tcrei-prompt

name	tcrei-prompt
description	TCREI prompt structuring tool. Rewrites rough prompts into Google's TCREI (Task, Context, References, Evaluate, Iterate) format so they produce consistent results when reused in the next session. Triggers: "TCREI", "structure this prompt", "prompt enhance", "make a prompt for next session", "rewrite as TCREI", "improve this prompt".
version	1.0.0
allowed-tools	["Read","Write","Glob","Grep","AskUserQuestion","Task"]

TCREI Prompt Transformer

Hermes Agent Compatibility

When this skill is loaded through Hermes as tcrei-prompt:tcrei-prompt, map Claude/Codex tool names to Hermes tools:

Claude/Codex term	Hermes tool
Read	read_file
Write	write_file
AskUserQuestion	clarify
Task	delegate_task

Treat $ARGUMENTS as the natural-language arguments supplied when the user asks Hermes to load the skill. Plugin-provided skills are explicit opt-in loads in Hermes; use skill_view("tcrei-prompt:tcrei-prompt") (or ask Hermes to load that qualified skill) rather than relying on bare text.

Rewrites rough prompts into Google's TCREI 5-step structure and outputs copy-paste-ready prompts for the next session. Based on Google's Prompting Essentials (Coursera): Task, Context, References, Evaluate, Iterate.

Trigger Examples

Context: User has a rough prompt and wants it structured user: "Turn this into a TCREI prompt: write a blog post" assistant: Loads TCREI skill, diagnoses missing elements, and starts interview. Explicit TCREI keyword with a target prompt provided - direct trigger. Context: User wants a reusable prompt for a future session user: "Make a prompt I can use next time for this task" assistant: Loads TCREI skill to produce a structured, session-transferable prompt. "next time" + "prompt" signals session-transfer intent. Context: User wants to improve an existing prompt user: "This prompt gives inconsistent results, make it better" assistant: Loads TCREI skill, diagnoses gaps, and interviews for missing elements. Prompt quality improvement request - systematic TCREI restructuring.

Critical Rules

1. Use AskUserQuestion for all questions - never ask in plain text
2. Skip elements already present - if Task is clear in the original, do not ask about Task
3. Interview only missing elements - do not ask about all 5; collect only what is missing
4. Output must be a copy-paste-ready prompt - not an explanation or tutorial
5. Self-verify with OMC verifier - auto-check quality after generation
6. Save to file - store the final prompt at .claude/prompts/{YYYY-MM-DD}-{name}.md

Phase 0: Diagnose the Original Prompt

Analyze the user's prompt and determine which TCREI elements are present vs missing.

Diagnosis Criteria

Element	Present if...	Missing if...
Task	Has 2+ of: action verb, output format, constraints	Only a vague verb ("do this", "write something")
Context	Specifies 1+ of: audience, purpose, source material	No background information at all
References	Includes samples, format specs, schemas, or tone examples	No hint about desired output shape
Evaluate	Has 1+ of: must-have, must-not, verification criteria	No quality standards defined
Iterate	Mentions improvement direction or version requests	One-shot request only

Diagnosis Output

Show the diagnosis to the user, then proceed to interview:

## TCREI Diagnosis

| Element | Status | Found |
|---------|--------|-------|
| Task | Partial | "write a blog post" - verb only, no output format or constraints |
| Context | Missing | - |
| References | Missing | - |
| Evaluate | Missing | - |
| Iterate | Missing | - |

Starting interview to fill in missing elements.

Phase 1: TCREI Interview (Interview Methodology)

Interview the user for missing elements only, using AskUserQuestion. Follow interview-methodology principles:

• No obvious questions
• Follow-up on each answer
• Provide options for easy answering

Interview Questions by Element

When Task is missing

Question: "What output format do you need?"
Options:
1. "Markdown document" - Structured text with headers
2. "JSON / structured data" - Machine-readable, tool-friendly
3. "Free-form text" - Essay, letter, narrative
4. "Table / comparison" - Side-by-side data

Question: "Any length or format constraints?"
Options:
1. "Short (5 sentences or less)" - Key points only
2. "Medium (1 page)" - Moderate detail
3. "Long (comprehensive)" - Full coverage
4. "Specific format" - Custom input

When Context is missing

Question: "Who is the audience for this output?"
Options:
1. "Myself only" - Personal reference
2. "Team / colleagues" - Internal sharing
3. "Clients / customers" - External delivery
4. "General public" - Published content

Question: "What is the purpose of this task?"
Options:
1. "Inform / educate" - Help understand something
2. "Persuade / market" - Drive action
3. "Document / archive" - Record keeping
4. "Support decisions" - Aid judgment

When References are missing

Question: "Do you have a preferred tone or style?"
Options:
1. "Professional / formal" - Reports, official docs
2. "Friendly / conversational" - Blog, social media
3. "Concise / bullet-point" - Summaries, briefs
4. "Provide a sample" - I'll give an example

Question: "Do you want a fixed output structure? (e.g., JSON schema)"
Options:
1. "Free form" - Let AI decide the structure
2. "Fixed key structure" - JSON/YAML schema
3. "Match existing template" - Provide a file path

When Evaluate is missing

Question: "What must the output include?"
multiSelect: true
Options:
1. "N key points" - Specify a number
2. "Call-to-action (CTA)" - Reader should do something
3. "Data / evidence" - Sourced information
4. "Custom requirement" - Specify directly

Question: "What must the output avoid?"
multiSelect: true
Options:
1. "Exaggeration / hype" - Facts only
2. "Jargon overload" - Keep it simple
3. "Unsourced claims" - Everything needs evidence
4. "Custom restriction" - Specify directly

When Iterate is missing

Iterate instructions are auto-generated (no interview needed). Generate 3 context-specific refinement prompts based on the assembled TCREI content.

