stress-test important artifacts with normal, minimal, ambiguous, adversarial, and off-scope inputs.
- Lock output shape when consistency matters.
- Report evidence, risks, and remaining failure modes.
Mode Routing
| Mode | Use When | Required Output |
|---|
| 1 Failure Repair | Prompt is broken or underperforming | Diagnosis, root causes, fixed prompt, retest matrix |
| 2 Skill Builder | Build a reusable prompt/skill/workflow | Goal, risks, production artifact, safeguards, tests |
| 3 Stress-Test | Determine breakability | Adversarial input set, failure log, severity, score |
| 4 Constraint Engineer | Instructions are vague | Before/after constraints and consistency score |
| 5 Output Lock | Outputs drift | Fixed sections, order, limits, allowed/forbidden content |
| 6 Workflow Architect | One prompt should become a pipeline | Phases, handoff formats, checkpoints |
| 7 Production Audit | Decide readiness | 5-dimension scorecard and priority fixes |
| 8 Language Pruning | Wording is soft or bloated | Flagged phrases, tightened version, word-count delta |
| 9 Self-Improving Skill | Artifact will be reused | Failure log schema and improvement cycle |
See modes.md for full mode procedures.
Reliability Rules
- Preserve original business goal, scope, audience, and required tone unless those are the problem.
- Fix root causes, not cosmetic symptoms.
- Make constraints measurable.
- Name output sections and order when format stability matters.
- Expose weaknesses before proposing fixes in stress-test mode.
- Keep uncertainty when truthfully required; remove hedging that weakens instructions.
- Apply existing anti-slop and public-safety rules from the active repo policy.
Mode Selection
- Broken prompt: Failure Repair -> Constraint Engineer -> Output Lock -> Audit.
- Build from scratch: Skill Builder -> Stress-Test -> Output Lock -> Audit.
- "Is this good enough?": Stress-Test -> Audit.
- Inconsistent outputs: Output Lock -> Constraint Engineer -> Language Pruning.
- Reusable system: Workflow Architect -> Self-Improving Skill.
- Stricter wording: Language Pruning.
If unspecified, infer the mode and state it briefly.
Output Expectations
- Mode selected and why.
- Diagnosis before rewrite.
- Revised prompt/skill/workflow only when repair/build is requested.
- Test or stress matrix for important systems.
- Reliability score or readiness score when auditing.
- Remaining risks and next recommended fix.
Reference Map
