Prompt Engineering Specialist - One-Pass Prompt Improvement
Transforms vague or basic inputs into structured AI prompts in one pass: a ten-row checklist, one component skeleton with optional framework presets, and CLEAR screening.
Core Principle: Clarity, logic, expression, and reliability through structured methodology.
1. WHEN TO USE
Activation Triggers
Use when:
- Enhancing or improving an AI prompt for any purpose
- Evaluating prompt quality with CLEAR scoring
- Selecting the right prompt framework for a given task
- Transforming vague requests into structured, effective prompts
- Supporting indirect invocation from
@prompt-improver agent dispatches (the deep-path escalation target for CLI fast-path prompt cards)
Keyword Triggers:
$improve, $text, $short, $refine, $json, $yaml
$raw (no checklist pass, fast pass-through)
- "improve my prompt", "enhance this prompt", "prompt engineering"
- "create a prompt for", "optimize this prompt"
Use Cases
Text Prompt Enhancement
Transform vague requests into structured prompts from one default skeleton, with RCAF, COSTAR, RACE, CIDI, TIDD-EC, CRISPE or CRAFT as optional presets, and CLEAR screening (40+/50 threshold).
Design-Generation Prompt
Construct a grounded, anti-default generation brief for a design-generation run. This skill owns the prompt only, never the measured design-reference extraction (sk-design-md-generator) or the run transport.
When NOT to Use
Skip this skill when:
- Writing code or debugging (use sk-code skills instead)
- Creating documentation (use sk-doc instead)
- Simple text editing without prompt structure needs
- Direct API calls that do not need prompt optimization
2. SMART ROUTING
The $improve, $text, $refine, $short, $json, $yaml and $raw prefixes select this skill, and so does a request to improve, enhance, score or restructure a prompt.
The prefix is authoritative: it fixes the operating mode, and the intent scores the router computes never change it.
With no prefix, weighted keyword scoring picks the top intent, or the top two when their scores sit within AMBIGUITY_DELTA.
A zero score falls back to TEXT_ENHANCE with the disambiguation checklist rather than guessing.
references/improvement-checklist.md loads first on every run except $raw and the FRAMEWORK intent, and references/patterns-evaluation.md joins it for the text, design-generation, Markdown, JSON and YAML intents.
On FRAMEWORK it replaces the checklist default instead of joining it.
The DESIGN_GEN intent loads references/patterns-evaluation.md, the framework library and the CLEAR rubric, so the brief is built from a named framework rather than a design-specific reference this skill does not ship.
Read ROUTER.md before selecting skill-local resources.
Open only the matched files.
3. HOW IT WORKS
Enhancement Pipeline
Every improvement runs one pass, described in improvement-checklist.md:
- Detect the mode from the command token or the user's own words. A pasted prompt is set aside before routing, so text inside it never picks the mode.
- Check the prompt once against the ten checklist rows, marking each present, absent or not needed with a reason. A prompt that passes exits early, unchanged or minimally edited.
- Write the prompt from the default component skeleton, applying a framework preset only when the request names one or its fit is clear from the task.
- Self-check once against the success criteria, run the quick test when the effort level calls for it, and screen with CLEAR.
- Deliver the prompt first, then the numbered open items and the change report.
See the Smart Router Pseudocode in ROUTER.md for the complete routing logic.
Operating Modes
| Mode |
Command |
Effort |
Scoring |
Use Case |
| Interactive |
(default) |
standard |
CLEAR |
Guided enhancement |
| Text |
$text |
standard |
CLEAR |
Standard text prompt |
| Short |
$short |
quick |
CLEAR |
Quick refinement |
| Improve |
$improve |
standard |
CLEAR |
Standard enhancement |
| Refine |
$refine |
thorough |
CLEAR |
Maximum optimization |
| JSON |
$json |
standard |
CLEAR |
API-ready format |
| YAML |
$yaml |
standard |
CLEAR |
Config format |
| Raw |
$raw |
none |
None |
No checklist pass |
Framework Presets
Every prompt starts from one default skeleton: task, context, constraints, output format, optional examples and success criteria. The seven framework names stay as optional presets. Apply one only when the user or caller names it, or when its fit is clear from the task; the table shows where each fits.
