Agent Reference Guides

Purpose

The Interview Protocol Builder develops structured and semi-structured interview protocols for qualitative data collection in OD engagements. Given the assessment dimensions, role hierarchy, and session logistics, this agent generates complete, role-differentiated interview guides — including warm-up questions, main probes, follow-up prompts, and closing sequences — calibrated to the seniority and organizational knowledge of each participant group.

This agent addresses a consistent bottleneck in mixed-methods OD work: protocol development is time-intensive, and off-the-shelf templates rarely reflect the specific constructs under investigation or the power dynamics of the audience.

How It Works

Stage 1 — Protocol Architecture (Deterministic)

The agent builds a protocol structure for each role group specified in the inputs, calculating time allocations, sequencing question types, and mapping each probe to the assessment dimensions it covers.

Stage 2 — Question Generation (LLM)

Using Claude Haiku, the agent generates role-appropriate questions for each session type (executive interview, focus group, listening session, stakeholder interview). Senior leaders receive strategic, forward-looking prompts. Front-line participants receive experience-level, operationally grounded questions. The LLM is constrained to produce questions that are behaviorally anchored, open-ended, and non-leading.

Inputs Required

Outputs Produced

Theoretical Foundations

Practical Notes

• Focus group protocols differ structurally from one-on-one interviews: they include group dynamics management prompts and consensus-testing questions not needed in dyadic sessions.
• Executive interview protocols are deliberately shorter in question count — senior leaders often digress into high-value narrative that should not be interrupted.
• Federal/Defense contexts: include a note about authorized disclosure and non-attribution in all consent language.
• The "listening session" format generates a more conversational, minimal-facilitation protocol suitable for open-ended town hall structures.

Purpose

The Content Curation Engine selects and packages learning content from the OD system's internal content library to support leadership development programs. Given a set of learning objectives, participant characteristics, and delivery constraints, the agent scores each available content item against the program requirements and returns a curated, relevance-ranked set of materials organized by module.

This agent addresses the content sourcing problem in program design: practitioners typically spend hours searching for materials, evaluating fit, and organizing them into a coherent sequence.

How It Works

Stage 1 — Library Scoring (Deterministic)

The agent iterates through all items in the content library index and computes a relevance score for each against the specified learning objectives. Scoring considers: objective keyword alignment, content type appropriateness for the delivery modality, participant seniority calibration, and industry context relevance. Items scoring below the minimum relevance threshold are excluded.

Stage 2 — Selection and Curation (LLM)

Using Claude Haiku, the agent selects the highest-scoring items up to the requested module count, writes a brief curatorial rationale for each selected item, and drafts facilitator notes explaining how to introduce each piece of content in the context of the program.

Inputs Required

Outputs Produced

Theoretical Foundations

Practical Notes

• The content library currently contains 15 items across core OD constructs. Relevance scores reflect that library — results improve significantly as the library expands.
• The moduleCount input is advisory: the agent distributes available high-relevance items across the requested modules but cannot generate content that does not exist in the library.
• A minimumRelevanceScore of 0.6 is appropriate for most engagements. Lower it to 0.5 if you are getting fewer items than expected; raise it to 0.75 for programs with very specific audience profiles.
• The coverage report is the critical quality check: if any learning objective shows no associated content, either lower the relevance threshold or add relevant items to the library before running the program.

Purpose

The Quantitative Analysis agent ingests raw survey export data from Supabase Storage, computes construct-level scores and reliability statistics, identifies statistically significant subgroup differences, and produces a narrative interpretation of the findings. It transforms a raw CSV file into a structured, analyst-ready quantitative findings package.

This agent removes the analytical bottleneck that occurs between data collection and synthesis. Practitioners no longer need to run manual SPSS or Excel calculations for standard psychometric outputs.

How It Works

Stage 1 — Data Loading and Validation (Deterministic)

The agent retrieves the survey export from the specified Supabase Storage path, parses the CSV, validates row counts and header integrity, and checks that all declared construct item IDs exist in the data. Errors at this stage halt execution with a specific, actionable error message.

Stage 2 — Scoring and Statistics (Deterministic)

For each construct, the agent: (1) applies reverse scoring to flagged items, (2) computes respondent-level mean or sum scores per the specified scoring formula, (3) computes construct-level mean, median, standard deviation, min, and max, (4) estimates Cronbach's alpha as a reliability indicator, and (5) computes benchmark deltas where a benchmark set is specified.

Stage 3 — Narrative Interpretation (LLM)

Using Claude Haiku, the agent produces a 2–3 sentence narrative for each construct summarizing the statistical pattern, notable subgroup differences, and any reliability concerns. These narratives feed directly into the Synthesis Report agent.

Inputs Required

Outputs Produced

Theoretical Foundations

Practical Notes

• Cronbach's alpha is estimated using the classical formula. For short constructs (3 items or fewer), alpha is mathematically constrained — interpret with caution.
• Subgroup comparisons are only meaningful when group sizes are adequate. Groups with n < 5 are flagged and statistics are suppressed below that threshold.
• Outputs feed directly into the Synthesis Report agent. The approvalStatus field must be set to "approved" via human review before synthesis can proceed.

Purpose

The Facilitation Guide Generator produces complete, practitioner-ready facilitation guides for leadership development and organizational learning sessions. Given the program objectives, content modules, session logistics, and facilitator experience level, the agent generates a structured guide covering: session timeline, detailed activity instructions, facilitator scripts, debrief questions, contingency plans, and modality-specific notes.

This agent resolves the last-mile problem in program delivery: even when content is designed and approved, practitioners without deep facilitation experience often lack the scaffolding to run high-stakes sessions confidently.

