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
| Need |
Read |
| Full procedures for Modes 1-9 |
modes.md |
| Mode selection algorithm and failure pattern library |
selection-and-failure-patterns.md |
| Testing protocol and prompt engineering standards |
testing-and-standards.md |
| Integration points and cross-references |
integrations.md |
1---2name: prompt-reliability-engine3description: 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.4---56# Prompt Reliability Engine78Use 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.910## Invocation Triggers1112- Weak, failing, inconsistent, or under-specified prompt.13- New skill, command, workflow, or reusable instruction set.14- Prompt outputs drift in format, tone, length, naming, or structure.15- User asks whether a prompt is production-ready.16- User wants stricter wording, test cases, edge-case coverage, or a feedback loop.1718Do not use for simple one-shot questions or code-only tasks with no prompt/instruction design component.1920## Core Workflow21221. Identify the target: prompt, skill, command, workflow, or output contract.232. Select one or more modes from the routing table.243. Diagnose before repairing.254. Convert vague goals into observable constraints.265. Stress-test important artifacts with normal, minimal, ambiguous, adversarial, and off-scope inputs.276. Lock output shape when consistency matters.287. Report evidence, risks, and remaining failure modes.2930## Mode Routing3132| Mode | Use When | Required Output |33|---|---|---|34| 1 Failure Repair | Prompt is broken or underperforming | Diagnosis, root causes, fixed prompt, retest matrix |35| 2 Skill Builder | Build a reusable prompt/skill/workflow | Goal, risks, production artifact, safeguards, tests |36| 3 Stress-Test | Determine breakability | Adversarial input set, failure log, severity, score |37| 4 Constraint Engineer | Instructions are vague | Before/after constraints and consistency score |38| 5 Output Lock | Outputs drift | Fixed sections, order, limits, allowed/forbidden content |39| 6 Workflow Architect | One prompt should become a pipeline | Phases, handoff formats, checkpoints |40| 7 Production Audit | Decide readiness | 5-dimension scorecard and priority fixes |41| 8 Language Pruning | Wording is soft or bloated | Flagged phrases, tightened version, word-count delta |42| 9 Self-Improving Skill | Artifact will be reused | Failure log schema and improvement cycle |4344See [modes.md](references/modes.md) for full mode procedures.4546## Reliability Rules4748- Preserve original business goal, scope, audience, and required tone unless those are the problem.49- Fix root causes, not cosmetic symptoms.50- Make constraints measurable.51- Name output sections and order when format stability matters.52- Expose weaknesses before proposing fixes in stress-test mode.53- Keep uncertainty when truthfully required; remove hedging that weakens instructions.54- Apply existing anti-slop and public-safety rules from the active repo policy.5556## Mode Selection5758- Broken prompt: Failure Repair -> Constraint Engineer -> Output Lock -> Audit.59- Build from scratch: Skill Builder -> Stress-Test -> Output Lock -> Audit.60- "Is this good enough?": Stress-Test -> Audit.61- Inconsistent outputs: Output Lock -> Constraint Engineer -> Language Pruning.62- Reusable system: Workflow Architect -> Self-Improving Skill.63- Stricter wording: Language Pruning.6465If unspecified, infer the mode and state it briefly.6667## Output Expectations6869- Mode selected and why.70- Diagnosis before rewrite.71- Revised prompt/skill/workflow only when repair/build is requested.72- Test or stress matrix for important systems.73- Reliability score or readiness score when auditing.74- Remaining risks and next recommended fix.7576## Reference Map7778| Need | Read |79|---|---|80| Full procedures for Modes 1-9 | [modes.md](references/modes.md) |81| Mode selection algorithm and failure pattern library | [selection-and-failure-patterns.md](references/selection-and-failure-patterns.md) |82| Testing protocol and prompt engineering standards | [testing-and-standards.md](references/testing-and-standards.md) |83| Integration points and cross-references | [integrations.md](references/integrations.md) |
kmshihab7878