SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillssubagent-instruction-operationalization

subagent-instruction-operationalization

Writing & research

View repository

Updated 10-15 09:46

About this skill

Subagent Instruction Operationalization

Overview

Translate a completed style specification into a structured LLM subagent prompt that can reproduce a target voice. The core principle: every constraint in the style specification must become either a concrete, measurable prompt directive or be explicitly documented as aspirational. A style specification is an analytical artifact (what the voice IS); a subagent instruction is a generative artifact (how to PRODUCE the voice). This skill bridges the gap by converting each upstream analysis output into a prompt section with verification criteria.

Why this is hard: LLMs respond to instructions with varying fidelity depending on constraint type. Lexical constraints ("use grade-level 12 vocabulary") are moderately enforceable. Structural constraints ("lead with evidence before claims") are somewhat enforceable. Distributional constraints ("maintain a 40/30/20/10 split of explaining/advising/challenging/asserting") are weakly enforceable without few-shot examples. The prompt must be designed with awareness of these enforcement gradients.

Research foundation: Persona-based prompting (role assignment) produces measurably different output style, tone, and vocabulary (Brim Labs, 2025; PromptHub, 2025). Detailed persona specifications consistently outperform vague ones, and LLM-generated personas often outperform human-written ones (LearnPrompting, 2025). Few-shot examples are often more effective than lengthy instructions for style constraints (Palantir, 2025). Stylometric research confirms that function-word patterns, sentence-length distributions, and punctuation habits are the most stable authorship markers -- and therefore the most important features to replicate (Mosteller & Wallace, 1963; Kestemont, 2014). Recent work on LLM style transfer shows that instruction-tuned models can approximate target styles with one-shot accuracy between 67.6% and 94.7% depending on prompting strategy (arXiv:2509.24930).

When to Use

  • Style specification from upstream pipeline is complete (or partially complete -- see Insufficient Data Handling)
  • Need to produce a prompt that an LLM subagent can use to generate text in a target voice
  • Translating psycholinguistic analysis (LIWC, Big Five, MDPI archetype, stylometrics, readability, CAT, rhetorical structure, speech acts) into generative instructions
  • Building a voice-replication prompt for ANY text corpus, not limited to any specific platform
  • Creating a testable, versioned persona prompt with verification criteria

When NOT to use:

  • No style specification exists yet (run the upstream analyses first)
  • Goal is to impersonate someone for deception, fraud, or harassment (ethical boundary -- this skill produces writing style approximations for legitimate purposes such as content assistance, ghostwriting with consent, or creative writing)
  • Target voice is "generic professional" or "standard formal" -- these do not require specification-driven prompting; a simple role prompt suffices
  • Style specification contains only one or two dimensions (e.g., only readability) -- the prompt would be too thin to constitute a voice; use the individual constraint directly
digraph operationalization_decision {
    rankdir=TB;
    "Style specification\nexists?" [shape=diamond];
    "3+ dimensions\ncompleted?" [shape=diamond];
    "Legitimate\nuse case?" [shape=diamond];
    "Few-shot examples\navailable?" [shape=diamond];
    "Full prompt\ngeneration" [shape=box];
    "Run upstream\nanalyses first" [shape=box];
    "Reduced prompt with\ngaps documented" [shape=box];
    "STOP: Ethical\nboundary" [shape=box];
    "Prompt without\nfew-shot section" [shape=box];

    "Style specification\nexists?" -> "3+ dimensions\ncompleted?" [label="yes"];
    "Style specification\nexists?" -> "Run upstream\nanalyses first" [label="no"];
    "3+ dimensions\ncompleted?" -> "Legitimate\nuse case?" [label="yes"];
    "3+ dimensions\ncompleted?" -> "Reduced prompt with\ngaps documented" [label="no, 1-2 only"];
    "Legitimate\nuse case?" -> "Few-shot examples\navailable?" [label="yes"];
    "Legitimate\nuse case?" -> "STOP: Ethical\nboundary" [label="no"];
    "Few-shot examples\navailable?" -> "Full prompt\ngeneration" [label="yes"];
    "Few-shot examples\navailable?" -> "Prompt without\nfew-shot section" [label="no"];
}

Quick Reference

Prompt Section Architecture

The subagent prompt is organized into sections, each derived from a specific upstream analysis. The order below reflects enforcement priority -- strongest constraints first.

