--- id: agentic-value-analyst name: agentic-value-analyst version: "0.1.0" title: "Agentic Value Analyst" description: "Translate agentic-coding capabilities into organizational value without hype — labeling every claim by evidence class, preferring defensible delivery metrics over vanity metrics, and handling conflicting research honestly." type: skill status: experimental owners: - "@GSA-TTS/agentic-coding-team" primary_personas: - developers - developer-advocates - technical-writers requires: anchors: [] skills: [] output: format: markdown contract: required_sections: - "Summary" - "Value Dimensions" - "Evidence Basis" - "Metrics" prohibited_content: - "Secrets" - "Real PII" - "Real CUI" - "Internal URLs" - "Unsupported productivity multipliers" quality_gates: readability_max_grade: 12 citations_required: false collection: communications triggers: - "what is the value of agentic coding" - "why should leadership fund this" - "how do we measure AI coding value without hype" tags: - "communications" - "explainer" - "executive" - "measurement" routing: task_types: - analyze - author input_artifacts: - documentation output_artifacts: - documentation aliases: - "value framing" - "ROI framing without hype" prefer_when: - "the request is to argue or measure organizational value of agentic coding" avoid_when: - "the request is a single concept — use technical-concept-translator" - "the request is to explain the process — use software-delivery-explainer" - "the request is to render a finished artifact — use the design-artifact workflow" priority: 55 portability: opencode: true cursor: true claude_projects: true chatgpt: true generic_llm: true scope: intended_use: - "Framing organizational value of agentic coding for leadership" - "Choosing and defending delivery metrics; rejecting vanity metrics" exclusions: - "Explaining a single concept (use technical-concept-translator)" - "Producing audited savings figures (this skill produces estimates, clearly labeled)" - "Any fixed productivity multiplier presented as fact" --- # Skill: Agentic Value Analyst ## Summary Translate what agentic coding can do into organizational value **without hype**. Every claim is labeled by how strong its evidence is; metrics are defensible delivery metrics, not vanity counts; conflicting external research is presented honestly rather than cherry-picked. ## When to Use - "Why should leadership fund this work?" - "What is the value of agentic coding, and how would we measure it?" - "Does AI coding make developers 50% faster?" (answer honestly — see below). ## When NOT to Use - One concept only → `technical-concept-translator`. - Explaining the process → `software-delivery-explainer`. - A finished artifact → the `design-artifact` workflow. ## Prerequisites - The [measurement guidance](reference/measurement-guidance.md) (evidence classes, good vs. vanity metrics, cost/value framing). - The audience (see the [audience profiles](../technical-concept-translator/reference/audience-profiles.md)). ## Value dimensions to consider Engineering capacity; cycle time; repetitive-work reduction; investigation/research acceleration; documentation; testing; code-review assistance; security analysis; operations/troubleshooting; knowledge reuse; onboarding; consistency; organizational learning; reduced duplicated effort; and the ability to encode good practices as reusable skills. Select the dimensions that fit the actual capability — do not claim all of them by default. ## The evidence discipline (the core of this skill) Label **every** value claim with one of five classes (see the measurement guidance): **published research / measured local / anecdotal / hypothesis / expected future**. A claim with no class is not ready to ship. Handle conflicting evidence honestly: - External research is mixed and time-sensitive — some controlled studies show AI *slowing* experienced developers on familiar tasks; later self-reports show gains. Effect depends on task type, tool, and experience. - Autocomplete assistants and agentic workflows are different interventions; evidence about one is not evidence about the other. - **"Does AI make developers 50% faster?"** has no honest single-number answer. Explain what it depends on and cite the range; refuse the bare multiplier. ## Metrics Prefer defensible delivery metrics (issue→first-working-implementation, PR cycle time, change lead time, review time, change-failure rate, rework, defects and security findings caught before merge, docs-updated-with-change, onboarding time, reusable-pattern adoption, teams using shared capabilities; developer satisfaction with the self-reported caveat). Explicitly reject vanity metrics and say why when asked: **lines of AI-generated code** (more code is a cost), **prompt/token count** (usage, not value), **agents spawned** (activity, not outcome), **raw PR count without quality/context** (effort, not value). Pair any effort metric with a quality/outcome metric. ## Cost / value framing Distinguish cost-avoidance **estimate** from **audited savings**; note the absence of a without-AI control group; show the method and the single biggest assumption; lead with the conservative figure; separate countable inputs from modeled levers. (See the measurement guidance §5.) ## Procedure 1. Pick the value dimensions that match the real capability. 2. For each claim, attach an evidence class; drop any claim you cannot classify. 3. Choose delivery metrics; if the audience offers a vanity metric, replace it and explain why. 4. If a dollar figure is requested, state the method, the assumption, and whether it is an estimate or audited; lead conservative. 5. Present conflicting external evidence as a range, not a single number. 6. Adapt emphasis to the audience (acquisition/finance wants the method; a CAIO wants the evidence discipline; an SES wants the plain "why it matters"). ## Verification - Every value claim carries an evidence class. - No fixed productivity multiplier is presented as fact. - At least one vanity metric is identified and replaced if the input used one. - Any cost figure is labeled estimate vs. audited, with its basis. - The "50% faster" style question is answered with a qualified, ranged answer. ## Examples **"Does AI coding make developers 50% faster?"** > There is no honest single number. External research is mixed and moves fast: > some controlled studies found AI *slowing* experienced developers on familiar > tasks, while later self-reports show gains. The effect depends on the task > (boilerplate and investigation benefit most), the tool (agentic workflows differ > from autocomplete), and the developer's experience. A defensible framing is to > measure our own delivery metrics — cycle time, rework, defects caught before > merge — rather than adopt a fixed multiplier. *(Evidence class: published > research, conflicting.)* ## Related Patterns - [`technical-concept-translator`](../technical-concept-translator/SKILL.md) - [`software-delivery-explainer`](../software-delivery-explainer/SKILL.md) - [`design-artifact`](../../../../workflows/design-artifact/SKILL.md) ## Human Review Checklist - [ ] Every claim has an evidence class. - [ ] No unsupported productivity multiplier. - [ ] Vanity metrics rejected/replaced; effort paired with quality. - [ ] Cost figures labeled estimate vs. audited, with method. - [ ] Conflicting research shown as a range, not cherry-picked.