Interview Guidelines

1. Max 2 questions per round - batch into AskUserQuestion's questions array
2. 1-2 missing elements = 1 round; 3-4 missing = 2 rounds - keep it short
3. Cross-pollinate answers - a Context answer may fill in Task details
4. Follow up on "Custom" selections - probe open-ended answers into specifics

Phase 2: Generate TCREI Prompt

Assemble the structured prompt from collected information.

Output Template

# TCREI Prompt: [Title]

> Structured with Google's TCREI framework. Copy and paste directly to any AI.

---

## Task

[Verb] to produce [output].

**Constraints:**
- [Format / length / tone constraint 1]
- [Constraint 2]
- [Constraint 3]

## Context

- **Audience**: [Who will read this]
- **Purpose**: [Why this is needed]
- **Source material**: [What to base it on - attach below or mark "to be attached"]

[Additional background if any]

## References

**Style guide:**
- [Tone / voice rule 1]
- [Tone / voice rule 2]

**Output samples (for tone reference):**
> "[Sample sentence 1]"
> "[Sample sentence 2]"

[If JSON schema is needed:]
```json
{
  "key1": "description",
  "key2": "description"
}

Evaluate

After completing the task, self-check against these criteria:

Must-have:

• [Required element 1]
• [Required element 2]

Must-not:

• [Prohibited element 1]
• [Prohibited element 2]

Verify:

• [Verification check 1]
• [Verification check 2]

Iterate

Use these refinement prompts if the output needs improvement:

1. "[Specific refinement 1 - includes numbers/metrics]"
2. "[Specific refinement 2 - includes action directive]"
3. "[Specific refinement 3 - includes format/structure change]"

### Generation Rules

1. **Each section must be self-contained** - understandable without reading other sections
2. **Use specific numbers** - "5 sentences" not "short"; "1 page" not "moderate"
3. **Iterate prompts must be context-aware** - not generic "make it better" but task-specific
4. **Must be copy-paste ready** - no meta-commentary or instructional text inside the prompt body

## Phase 3: Self-Verification (OMC Iterate)

Verify the generated prompt using the OMC verifier agent.

### Verification Process

Delegate to verifier via Task tool:

Task( subagent_type="claude", model="haiku", prompt="Verify this TCREI prompt for quality: 1. Task: Has action verb + output format + constraints? 2. Context: Has 2+ of audience / purpose / source material? 3. References: Has 1+ of tone / format / sample? 4. Evaluate: Has 2+ of must-have / must-not / verify? 5. Iterate: Has 2+ refinement prompts with specific numbers/rules/actions? 6. Is the entire prompt copy-paste ready (no meta-commentary)?

Pass threshold: 5 of 6 criteria met
On fail: List deficient items with specific improvement suggestions

[Full generated prompt here]"

)

### On Verification Failure

1. Auto-fix the deficient items identified by the verifier
2. Re-verify (max 2 retry rounds)
3. If still below threshold after 2 rounds, output current state and notify user of gaps

## Phase 4: Final Output and Save

### Output

1. Display the completed TCREI prompt in the conversation
2. Save to `.claude/prompts/{YYYY-MM-DD}-{name}.md`
   - Create `.claude/prompts/` directory if it does not exist
   - Filename in kebab-case
3. Show verification summary:

Verification Results

Criterion	Status
Task (verb + output + constraints)	Pass
Context (audience + purpose + source)	Pass
References (tone + format + sample)	Pass
Evaluate (must-have + must-not + verify)	Pass
Iterate (numbers + rules + actions)	Pass
Copy-paste ready	Pass

Saved to: .claude/prompts/2026-02-16-blog-post.md

## Domain-Specific TCREI Gap Patterns

Common gaps by domain with default remediation:

| Domain | Commonly missing | Default remediation |
|--------|-----------------|-------------------|
| **Development** | References (code style), Evaluate (test criteria) | Match existing codebase patterns, include test cases |
| **Marketing** | Context (target persona), Evaluate (conversion criteria) | Define persona, add CTA verification |
| **Education** | References (difficulty samples), Iterate (learning progression) | Separate beginner/intermediate, add step-by-step deepening |
| **Documents** | Context (reader level), References (existing templates) | Distinguish exec/practitioner, reference company templates |
| **Translation** | References (glossary), Evaluate (consistency) | Unify technical terms, verify tone consistency |

## Interview Anti-Patterns

1. **Asking about all 5 elements** - skip what is already present; it wastes time
2. **Abstract questions** - offer concrete options instead of "what vibe do you want?"
3. **5 questions at once** - batch max 2 per AskUserQuestion call
4. **Ignoring cross-signals** - a Context answer may reveal Task or Reference details
5. **Interviewing for Iterate** - auto-generate is more effective (context-aware specifics beat generic user input)