Section Source Analysis Enforcement Strength What It Controls
Role Frame Persona archetype + MDPI archetype Strong Overall persona identity and emotional register
Few-Shot Examples Original corpus samples Strong Concrete style demonstrations (most powerful constraint)
Complexity Targets Readability + Lexical Diversity Strong (measurable) Grade level, sentence length, vocabulary diversity
Syntactic Stability Stylometric Fingerprint Moderate-Strong Function-word rates, punctuation patterns, sentence-length distribution
Emotional Valence MDPI Archetype + VADER Moderate Affective register boundaries, toxicity limits
Informational Density LIWC dimensions + Archetype Moderate Noun-heavy vs. conversational, process/content word ratio
Evidence Framing Big Five (Conscientiousness) + Archetype Moderate Degree claims require support, qualification habits
Rhetorical Structure Rhetorical/Discourse Analysis Moderate-Weak Argument ordering, hedging, post structure
Pragmatic Distribution Speech Act Analysis Weak (needs examples) Ratio of explaining vs. advising vs. challenging
Register Rules Register Variation Analysis Conditional Context-dependent style switching rules
Community Convergence CAT / LSM Weak (ambient) Target community standards for function-word convergence

Enforcement Gradient

Strength Meaning Prompt Strategy
Strong LLM reliably follows Direct numerical constraints + few-shot examples
Moderate LLM follows with drift Constraints + self-check instructions + exemplars
Weak LLM approximates Few-shot examples as primary driver; instructions as secondary
Conditional Context-triggered If/then rules with explicit context detection

Minimum Specification Completeness

Dimensions Completed Viability Action
7+ of 9 dimensions Full prompt Generate complete subagent instruction
4-6 dimensions Partial prompt Generate prompt with gap documentation; mark missing sections as "DEFAULT: use natural style"
1-3 dimensions Minimal prompt Generate targeted constraints only; do NOT frame as a full voice replication prompt
0 dimensions Not viable STOP. Run upstream analyses.

Workflow

Copy this checklist and track progress:

Subagent Instruction Operationalization Progress:
- [ ] Step 1: Inventory the style specification and assess completeness
- [ ] Step 2: Select few-shot examples from the source corpus
- [ ] Step 3: Compose the Role Frame
- [ ] Step 4: Translate complexity targets into numerical constraints
- [ ] Step 5: Encode syntactic stability constraints
- [ ] Step 6: Define emotional valence boundaries
- [ ] Step 7: Set informational density directives
- [ ] Step 8: Specify evidence framing requirements
- [ ] Step 9: Encode rhetorical structure patterns
- [ ] Step 10: Define pragmatic distribution targets
- [ ] Step 11: Write register rules (if context-dependent)
- [ ] Step 12: Encode community convergence as ambient style guidance
- [ ] Step 13: Add self-verification instructions
- [ ] Step 14: Assemble the complete prompt and check token budget
- [ ] Step 15: Test the prompt against source material
- [ ] Step 16: Write findings to docs/analysis/25-subagent-instruction.md

Step 1: Inventory the Style Specification

Before writing any prompt directives, audit which upstream analyses are available and their confidence levels.

Inventory table:

Dimension Source Report Status Confidence Key Constraints to Extract
Informational Density 16-liwc-psycholinguistic.md [done/partial/missing] [high/moderate/low] Process/content word ratio, dominant dimension
Emotional Valence 14-mdpi-hypernetwork-archetype.md [done/partial/missing] [high/moderate/low] Archetype label, valence constraints
Evidence Framing 15-big-five-personality.md [done/partial/missing] [high/moderate/low] Conscientiousness score, assertion style
Syntactic Stability 18-stylometric-fingerprinting.md [done/partial/missing] [high/moderate/low] Function-word rates, sentence-length distribution
Complexity Targets 19-readability-lexical-diversity.md [done/partial/missing] [high/moderate/low] FK grade, MTLD, sentence length range
Rhetorical Structure 20-rhetorical-discourse-structure.md [done/partial/missing] [high/moderate/low] Argument ordering, hedging rate, structure patterns
Pragmatic Distribution 22-speech-act-pragmatic.md [done/partial/missing] [high/moderate/low] Speech act ratios
Register Rules 21-register-variation.md [done/partial/missing] [high/moderate/low] Stable vs. context-dependent classification
Community Convergence 17-cat-linguistic-style-matching.md [done/partial/missing] [high/moderate/low] LSM targets, top-accommodated communities

If fewer than 4 dimensions are complete: Generate a reduced prompt. Document which sections are missing and what the default behavior will be. Do NOT fabricate constraints from missing data.

Step 2: Select Few-Shot Examples

Few-shot examples are the single most powerful style constraint. They demonstrate what instructions can only describe.

Selection criteria:

Criterion Requirement
Typicality Choose examples that are stylistically typical, not outliers. Use typicality scores from archetype analysis if available.
Topic diversity Select examples spanning at least 2-3 different topics to show that the style persists across subjects.
Length variety Include both short (1-3 sentences) and medium (1-2 paragraphs) examples to demonstrate the voice at different scales.
Recency Prefer recent examples that reflect the current voice, unless temporal analysis shows style stability.
Count 3-5 examples is optimal. More than 7 consumes excessive context without proportional benefit.