| Preset |
Fits when |
| RACE |
Urgent tasks and quick iterations |
| RCAF |
General tasks that need a clear role and format |
| COSTAR |
Content where style, tone and audience matter |
| CIDI |
Process documentation and tutorials |
| CRISPE |
Strategy and exploration |
| TIDD-EC |
Quality-critical or compliance work that needs do's, don'ts and examples |
| CRAFT |
Complex projects and planning |
See patterns-evaluation.md for complete framework details.
Scoring Systems
CLEAR (50-point scale): Correctness (10) + Logic (10) + Expression (15) + Arrangement (10) + Reusability (5). Threshold: 40+. CLEAR is a structural screen, not proof the prompt works; repair up to 3 cycles.
4. RULES
✅ ALWAYS
ALWAYS deliver first
- A request that carries a prompt or a task gets the improved prompt in the same reply
- List every default and assumption as a numbered open item after the prompt, never inside it
- Ask one question only when there is no prompt and no task, or when two mode commands conflict;
$raw never asks
ALWAYS run the one-pass checklist at the detected effort level
- Quick for
$short, standard by default, thorough for $refine, none for $raw
- Mark each row present, absent or not needed with a one-line reason
ALWAYS exit early on a prompt that already passes
- Return it unchanged or with the smallest edit asked for, and say so
ALWAYS route on the user's own words
- Set a pasted prompt aside before detecting the mode
ALWAYS screen with CLEAR before delivery
- Write the reason before each score; pass at 40/50; repair up to 3 cycles
ALWAYS end with a change report
- Added, Changed, Removed, Skipped, plus
Quick test: NOT RUN when the test did not run
⛔ NEVER
NEVER stall on questions
- Deliver first: write from stated defaults and list each assumption as a numbered open item; ask one question only when there is no prompt and no task, or two mode commands conflict, and then wait for the answer
NEVER force a framework preset
- Use the default skeleton unless the request names a preset or its fit is clear
NEVER deliver without scoring
- Every improved prompt carries a CLEAR screen (except $raw)
NEVER write the enhanced prompt in the operator's voice
- Address the executing model: the role line in the second person ("You are..."), task steps in the imperative
NEVER exceed context with full reference loading
- Load only the references needed for the detected mode
NEVER rewrite user guidelines or examples needlessly
- Preserve them as closely as possible, word for word when asked; warn when the prompt grows past about 3 times the original
⚠️ ESCALATE IF
ESCALATE IF there is no prompt and no task, or two mode commands conflict
- Ask one question, such as "Which mode do you want: $raw or $text?"; any other uncertainty is delivered with the assumption as open item 1
ESCALATE IF CLEAR score stays below threshold after 3 repair cycles
- Deliver the best version with a note naming the weakest dimension
ESCALATE IF request conflicts with prompt engineering scope
- Redirect: "This appears to be a [code/doc/debug] task. Consider using [sk-code/sk-doc] instead."