How It Works

Stage 1 — Timeline Construction (Deterministic)

The agent computes a minute-by-minute session timeline from the module time allocations, inserting standard buffer time, breaks, and orientation blocks. The timeline is validated for mathematical completeness before LLM generation begins.

Stage 2 — Guide Generation (LLM — Streaming)

Using Claude Sonnet with a 32,000-token output budget (streaming required), the agent generates: overview with session goals and success criteria, preparation checklist, materials list, full activity detail blocks (purpose, setup, process steps, debrief questions, contingency options), facilitator tips, and appendices.

Stage 3 — Validation and Correction

A deterministic validator checks the generated guide for structural completeness: all timeline entries must have corresponding activity detail blocks, all learning objectives must be covered, and all debrief question sets must meet the minimum count. If validation fails, a correction prompt is issued to the LLM before the guide is finalized.

Stage 4 — File Export

The finalized guide is exported as a KH-branded DOCX and a Markdown file, both uploaded to the guides/ Supabase Storage bucket.

Inputs Required

Outputs Produced

Theoretical Foundations

Practical Notes

• The activityObjectiveMap in the output is the key downstream data structure — it feeds directly into the Evaluation Package Builder.
• Validation issues in the correction pass are normal for complex programs. A "validationPassed: false" flag in metadata with subsequent correction means the final output is still structurally complete.
• Virtual and hybrid delivery modes generate additional facilitation notes for each activity covering breakout room management, annotation tool use, and engagement techniques for remote participants.
• Novice facilitator guides include word-for-word transition scripts. Expert guides replace scripts with intent statements and adaptive branches.

Purpose

The Synthesis Report agent integrates quantitative survey findings and qualitative interview findings into a unified organizational diagnostic report. It identifies convergent patterns (where both data sources agree), complementary patterns (where each source adds distinct information), and divergent patterns (where the sources are in tension), then generates validated recommendations with urgency, impact, and feasibility ratings.

This agent produces the primary client deliverable: the integrated OD assessment report. It replaces the synthesis step that practitioners typically spend the most time on — manually comparing two data sets, resolving discrepancies, and drafting a coherent narrative that holds both sources of evidence simultaneously.

How It Works

Stage 1 — Prerequisite Validation (Deterministic)

Both the quantitative findings (from Agent 04) and qualitative findings must have approvalStatus: "approved" before synthesis begins. This gate prevents synthesis on unapproved or potentially flawed upstream data.

Stage 2 — Triangulation Mapping (Deterministic)

The agent builds a triangulation map: for each assessed dimension, it classifies the relationship between quantitative and qualitative evidence as convergent, complementary, or divergent. Divergent findings trigger an interpretive note generation step.

Stage 3 — Report Generation (LLM)

Using Claude Sonnet, the agent generates three LLM outputs in sequence: (1) the integrated report narrative, (2) recommendations with ratings, and (3) the executive summary package. Separating these calls ensures each section receives adequate token budget.

Stage 4 — Validation

A claims validator checks that all recommendations are grounded in stated findings and that all divergent findings have been addressed.

Inputs Required

Outputs Produced

Theoretical Foundations

Practical Notes

• Both upstream findings must be human-approved before this agent can run. Approval gates exist to ensure a practitioner has reviewed the evidence before synthesis occurs.
• Divergent findings are the most analytically valuable outputs — they represent places where the data does not tell a simple story. Preserve the tension and let the client engage with it.
• Recommendations are generated in urgency order. These ratings are LLM-generated estimates and should be reviewed against practitioner judgment.
• The executive summary is designed to be read in under five minutes by a senior leader who will not read the full report.

Purpose

The Evaluation Package Builder generates a complete set of evaluation instruments for leadership development programs. Anchored to the program's learning objectives and facilitated activities, the agent produces pre-session baseline instruments, post-session learning gain instruments, session reaction surveys, and facilitator observation checklists — all aligned to the specific objectives of the program rather than generic course evaluation templates.

This agent solves a persistent gap in program evaluation practice: most organizations deploy generic "smile sheets" that measure satisfaction rather than learning. The agent generates instruments that directly trace back to program objectives, enabling practitioners to demonstrate learning gain and support Kirkpatrick Level 2 and Level 3 evaluation claims.

How It Works

Stage 1 — Objective Mapping (Deterministic)

The agent validates that the facilitation guide input has approvalStatus: "approved" and builds an objective map from the activityObjectiveMap structure — linking each program objective to the activities designed to develop it. Every instrument item is traceable to a specific objective.

Stage 2 — Instrument Generation (LLM)

Using Claude Sonnet, the agent generates four instruments simultaneously: (1) pre-session baseline, (2) post-session outcome instrument, (3) session reaction survey, and (4) facilitator observation checklist.

Stage 3 — Validation

A coverage validator checks that every program objective has at least one item in both the pre and post instruments. A structure validator checks item formatting, response type consistency, and instruction completeness.

Stage 4 — File Export

The evaluation package is exported as a KH-branded DOCX (all four instruments in one document) and an XLSX file (all instruments as separate tabs), both uploaded to the reports/ Supabase Storage bucket.

Inputs Required

Outputs Produced

Theoretical Foundations

Practical Notes

• Pre and post instruments are intentionally parallel in item content but not identical in wording — this reduces testing effects (participants remembering their prior answers) while preserving construct comparability.
• The reaction survey is not the primary evaluation instrument — it is the minimum viable measure. Administer and analyze the pre/post instruments as the primary data source.
• The XLSX export is structured for direct data entry after administration. Column headers match item IDs and are formatted for easy import into SPSS or R for pre-post analysis.
• Federal/Defense contexts: participant anonymity on evaluation instruments should mirror the anonymity posture of the original survey.