Example framing in the prompt:

Here are examples of the target voice. Study the sentence structure,
word choice, hedging patterns, and argument flow -- not the topic content.

Example 1 (technical context):
[verbatim text from corpus]

Example 2 (casual context):
[verbatim text from corpus]

Example 3 (disagreement context):
[verbatim text from corpus]

When few-shot examples are unavailable: If the original corpus is not accessible, rely on the specification constraints alone but note in the prompt: "No reference examples available. Follow the constraints below as closely as possible. Err toward the conservative end of each range."

Step 3: Compose the Role Frame

The role frame is the first section of the prompt. It establishes identity, not mechanics.

Template:

You are a writing voice replication agent. Your task is to generate text
that matches a specific author's documented writing style. You are NOT
the author -- you are producing an approximation based on a detailed
style analysis. The output should be indistinguishable in STYLE from
the author's writing, while the CONTENT is determined by the task at hand.

Voice profile: [archetype label, e.g., "HHL -- high-visibility, positive,
constructive communicator"] with [personality summary, e.g., "moderate-high
Openness, high Conscientiousness"].

Primary register: [from LIWC, e.g., "Analytical-reflective with
social-relational secondary emphasis"].

Rules for the role frame:

  • State explicitly that this is an approximation, not identity theft
  • Include the MDPI archetype label and its behavioral signature
  • Include the Big Five summary (trait positions, not raw scores)
  • Include the LIWC register classification
  • Keep under 150 words -- detail belongs in constraint sections below

Step 4: Complexity Targets

Translate readability and lexical diversity measurements into numerical prompt constraints. These are among the most reliably enforceable constraints.

Directive template:

## Complexity Constraints

Target readability: Flesch-Kincaid grade level [X-Y] (consensus range
from [N] formulas). The measured corpus average is [X.X] with standard
deviation [X.X].

Sentence length: Target [X-Y] words per sentence on average. Vary
individual sentences between [min] and [max] words. Do NOT produce
uniform sentence lengths.

Vocabulary diversity: Maintain lexical diversity (MTLD) in the range
[X-Y]. This means introducing new vocabulary at a rate consistent with
[high/moderate/low] diversity. In practice: [concrete guidance, e.g.,
"avoid repeating the same content word within 3 sentences unless it is
a technical term"].

Word length: Target average word length of [X.X] characters.
[X%] of words should be 6+ characters (long words).

Mapping from upstream data:

Upstream Metric Prompt Constraint
FK grade level (consensus range across formulas) "Target FK grade [X-Y]"
Average sentence length + std dev "Average [X-Y] words/sentence, range [min-max]"
MTLD score "MTLD [X-Y]" with concrete vocabulary reuse guidance
MATTR score Secondary check: "Moving-average TTR should approximate [X.XX]"
Hapax ratio Informational: "Approximately [X%] of words should appear only once"
Average word length + long-word % "Average [X.X] chars/word; [X%] words 6+ chars"

Step 5: Syntactic Stability

Encode the stylometric fingerprint as structural constraints. These maintain the author's unconscious grammatical habits.

Directive template:

## Syntactic Fingerprint

Function word targets (percentage of total words):
- Articles (a, an, the): [X.X%] (+/- [X.X%])
- Prepositions: [X.X%] (+/- [X.X%])
- Personal pronouns: [X.X%] (+/- [X.X%])
  - I/me/my: [X.X%] | we/us/our: [X.X%] | you/your: [X.X%]
- Conjunctions: [X.X%] (+/- [X.X%])
- Auxiliary verbs: [X.X%] (+/- [X.X%])

Punctuation patterns:
- Comma rate: [X.X] per sentence
- Semicolons: [rare/occasional/frequent]
- Dashes (em/en): [rare/occasional/frequent]
- Exclamation marks: [rare/occasional/frequent]
- Parenthetical asides: [rare/occasional/frequent]

Sentence length distribution:
- Mean: [X] words | Median: [X] words | Std dev: [X]
- [X%] short sentences (< 10 words)
- [X%] medium sentences (10-25 words)
- [X%] long sentences (> 25 words)

Paragraph length: Average [X] sentences per paragraph. Range [X-Y].

Only include features with stability coefficient CV < 0.30 from the stylometric fingerprint. Unstable features (CV >= 0.30) should be omitted or marked as aspirational.

Step 6: Emotional Valence Boundaries

Translate the MDPI archetype into emotional register rules.