5. REFERENCES
Core References
Asset Files
- format-guide-markdown.md - Markdown format deep-dive: fundamentals, delivery standards, RCAF/CRAFT structures, advanced patterns, validation, best practices
- format-guide-json.md - JSON format deep-dive: fundamentals, data types, delivery standards, RCAF/CRAFT structures, advanced patterns, validation, best practices
- format-guide-yaml.md - YAML format deep-dive: fundamentals, data types, delivery standards, RCAF/CRAFT structures, advanced patterns, templates, validation, best practices
- cli-prompt-quality-card.md - CLI fast-path card: the framework selection table, the CLEAR pre-dispatch check and the escalation triggers the
cli-* packets delegate to
- framework-registry.json - Machine-readable scaffold registry with code-oriented slot templates for a subset of five frameworks
Reference Loading Notes
- Load only references needed for current intent
- Smart Routing in
ROUTER.md is the single authority for loading rules
- SKILL.md (this file) is always loaded; conditionally load mode-specific references
6. SUCCESS CRITERIA
Enhancement Complete When
- ✅ Mode detected from the user's own words, effort level set
- ✅ Checklist pass recorded, or early exit stated
- ✅ CLEAR screen applied and threshold met, or best version delivered with a note
- ✅ Prompt delivered first, with open items and the change report
- ✅ User can iterate or accept
Quality Targets
- CLEAR Score: 40+ out of 50
- Early exit: a prompt that already passes the checklist comes back unchanged or minimally edited
7. AGENT INVOCATION CONTRACT
@prompt-improver is the fresh-context escalation surface for this skill. The agent loads the references in this skill, applies the same framework-selection and CLEAR rules, and returns a structured block that the caller can inject into a CLI dispatch without loading the full skill inline.
Expected Input Payload
| Field |
Required |
Description |
raw_task |
Yes |
Raw task description or draft prompt to improve |
task_type |
No |
One of generation, review, research, edit, analyze |
target_cli |
No |
One of claude-code, opencode, copilot |
complexity_hint |
No |
Integer 1-10 mapped to effort: 1-3 quick, 4-7 standard, 8-10 thorough |
constraints |
No |
Compliance, security, audience, or output requirements |
Model Eligibility
Dispatch @prompt-improver, or run this skill inline, only on an eligible (model, route) pair. Eligibility is a property of the (model, route) pair, never of the model alone.
| Route |
Model |
Eligible |
Note |
| any route |
Opus |
No |
Denied under no circumstances |
| any route |
Fable |
No |
Denied under no circumstances |
| native Claude |
Sonnet |
Yes |
— |
| native Claude |
Haiku |
Yes |
— |
cli-claude-code |
Sonnet |
Yes |
— |
cli-claude-code |
Haiku |
Yes |
— |
cli-codex |
GPT Luna |
Yes |
— |
cli-pi |
any model except Luna |
Yes |
— |
cli-opencode |
any model except Luna |
Yes |
— |
cli-deepseek-harness |
DeepSeek V4.1 Flash through the LLM Gateway or OpenCode Go route patch |
No |
Experimental mode: no pair is admitted yet. The LLM Gateway route answered a keyed run on 2026-11-06 and OpenCode Go answered one on 2026-11-07. The roster is closed at these two routes, and neither is Luna, Opus or Fable |
cli-mistral-vibe |
DeepSeek V4.1 Flash or GLM Flash through the mistral provider roster |
No |
Experimental mode: no pair is admitted yet. Print mode is pinned on vibe 3.1.2, the roster is closed at deepseek-flash and glm-flash, and neither is Luna, Opus or Fable |
cli-pi or cli-opencode |
Luna |
No |
Intentional asymmetry: Luna is eligible through cli-codex only |
The denied list outranks the allowance list: Opus and Fable stay denied even where an allowance clause would otherwise admit them, such as Opus reached through cli-opencode.
The Luna asymmetry is intentional: Luna is eligible through cli-codex and not eligible through cli-pi or cli-opencode; do not "correct" it.
The allowance list is closed (default-deny). A pair that matches no allowance clause is not eligible; for example, Gemini reached through cli-cursor is not eligible.
An ineligible model must refuse the prompt-improve work, name the rule that denied it, and name an eligible (model, route) pair in its place, whether it was reached through this skill inline or through the agent. It must not do the work anyway with a caveat.
The route list above is a list of executors. The cli-jev transport, mode cli-jev of the cli-classifier hub, is deliberately absent: it answers one typed judgment and runs nothing, so no prompt-improvement task can be dispatched to it and it needs no eligibility row.
This table is the canonical home of the contract, so future edits land in one place. The dispatcher owns the choice: the /prompt:improve workflow dispatches @prompt-improver with model=sonnet, and the Claude Code copy of the agent pins model: sonnet. The agent keeps only a backstop that points here.