Directive template based on archetype:

Archetype Valence Directive
HHH "Express positive sentiment as the baseline. Aggressive emphasis (strong language, exclamations, intensifiers) is permitted and expected. High emotional arousal. Positive intent with forceful delivery."
HHL "Maintain a consistently positive, constructive tone. Avoid aggressive language, profanity, or confrontational framing. Supportive and measured. Low arousal, high warmth."
HLH "Express criticism and negative assessments directly. Confrontational and assertive. Toxic/aggressive language markers are part of the voice. High arousal, combative register."
HLL "Express criticism through substantive analysis, not aggressive language. Skeptical and evaluative but restrained. Analytical negativity without hostility."
LHH "Positive intent expressed through confrontational or aggressive delivery. 'Tough love' framing. Supportive conclusions reached through blunt, sometimes profane, language."
LHL "Quietly positive and encouraging. Gentle, affirming tone. Low-key delivery. Avoid assertiveness or dominance."
LLH "Negative and hostile. Bitter or resentful framing. Express discontent through aggressive language. Low-visibility, high-toxicity patterns."
LLL "Subdued negativity. Resigned or withdrawn tone. Express dissatisfaction through measured, low-energy commentary. Avoid both enthusiasm and aggression."

Include VADER calibration: "The target sentiment compound score across outputs should average approximately [X.XX] (range [X.XX to X.XX])."

Step 7: Informational Density

Translate LIWC dimension profile and process/content word split into density directives.

Directive template:

## Informational Density

Content density: [content-dense / balanced / process-heavy]
- Function word percentage target: [X%] (baseline: 55-65%)
- Content word percentage target: [X%]
- This voice [is / is not] noun-heavy.

Dominant psycholinguistic registers (from LIWC):
1. [Primary register, e.g., "Cognitive-analytical: elevated insight
   and causation language"]
2. [Secondary register, e.g., "Social-relational: above-baseline
   social process words"]

Register guidance:
- [Elevated dimension]: Use [specific word types] more frequently
  than a generic voice would. Target [X%] vs. baseline [Y%].
- [Reduced dimension]: Use [specific word types] less frequently.
  Target [X%] vs. baseline [Y%].

Step 8: Evidence Framing

Derive from Big Five Conscientiousness score and archetype.

Mapping:

Conscientiousness Archetype Score Evidence Framing Directive
High (65+) High "Support claims with evidence, examples, or reasoning. Use discourse markers (therefore, because, given that). Qualify assertions with scope limitations. Structure arguments with clear premises before conclusions."
High (65+) Low "Support claims when possible but accept that lower-visibility contexts permit more casual assertion."
Average (35-65) Any "Mix supported and unsupported claims naturally. Some assertions can stand alone; complex or controversial claims should include reasoning."
Low (0-35) Any "State opinions directly without obligatory evidence scaffolding. Casual assertion is the default. Include reasoning only when the point is complex or when challenged."

Step 9: Rhetorical Structure

Encode argument ordering and rhetorical device usage.

Directive template:

## Rhetorical Structure

Argument ordering: [claim-first / evidence-first / dialectical /
exploratory / narrative-embedded]
- [Concrete description: e.g., "State your position in the first
  sentence, then provide 2-3 supporting points, then anticipate
  a counter-argument."]

Hedging frequency: [X.X%] of assertions include hedging markers
(I think, perhaps, might, arguably, it seems).
- [Concrete guidance: e.g., "Hedge approximately 1 in every 4
  evaluative claims. Never hedge factual statements."]

Concession patterns: [rare / occasional / frequent]
- [If frequent: "Acknowledge opposing views before presenting
  your own position. Use 'to be fair', 'granted', 'I see your
  point but'."]

Rhetorical questions: [rare / occasional / frequent]
- [If used: "Approximately [X] per [Y] paragraphs."]

List vs. prose ratio: [X%] of multi-point arguments use enumerated
lists vs. flowing prose.

Post structure:
- Opening: [direct claim / question / anecdote / context-setting]
- Body: [sequential points / point-counterpoint / narrative]
- Closing: [summary / call to action / open question / none]

Step 10: Pragmatic Distribution

Translate speech act analysis into output ratios.

Directive template:

## Pragmatic Distribution (Speech Act Targets)

Target distribution across output:
- Explaining (causal reasoning, teaching): ~[X%]
- Advising (recommendations, suggestions): ~[X%]
- Asserting (declarative claims, opinions): ~[X%]
- Challenging (counter-arguments, disagreement): ~[X%]
- Questioning (interrogatives, prompts): ~[X%]
- Supporting (agreement, affirmation): ~[X%]

Primary mode: [The dominant speech act, e.g., "This voice primarily
EXPLAINS -- most responses walk through reasoning rather than
declaring conclusions."]

Note: These are approximate targets. Do not mechanically count
speech acts. Instead, use the primary mode as the default and
deviate when context requires it.

Enforcement note: Pragmatic distribution is weakly enforceable through instructions alone. The few-shot examples (Step 2) are the primary mechanism for establishing this distribution. The directive above provides a calibration target.

Step 11: Register Rules (Conditional)

Only include this section if the register variation analysis classified the voice as context-dependent.

Directive template (for context-dependent voices):

## Register Rules

This voice shifts style based on context. Apply these rules:

WHEN writing about [technical/professional domain]:
- Increase formality by [specific changes]
- Reduce hedging frequency to [X%]
- Increase evidence framing density
- Target FK grade [X-Y] (higher than default)

WHEN writing in [casual/social context]:
- Decrease sentence length to [X-Y] words average
- Increase pronoun usage (I/you) by [X%]
- Permit informal constructions and contractions
- Target FK grade [X-Y] (lower than default)

WHEN disagreeing or challenging:
- [Archetype-specific: e.g., "Maintain analytical tone (HLL),
  do not escalate to aggressive framing"]

DEFAULT (no specific context detected):
- Use the primary register described above.

For stable-register voices: Replace this section with: "This voice maintains a consistent register across contexts. Do not adjust style based on topic or audience. Apply all constraints uniformly."

Step 12: Community Convergence

Translate CAT/LSM findings into ambient style guidance. This is not a mechanical constraint but a generative principle.

Directive template:

## Community Convergence

This voice naturally converges toward the norms of [top 1-2
communities by LSM score]. When generating text:

- Function word usage should approximate the norms of [community
  name] (composite LSM: [X.XX])
- Specifically, match these function-word rates from the target
  community: [list 2-3 highest-convergence categories with rates]
- The voice [accommodates / resists accommodation to] [community
  name], so [match / diverge from] their norms for [specific
  categories]

This is a soft constraint. It shapes the ambient tone rather than
overriding the specific structural constraints above.

Step 13: Self-Verification Instructions

Build verification steps into the prompt itself. This catches drift during generation.

Directive template:

## Self-Verification

After generating text, verify these checkpoints before finalizing:

1. READABILITY: Is the output within FK grade [X-Y]? Are sentences
   averaging [X-Y] words?
2. TONE: Does the sentiment match the [archetype] profile? Is the
   emotional register [specific description]?
3. STRUCTURE: Does the argument follow [ordering pattern]? Are
   hedges present at approximately the target frequency?
4. VOICE MARKERS: Check for the presence of characteristic patterns:
   [list 3-5 specific markers from the fingerprint, e.g., "frequent
   use of parenthetical asides", "semicolons linking independent
   clauses", "rhetorical questions before key points"]
5. ANTI-MARKERS: Check for the ABSENCE of patterns this voice
   does NOT use: [list 2-3 patterns to avoid, e.g., "never uses
   exclamation marks", "avoids bullet-point lists", "does not
   use emoji"]

If any checkpoint fails, revise the output before presenting it.

Step 14: Assemble and Check Token Budget

Combine all sections into the final prompt. Check total length against context limits.

Token budget guidelines:

Prompt Component Target Tokens Maximum Tokens
Role Frame 100-150 200
Few-Shot Examples 300-800 1,500
Complexity Targets 100-150 200
Syntactic Stability 150-250 400
Emotional Valence 50-100 200
Informational Density 80-120 200
Evidence Framing 50-80 150
Rhetorical Structure 100-200 300
Pragmatic Distribution 80-120 200
Register Rules 0-200 300
Community Convergence 50-100 150
Self-Verification 100-150 250
Total 1,160-2,420 4,050

If over budget: Prioritize by enforcement strength. Cut community convergence and pragmatic distribution directives first (they are weakly enforceable). Never cut few-shot examples or complexity targets (they are strongly enforceable).

Compression strategies:

  • Merge register rules into other sections when only 1-2 context rules exist
  • Reduce few-shot examples from 5 to 3
  • Combine syntactic stability and complexity targets into one section
  • Remove aspirational constraints that cannot be verified

Step 15: Test the Prompt

Generate 3-5 test outputs using the assembled prompt and compare against the source corpus.

Test procedure:

  1. Generate text on 3 different topics using the prompt
  2. Measure readability (FK grade, sentence length) of generated output
  3. Compare function-word frequencies between generated output and source corpus fingerprint
  4. Assess emotional valence of generated output against archetype target
  5. Check structural patterns: argument ordering, hedging frequency, post structure
  6. Calculate a style similarity score if possible (cosine distance on stylometric feature vectors)

Pass/fail criteria:

Metric Pass Marginal Fail
FK grade level Within +/- 1 of target range Within +/- 2 Outside +/- 2
Sentence length avg Within +/- 3 words of target Within +/- 5 Outside +/- 5
Sentiment compound Within +/- 0.15 of target Within +/- 0.25 Outside +/- 0.25
Function word rates (top 5) Within +/- 1.5% of fingerprint Within +/- 3% Outside +/- 3%
Argument ordering Matches specified pattern Partial match Wrong pattern
Hedging frequency Within +/- 50% of target rate Within +/- 100% No hedging or all hedging

If tests fail: Revise the failing constraint section. The most common fixes:

  • Add or improve few-shot examples (if structural constraints fail)
  • Make numerical targets more explicit (if readability drifts)
  • Add anti-markers to self-verification (if unwanted patterns appear)
  • Simplify over-constrained sections that the LLM cannot satisfy simultaneously

Step 16: Write Report

Write all findings to docs/analysis/25-subagent-instruction.md. See the report template below.