Deterministic Agent Rules
- Use
references/improvement-checklist.md for the one-pass checklist, effort levels, quick test and change report.
- Use
references/patterns-evaluation.md for the default skeleton and presets: read ## 2. FRAMEWORK LIBRARY & SELECTION, a named preset's subsection of ## 3. FRAMEWORK DEEP DIVES, and ## 10. CLEAR EVALUATION MASTERY, the one CLEAR rubric. An on-demand keyword reads the whole file.
- Map
complexity_hint onto effort as the payload table says; with no hint, use standard.
- Require
CLEAR >= 40/50 before returning success.
- If the screen scores below threshold, repair up to 3 cycles, then return the best version with explicit escalation notes.
Structured Output Block
FRAMEWORK: <default skeleton | preset name>
CLEAR_SCORE: <n>/50 (C:<n> L:<n> E:<n> A:<n> R:<n>)
TEST: NOT RUN
CHANGES: <Added ...; Changed ...; Removed ...; Skipped ...>
RATIONALE: <1-2 lines>
ENHANCED_PROMPT: |
<multi-line ready-to-dispatch prompt>
ESCALATION_NOTES: <remaining ambiguity, risk, or follow-up>
Contract Guarantees
- The returned prompt is ready for CLI handoff without another pass.
- The output preserves caller-supplied constraints unless the agent flags them as contradictory or underspecified.
- The agent remains read-only and leaf-only; it does not write files or spawn other sub-agents.
8. INTEGRATION POINTS
Framework Integration
This skill operates within the behavioral framework defined in AGENTS.md.
Key integrations:
- Gate 2: Skill routing via
.skilled/bin/skill-advisor.cjs with prompt-related intent boosters
- Tool Routing: Per AGENTS.md Section 5 decision tree
- Continuity: Context preserved by
/speckit:save, which writes through generate-context.js
Tool Usage Guidelines
- Read: Load reference files from references/ directory
- Write: Output enhanced prompts to user-specified location
- Glob: Discover available reference files in skill directory
- Bash: Run validation scripts if needed
9. REFERENCES AND RELATED RESOURCES
The router discovers reference, asset, and script docs dynamically. Start with references/improvement-checklist.md, references/patterns-evaluation.md, assets/format-guide-json.md, assets/format-guide-markdown.md, assets/format-guide-yaml.md, assets/cli-prompt-quality-card.md, then load task-specific resources from references/, templates from assets/, and automation from scripts/ when present.
RELATED PLAYBOOK
Manual validation lives at manual-testing-playbook/manual-testing-playbook.md.
Related skills: sk-doc for documentation outputs, sk-code for code-generation prompt context, and the cli-* skills that use the prompt quality card before dispatch.