Report Output Template

The final report MUST be written to docs/analysis/25-subagent-instruction.md with this structure:

# Subagent Instruction Operationalization

## Style Specification Inventory

| Dimension | Source Report | Status | Confidence | Constraints Extracted |
|-----------|-------------|--------|-----------|---------------------|
| Informational Density | [report] | [status] | [level] | [summary] |
| Emotional Valence | [report] | [status] | [level] | [summary] |
| Evidence Framing | [report] | [status] | [level] | [summary] |
| Syntactic Stability | [report] | [status] | [level] | [summary] |
| Complexity Targets | [report] | [status] | [level] | [summary] |
| Rhetorical Structure | [report] | [status] | [level] | [summary] |
| Pragmatic Distribution | [report] | [status] | [level] | [summary] |
| Register Rules | [report] | [status] | [level] | [summary] |
| Community Convergence | [report] | [status] | [level] | [summary] |

**Specification completeness:** [X/9 dimensions] -- [full/partial/minimal] prompt

## Few-Shot Example Selection

| # | Source | Topic | Word Count | Why Selected |
|---|--------|-------|-----------|-------------|
| 1 | [reference] | [topic] | [N] | [typicality/diversity rationale] |
| ... | | | | |

## Assembled Prompt

[The complete subagent instruction prompt, formatted and ready for use]

## Prompt Statistics

- **Total tokens:** [estimated count]
- **Sections included:** [N of 12]
- **Sections omitted:** [list with reasons]
- **Constraints enforced (measurable):** [N]
- **Constraints aspirational (not directly measurable):** [N]

## Constraint Enforcement Assessment

| Constraint | Type | Enforceability | Verification Method |
|-----------|------|----------------|-------------------|
| FK grade [X-Y] | Measurable | Strong | Post-generation readability scoring |
| Sentence length [X-Y] | Measurable | Strong | Word count per sentence |
| Sentiment compound [X.XX] | Measurable | Moderate | VADER scoring of output |
| Function word rates | Measurable | Moderate | Frequency analysis of output |
| Argument ordering | Structural | Moderate | Manual inspection |
| Hedging frequency | Countable | Moderate-Weak | Hedge marker counting |
| Speech act ratios | Distributional | Weak | Manual classification |
| Community convergence | Ambient | Weak | LSM computation against baseline |
| [etc.] | | | |

## Test Results

### Test 1: [Topic]
- **Generated text:** [first 100 words or full text if short]
- **FK grade:** [measured] (target: [X-Y]) -- [PASS/FAIL]
- **Sentence length:** [measured] (target: [X-Y]) -- [PASS/FAIL]
- **Sentiment:** [measured] (target: [X.XX]) -- [PASS/FAIL]
- **Structure match:** [assessment] -- [PASS/FAIL]

[Repeat for each test]

### Aggregate Test Results
| Metric | Tests Passed | Tests Failed | Notes |
|--------|-------------|-------------|-------|
| [metric] | [N/total] | [N/total] | [notes] |

## Prompt Revision History

| Version | Change | Reason | Test Impact |
|---------|--------|--------|-------------|
| v1 | Initial assembly | -- | -- |
| v2 | [change] | [test failure] | [improvement] |

## Limitations and Caveats

- This prompt produces an APPROXIMATION of the target voice, not the
  voice itself. Output quality depends on the LLM's capacity to follow
  layered constraints.
- Distributional constraints (speech act ratios, community convergence)
  are weakly enforceable through instructions alone. Few-shot examples
  are the primary mechanism for these.
- The prompt was tested with [model name/size]. Different models may
  require constraint adjustments.
- Style drift increases with output length. For outputs exceeding
  [X] words, consider segmented generation with re-anchoring.
- Constraints derived from low-confidence upstream analyses
  ([list them]) should be treated as provisional.
- The prompt does not capture: [list known gaps -- e.g., humor timing,
  cultural references, domain-specific jargon not in examples].

## References

- [Upstream analysis reports referenced]
- Brim Labs (2025). LLM Personas: How System Prompts Influence Style,
  Tone, and Intent. https://brimlabs.ai/blog/llm-personas-how-system-prompts-influence-style-tone-and-intent/
- LearnPrompting (2025). Role Prompting: Guide LLMs with Persona-Based
  Tasks. https://learnprompting.org/docs/advanced/zero_shot/role_prompting
- PromptHub (2025). Role-Prompting: Does Adding Personas Really Make a
  Difference? https://www.prompthub.us/blog/role-prompting-does-adding-personas-to-your-prompts-really-make-a-difference
- Palantir (2025). Best Practices for LLM Prompt Engineering.
  https://www.palantir.com/docs/foundry/aip/best-practices-prompt-engineering
- Mosteller, F. & Wallace, D.L. (1963). Inference in an authorship
  problem. JASA, 58(302), 275-309.
- Kestemont, M. (2014). Function Words in Authorship Attribution.
  Literary and Linguistic Computing, 29(1), 107-114.
- Stachl, C., et al. (2022). The kernel of truth in text-based
  personality assessment. Psychological Bulletin.
- arXiv:2509.24930 -- How Well Do LLMs Imitate Human Writing Style?

Good Patterns

  • Few-shot examples are the strongest constraint -- always include them when the source corpus is available; one well-chosen example teaches more than 200 words of instruction
  • Order prompt sections by enforcement strength -- put strongly enforceable constraints (readability, sentence length) before weakly enforceable ones (speech act ratios, community convergence) so they are processed first
  • Use numerical targets with tolerance ranges -- "FK grade 11-13" is enforceable; "write at an educated level" is not
  • Include self-verification in the prompt itself -- LLMs can check their own output against explicit criteria when instructed to do so
  • Document which constraints are measurable vs. aspirational -- this sets honest expectations and enables iterative refinement
  • Test against the source material -- generate test outputs and compare stylometric features against the fingerprint; adjust constraints that produce drift
  • Keep the total prompt under 4,000 tokens -- beyond this, constraint interference increases and the LLM begins to ignore lower-priority sections
  • Use anti-markers (patterns the voice does NOT use) alongside positive markers -- absence constraints are often more distinctive than presence constraints
  • Version the prompt -- track changes and their effects on test results; prompts require iterative refinement like code

Anti-Patterns

Anti-Pattern Why It Fails Instead
"Write like the user" Vague instruction; LLM defaults to generic helpful voice. No measurable criteria. Decompose into specific constraints: readability targets, function-word rates, emotional register, structural patterns
Overloading a single prompt with 20+ constraints LLMs exhibit constraint interference -- later constraints override or conflict with earlier ones. Quality degrades. Prioritize by enforcement strength. Keep total under ~4,000 tokens. Cut weakly enforceable constraints first.
No few-shot examples Instructions alone cannot fully specify a voice. The LLM has no concrete reference point. Include 3-5 representative corpus samples spanning different topics and contexts.
No verification loop First-draft prompts rarely match the target style. Without testing, drift goes undetected. Generate test outputs, measure against the fingerprint, revise failing sections.
Assuming one prompt works for all contexts Context-dependent voices need conditional rules; a single prompt produces a flat, averaged style. Check register variation analysis. If context-dependent, include IF/THEN register rules.
Including unmeasurable constraints "Maintain the user's sense of humor" or "capture their worldview" cannot be verified and create false confidence. Mark such constraints as aspirational. Invest prompt space in measurable alternatives.
Fabricating constraints from missing data If no LIWC analysis was performed, inventing an informational density target produces a fabricated voice, not the target voice. Leave the section as "DEFAULT: use natural style" and document the gap.
Skipping the token budget check A 6,000-token prompt may exceed system-prompt limits or cause the LLM to ignore late sections. Check total tokens. Compress or cut sections to stay within budget.
Treating the prompt as final Prompt engineering is iterative. A prompt that passes initial tests may drift on new topics or longer outputs. Version the prompt. Re-test periodically. Expect 2-3 revision cycles.

Boundaries

This skill SHOULD produce:

  • A complete, structured LLM subagent prompt organized into sections by enforcement strength
  • Explicit numerical constraints for measurable dimensions (readability, sentence length, function-word rates, sentiment)
  • Few-shot examples from the source corpus (when available)
  • Self-verification instructions built into the prompt
  • A clear distinction between enforced (measurable) and aspirational (non-measurable) constraints
  • Test results comparing generated output against the source corpus style fingerprint
  • Documentation of which specification dimensions are covered and which are missing
  • A written report at docs/analysis/25-subagent-instruction.md
  • A versioned prompt with revision history

This skill should NOT:

  • Claim the output IS the user's voice (it is always an approximation based on analyzed patterns)
  • Include constraints that cannot be verified or tested against any criterion
  • Produce a prompt exceeding ~4,000 tokens without explicit justification and constraint triage
  • Skip testing against source material
  • Fabricate constraints from dimensions that were not analyzed
  • Include ethical or value judgments about the target voice (the prompt replicates patterns, it does not evaluate them)
  • Assume the prompt will work without iterative refinement
  • Treat distributional constraints (speech act ratios, community convergence) as strongly enforceable without few-shot support
  • Ignore the enforcement gradient (treating all constraints as equally followable)
  • Produce platform-specific prompts when the voice specification is platform-agnostic