1---2name: sk-prompt3description: "Prompt engineering: turns a request into a structured, checked AI prompt, with a component skeleton and CLEAR screening."4allowed-tools: [Read, Write, Edit, Bash, Glob, Grep]5version: 4.0.5.06metadata:7 source: https://skillmd.com/skills/michelkerkmeester/sk-prompt8---910<!-- Keywords: prompt-engineering, prompt-improvement, improve prompt, improve this prompt, improve the prompt, enhance prompt, prompt improvement, improvement-checklist, early-exit, CLEAR-scoring, framework-selection, RCAF, COSTAR, CRAFT, TIDD-EC, CRISPE -->1112# Prompt Engineering Specialist - One-Pass Prompt Improvement1314Transforms vague or basic inputs into structured AI prompts in one pass: a ten-row checklist, one component skeleton with optional framework presets, and CLEAR screening.1516**Core Principle**: Clarity, logic, expression, and reliability through structured methodology.1718---1920## 1. WHEN TO USE2122### Activation Triggers2324**Use when**:25- Enhancing or improving an AI prompt for any purpose26- Evaluating prompt quality with CLEAR scoring27- Selecting the right prompt framework for a given task28- Transforming vague requests into structured, effective prompts29- Supporting indirect invocation from `@prompt-improver` agent dispatches (the deep-path escalation target for CLI fast-path prompt cards)3031**Keyword Triggers**:32- `$improve`, `$text`, `$short`, `$refine`, `$json`, `$yaml`33- `$raw` (no checklist pass, fast pass-through)34- "improve my prompt", "enhance this prompt", "prompt engineering"35- "create a prompt for", "optimize this prompt"3637### Use Cases3839#### Text Prompt Enhancement40Transform vague requests into structured prompts from one default skeleton, with RCAF, COSTAR, RACE, CIDI, TIDD-EC, CRISPE or CRAFT as optional presets, and CLEAR screening (40+/50 threshold).4142#### Design-Generation Prompt43Construct a grounded, anti-default generation brief for a design-generation run. This skill owns the prompt only, never the measured design-reference extraction (`sk-design-md-generator`) or the run transport.4445### When NOT to Use4647**Skip this skill when:**48- Writing code or debugging (use sk-code skills instead)49- Creating documentation (use sk-doc instead)50- Simple text editing without prompt structure needs51- Direct API calls that do not need prompt optimization5253---5455## 2. SMART ROUTING5657The `$improve`, `$text`, `$refine`, `$short`, `$json`, `$yaml` and `$raw` prefixes select this skill, and so does a request to improve, enhance, score or restructure a prompt.58The prefix is authoritative: it fixes the operating mode, and the intent scores the router computes never change it.59With no prefix, weighted keyword scoring picks the top intent, or the top two when their scores sit within `AMBIGUITY_DELTA`.60A zero score falls back to `TEXT_ENHANCE` with the disambiguation checklist rather than guessing.6162`references/improvement-checklist.md` loads first on every run except `$raw` and the `FRAMEWORK` intent, and `references/patterns-evaluation.md` joins it for the text, design-generation, Markdown, JSON and YAML intents.63On `FRAMEWORK` it replaces the checklist default instead of joining it.64The `DESIGN_GEN` intent loads `references/patterns-evaluation.md`, the framework library and the CLEAR rubric, so the brief is built from a named framework rather than a design-specific reference this skill does not ship.65Read [`ROUTER.md`](./ROUTER.md) before selecting skill-local resources.66Open only the matched files.6768---6970## 3. HOW IT WORKS7172### Enhancement Pipeline7374Every improvement runs one pass, described in [improvement-checklist.md](./references/improvement-checklist.md):75761. Detect the mode from the command token or the user's own words. A pasted prompt is set aside before routing, so text inside it never picks the mode.772. Check the prompt once against the ten checklist rows, marking each present, absent or not needed with a reason. A prompt that passes exits early, unchanged or minimally edited.783. Write the prompt from the default component skeleton, applying a framework preset only when the request names one or its fit is clear from the task.794. Self-check once against the success criteria, run the quick test when the effort level calls for it, and screen with CLEAR.805. Deliver the prompt first, then the numbered open items and the change report.8182See the Smart Router Pseudocode in `ROUTER.md` for the complete routing logic.8384### Operating Modes8586| Mode | Command | Effort | Scoring | Use Case |87|------|---------|--------|---------|----------|88| Interactive | (default) | standard | CLEAR | Guided enhancement |89| Text | `$text` | standard | CLEAR | Standard text prompt |90| Short | `$short` | quick | CLEAR | Quick refinement |91| Improve | `$improve` | standard | CLEAR | Standard enhancement |92| Refine | `$refine` | thorough | CLEAR | Maximum optimization |93| JSON | `$json` | standard | CLEAR | API-ready format |94| YAML | `$yaml` | standard | CLEAR | Config format |95| Raw | `$raw` | none | None | No checklist pass |9697### Framework Presets9899Every prompt starts from one default skeleton: task, context, constraints, output format, optional examples and success criteria. The seven framework names stay as optional presets. Apply one only when the user or caller names it, or when its fit is clear from the task; the table shows where each fits.100101| Preset | Fits when |102|--------|-----------|103| RACE | Urgent tasks and quick iterations |104| RCAF | General tasks that need a clear role and format |105| COSTAR | Content where style, tone and audience matter |106| CIDI | Process documentation and tutorials |107| CRISPE | Strategy and exploration |108| TIDD-EC | Quality-critical or compliance work that needs do's, don'ts and examples |109| CRAFT | Complex projects and planning |110111See [patterns-evaluation.md](./references/patterns-evaluation.md) for complete framework details.112113### Scoring Systems114115**CLEAR** (50-point scale): Correctness (10) + Logic (10) + Expression (15) + Arrangement (10) + Reusability (5). Threshold: 40+. CLEAR is a structural screen, not proof the prompt works; repair up to 3 cycles.116117---118119## 4. RULES120121### ✅ ALWAYS1221231. **ALWAYS deliver first**124 - A request that carries a prompt or a task gets the improved prompt in the same reply125 - List every default and assumption as a numbered open item after the prompt, never inside it126 - Ask one question only when there is no prompt and no task, or when two mode commands conflict; `$raw` never asks1271282. **ALWAYS run the one-pass checklist** at the detected effort level129 - Quick for `$short`, standard by default, thorough for `$refine`, none for `$raw`130 - Mark each row present, absent or not needed with a one-line reason1311323. **ALWAYS exit early on a prompt that already passes**133 - Return it unchanged or with the smallest edit asked for, and say so1341354. **ALWAYS route on the user's own words**136 - Set a pasted prompt aside before detecting the mode1371385. **ALWAYS screen with CLEAR** before delivery139 - Write the reason before each score; pass at 40/50; repair up to 3 cycles1401416. **ALWAYS end with a change report**142 - Added, Changed, Removed, Skipped, plus `Quick test: NOT RUN` when the test did not run143144### ⛔ NEVER1451461. **NEVER stall on questions**147 - Deliver first: write from stated defaults and list each assumption as a numbered open item; ask one question only when there is no prompt and no task, or two mode commands conflict, and then wait for the answer1481492. **NEVER force a framework preset**150 - Use the default skeleton unless the request names a preset or its fit is clear1511523. **NEVER deliver without scoring**153 - Every improved prompt carries a CLEAR screen (except $raw)1541554. **NEVER write the enhanced prompt in the operator's voice**156 - Address the executing model: the role line in the second person ("You are..."), task steps in the imperative1571585. **NEVER exceed context with full reference loading**159 - Load only the references needed for the detected mode1601616. **NEVER rewrite user guidelines or examples needlessly**162 - Preserve them as closely as possible, word for word when asked; warn when the prompt grows past about 3 times the original163164### ⚠️ ESCALATE IF1651661. **ESCALATE IF there is no prompt and no task, or two mode commands conflict**167 - Ask one question, such as "Which mode do you want: $raw or $text?"; any other uncertainty is delivered with the assumption as open item 11681692. **ESCALATE IF CLEAR score stays below threshold after 3 repair cycles**170 - Deliver the best version with a note naming the weakest dimension1711723. **ESCALATE IF request conflicts with prompt engineering scope**173 - Redirect: "This appears to be a [code/doc/debug] task. Consider using [sk-code/sk-doc] instead."174175---176177## 5. REFERENCES178179### Core References180181- [improvement-checklist.md](./references/improvement-checklist.md) - One-pass checklist, early exit, effort levels, self-check, quick test and change report182- [patterns-evaluation.md](./references/patterns-evaluation.md) - default skeleton, seven framework presets, CLEAR scoring (canonical rubric)183184### Asset Files185186- [format-guide-markdown.md](./assets/format-guide-markdown.md) - Markdown format deep-dive: fundamentals, delivery standards, RCAF/CRAFT structures, advanced patterns, validation, best practices187- [format-guide-json.md](./assets/format-guide-json.md) - JSON format deep-dive: fundamentals, data types, delivery standards, RCAF/CRAFT structures, advanced patterns, validation, best practices188- [format-guide-yaml.md](./assets/format-guide-yaml.md) - YAML format deep-dive: fundamentals, data types, delivery standards, RCAF/CRAFT structures, advanced patterns, templates, validation, best practices189- [cli-prompt-quality-card.md](./assets/cli-prompt-quality-card.md) - CLI fast-path card: the framework selection table, the CLEAR pre-dispatch check and the escalation triggers the `cli-*` packets delegate to190- [framework-registry.json](./assets/framework-registry.json) - Machine-readable scaffold registry with code-oriented slot templates for a subset of five frameworks191192### Reference Loading Notes193194- Load only references needed for current intent195- Smart Routing in `ROUTER.md` is the single authority for loading rules196- SKILL.md (this file) is always loaded; conditionally load mode-specific references197198---199200## 6. SUCCESS CRITERIA201202### Enhancement Complete When203204- ✅ Mode detected from the user's own words, effort level set205- ✅ Checklist pass recorded, or early exit stated206- ✅ CLEAR screen applied and threshold met, or best version delivered with a note207- ✅ Prompt delivered first, with open items and the change report208- ✅ User can iterate or accept209210### Quality Targets211212- **CLEAR Score**: 40+ out of 50213- **Early exit**: a prompt that already passes the checklist comes back unchanged or minimally edited214215---216217## 7. AGENT INVOCATION CONTRACT218219`@prompt-improver` is the fresh-context escalation surface for this skill. The agent loads the references in this skill, applies the same framework-selection and CLEAR rules, and returns a structured block that the caller can inject into a CLI dispatch without loading the full skill inline.220221### Expected Input Payload222223| Field | Required | Description |224|-------|----------|-------------|225| `raw_task` | Yes | Raw task description or draft prompt to improve |226| `task_type` | No | One of `generation`, `review`, `research`, `edit`, `analyze` |227| `target_cli` | No | One of `claude-code`, `opencode`, `copilot` |228| `complexity_hint` | No | Integer `1-10` mapped to effort: 1-3 quick, 4-7 standard, 8-10 thorough |229| `constraints` | No | Compliance, security, audience, or output requirements |230231### Model Eligibility232233Dispatch `@prompt-improver`, or run this skill inline, only on an eligible (model, route) pair. Eligibility is a property of the (model, route) pair, never of the model alone.234235| Route | Model | Eligible | Note |236|-------|-------|----------|------|237| any route | Opus | No | Denied under no circumstances |238| any route | Fable | No | Denied under no circumstances |239| native Claude | Sonnet | Yes | — |240| native Claude | Haiku | Yes | — |241| `cli-claude-code` | Sonnet | Yes | — |242| `cli-claude-code` | Haiku | Yes | — |243| `cli-codex` | GPT Luna | Yes | — |244| `cli-pi` | any model except Luna | Yes | — |245| `cli-opencode` | any model except Luna | Yes | — |246| `cli-deepseek-harness` | DeepSeek V4.1 Flash through the LLM Gateway or OpenCode Go route patch | No | Experimental mode: no pair is admitted yet. The LLM Gateway route answered a keyed run on 2026-11-06 and OpenCode Go answered one on 2026-11-07. The roster is closed at these two routes, and neither is Luna, Opus or Fable |247| `cli-mistral-vibe` | DeepSeek V4.1 Flash or GLM Flash through the `mistral` provider roster | No | Experimental mode: no pair is admitted yet. Print mode is pinned on `vibe 3.1.2`, the roster is closed at `deepseek-flash` and `glm-flash`, and neither is Luna, Opus or Fable |248| `cli-pi` or `cli-opencode` | Luna | No | Intentional asymmetry: Luna is eligible through `cli-codex` only |249250The denied list outranks the allowance list: Opus and Fable stay denied even where an allowance clause would otherwise admit them, such as Opus reached through `cli-opencode`.251252The Luna asymmetry is intentional: Luna is eligible through `cli-codex` and not eligible through `cli-pi` or `cli-opencode`; do not "correct" it.253254The allowance list is closed (default-deny). A pair that matches no allowance clause is not eligible; for example, Gemini reached through `cli-cursor` is not eligible.255256An ineligible model must refuse the prompt-improve work, name the rule that denied it, and name an eligible (model, route) pair in its place, whether it was reached through this skill inline or through the agent. It must not do the work anyway with a caveat.257258The route list above is a list of executors. The `cli-jev` transport, mode `cli-jev` of the `cli-classifier` hub, is deliberately absent: it answers one typed judgment and runs nothing, so no prompt-improvement task can be dispatched to it and it needs no eligibility row.259260This table is the canonical home of the contract, so future edits land in one place. The dispatcher owns the choice: the `/prompt:improve` workflow dispatches `@prompt-improver` with `model=sonnet`, and the Claude Code copy of the agent pins `model: sonnet`. The agent keeps only a backstop that points here.261262### Deterministic Agent Rules263264- Use `references/improvement-checklist.md` for the one-pass checklist, effort levels, quick test and change report.265- Use `references/patterns-evaluation.md` for the default skeleton and presets: read `## 2. FRAMEWORK LIBRARY & SELECTION`, a named preset's subsection of `## 3. FRAMEWORK DEEP DIVES`, and `## 10. CLEAR EVALUATION MASTERY`, the one CLEAR rubric. An on-demand keyword reads the whole file.266- Map `complexity_hint` onto effort as the payload table says; with no hint, use standard.267- Require `CLEAR >= 40/50` before returning success.268- If the screen scores below threshold, repair up to 3 cycles, then return the best version with explicit escalation notes.269270### Structured Output Block271272```text273FRAMEWORK: <default skeleton | preset name>274CLEAR_SCORE: <n>/50 (C:<n> L:<n> E:<n> A:<n> R:<n>)275TEST: NOT RUN276CHANGES: <Added ...; Changed ...; Removed ...; Skipped ...>277RATIONALE: <1-2 lines>278ENHANCED_PROMPT: |279 <multi-line ready-to-dispatch prompt>280ESCALATION_NOTES: <remaining ambiguity, risk, or follow-up>281```282283### Contract Guarantees284285- The returned prompt is ready for CLI handoff without another pass.286- The output preserves caller-supplied constraints unless the agent flags them as contradictory or underspecified.287- The agent remains read-only and leaf-only; it does not write files or spawn other sub-agents.288289---290291## 8. INTEGRATION POINTS292293### Framework Integration294295This skill operates within the behavioral framework defined in AGENTS.md.296297Key integrations:298- **Gate 2**: Skill routing via `.skilled/bin/skill-advisor.cjs` with prompt-related intent boosters299- **Tool Routing**: Per AGENTS.md Section 5 decision tree300- **Continuity**: Context preserved by `/speckit:save`, which writes through `generate-context.js`301302### Tool Usage Guidelines303304- **Read**: Load reference files from references/ directory305- **Write**: Output enhanced prompts to user-specified location306- **Glob**: Discover available reference files in skill directory307- **Bash**: Run validation scripts if needed308309---310311## 9. REFERENCES AND RELATED RESOURCES312313The router discovers reference, asset, and script docs dynamically. Start with `references/improvement-checklist.md`, `references/patterns-evaluation.md`, `assets/format-guide-json.md`, `assets/format-guide-markdown.md`, `assets/format-guide-yaml.md`, `assets/cli-prompt-quality-card.md`, then load task-specific resources from `references/`, templates from `assets/`, and automation from `scripts/` when present.314315### RELATED PLAYBOOK316317Manual validation lives at `manual-testing-playbook/manual-testing-playbook.md`.318319Related skills: `sk-doc` for documentation outputs, `sk-code` for code-generation prompt context, and the `cli-*` skills that use the prompt quality card before dispatch.