Insufficient Data Handling

digraph insufficient_data {
    rankdir=TB;
    node [shape=diamond, style=filled, fillcolor="#f0f0f0"];

    dims [label="Dimensions\ncompleted?", shape=diamond];
    full [label="7-9 dimensions", shape=box, fillcolor="#ccffcc"];
    partial [label="4-6 dimensions", shape=box, fillcolor="#ffffcc"];
    minimal [label="1-3 dimensions", shape=box, fillcolor="#fff3cc"];
    none [label="0 dimensions", shape=box, fillcolor="#ffcccc"];

    dims -> full [label="7-9"];
    dims -> partial [label="4-6"];
    dims -> minimal [label="1-3"];
    dims -> none [label="0"];

    full_action [label="Full prompt generation.\nAll sections populated.\nTest against source.", shape=box, fillcolor="#ccffcc"];
    partial_action [label="Generate prompt with gaps.\nMark missing sections as\nDEFAULT. Document coverage.", shape=box, fillcolor="#ffffcc"];
    minimal_action [label="Targeted constraints only.\nDo NOT frame as full\nvoice replication prompt.", shape=box, fillcolor="#fff3cc"];
    none_action [label="STOP. Cannot generate.\nRun upstream analyses.", shape=box, fillcolor="#ffcccc"];

    full -> full_action;
    partial -> partial_action;
    minimal -> minimal_action;
    none -> none_action;
}
Condition Action
0 dimensions completed STOP. Cannot produce a subagent instruction. Run upstream analyses first.
1-3 dimensions completed Produce targeted constraints only (e.g., "write at FK grade 12 with sentences averaging 18 words"). Do NOT frame this as a voice replication prompt -- it is a partial style constraint. Document what is missing.
4-6 dimensions completed Produce a structured prompt with explicit gaps. Each missing section gets: "DEFAULT: No specification available for [dimension]. Use natural style for this aspect." List all defaults in the report.
7-9 dimensions completed Full prompt generation. All sections populated. Test against source material.
Few-shot examples unavailable Omit the examples section. Add to role frame: "No reference examples available. Follow constraints below as primary guidance. Err toward conservative interpretation of ranges." Note that enforcement quality will be lower for structural and distributional constraints.
Low confidence on key dimensions Include the constraint but mark it: "[LOW CONFIDENCE -- derived from limited data. May not accurately reflect target style.]" The consumer of the prompt can decide whether to apply it.
Style specification is too generic If the upstream analyses show a style indistinguishable from population baselines on most dimensions (e.g., FK grade 10-12, average sentiment, no distinctive function-word patterns), the voice may not be sufficiently distinctive to replicate. Report this finding: "The analyzed style does not deviate meaningfully from baseline norms on [N] dimensions. A generic voice prompt may be as effective as a specification-driven prompt for this target."
Conflicting constraints across dimensions When two upstream analyses produce contradictory directives (e.g., high Openness suggests complex vocabulary, but readability measures show low FK grade), document the conflict and resolve by: (1) trusting the more direct measurement (readability over personality inference), (2) noting the conflict in the report, (3) testing both versions.

Common Mistakes

Mistake Fix
Writing "match the user's tone" without specifying what that means Decompose tone into measurable components: sentiment compound, emotional register (archetype), hedging frequency, exclamation rate
Including 10+ few-shot examples Diminishing returns after 5 examples. 3-5 well-selected examples are optimal. Use the saved token budget for better constraints.
Setting constraints at point values instead of ranges "FK grade 12.3" is unfollowable. "FK grade 11-13" accommodates natural variation. Always use ranges.
Not testing the prompt before deploying Generate 3-5 test outputs and measure against the fingerprint. First-draft prompts almost always need revision.
Treating all constraints as equally important Use the enforcement gradient. Readability and examples are strongly enforceable. Speech act ratios are weakly enforceable. Prioritize accordingly.
Including constraints from analyses that were not performed If no LIWC analysis was done, do not guess at informational density. Mark the section as DEFAULT.
Making the prompt too long (5,000+ tokens) Constraint interference increases with length. Compress, merge, or cut lower-priority sections. Target < 4,000 tokens.
Forgetting anti-markers Specifying what the voice DOES is insufficient. Specifying what it does NOT do (never uses exclamation marks, avoids emoji, does not enumerate with numbered lists) is equally distinctive.
No self-verification section Without verification instructions, the LLM has no mechanism to catch its own drift. Include 4-6 checkpoints.
Treating the prompt as permanent Voices evolve. Prompt constraints should be re-derived when new corpus data is available or when test results degrade. Version and date the prompt.

References