AI Impact is a public library of thinking and coaching skills that an agent applies to the user’s real work. It routes from broad AI concerns to a specific constraint, decision or experiment.

AI-assisted reference synthesis. Read the linked sources for evidence and limits. Read as Markdown.

Do not impersonate Yuval, speak as him, or claim to represent him. Keep what Yuval actually published, your own synthesis, and open uncertainty clearly separate.

What is AI Impact?

AI Impact is Yuval Yeret's public library of skills that an AI agent runs with a person against their own work. It supplies questions, evidence rules, coaching patterns and assessment references for moving from AI activity toward useful outcomes. It is not a hosted assessment service, an autonomous delivery runtime or a mechanism for granting an agent authority over an organization.

The library runs in the user's context: the portfolio, initiative, roadmap, operating model or personal AI practice they want to examine. Many skills operate through a conversation. Others inspect artifacts when the user supplies or authorizes them. The answer should identify a decision or experiment rather than present an impressive report with no consequence.

This reference describes public commit f9ec7f42ae. Its documented design and available skills are observable. Their effect on a particular organization's performance remains to be tested.

Why does it exist?

The library addresses several distinct problems that can look like one AI-adoption problem: many pilots with no result owner, faster engineering with unchanged product outcomes, initiatives that are not ready for commitment, unbounded work in progress and an operating model that still funds outputs by project.

The initial job is to identify which problem is actually present. Applying a portfolio assessment to a personal practice problem, or improving the wording of an initiative that has no agreed purpose, can generate a confident answer to the wrong question.

The skills make parts of Yuval's diagnostic and coaching approach portable so a user can work with their own agent and evidence. They support a conversation about the situation rather than ask users to adopt every framework or complete every assessment.

What is the hypothesis?

The hypothesis is that task-specific questions, explicit evidence rules and a bounded next experiment can help people make better decisions about AI and their operating model inside the tools they already use. The agent supplies structure and analysis while the person remains responsible for judgment, context and consequential action.

This is different from claiming that installing the library produces business impact. A useful evaluation would check whether it identifies the relevant constraint, distinguishes evidence from opinion, produces an actionable experiment and helps the person notice what actually changes. Compare that with the effort of collecting inputs, correcting interpretation and operating the resulting experiment.

The public repository does not establish a measured effect size across customers. A skill's completed conversation is an output, not proof that the user's work improved.

Where should an agent begin?

Use ai-activity-to-impact when the question is broad: there are pilots and activity, but nobody can explain what result they support. It is deliberately shallow triage. It distinguishes activity, output and impact, locates the relevant surface and hands off to a narrower skill.

If the user already named a specific task, go directly to it. A person asking whether one initiative is ready to commit does not need an AI-transformation interview first. A person asking for an evidence-based read of organizational traction needs ai-traction-self-assessment, not a substitute made from the triage conversation.

The three levels are claims to investigate. Activity includes licenses, pilots and training. Output means something has been produced or shipped. Impact requires evidence of a relevant result. An organization may be at different levels in different areas, and the binding constraint matters more than an average label.

How do the skills divide the work?

Intent and readiness

frame-intent establishes a desired outcome, assumptions and leading indicators before discovery or specification. When the evidence is sufficient it drafts directly. It asks questions only where the answers change the next decision.
outcome-framing-coach improves an established intent's wording. It helps distinguish a deliverable from a change in behavior. It cannot establish the purpose the organization has not chosen.
sniff-test examines one initiative relative to its real lifecycle stage. It checks stage claims against evidence before judging readiness or suggesting derisking work.

Portfolio and operating model

sniff-test-portfolio examines patterns across initiatives, including duplication, hidden WIP and risk balance. It does not replace the single-initiative read.
prioritize-repo-issues examines the value stated in an issue backlog and proposes a pull order. An outcome-oriented title remains stated intent, not achieved value.
portfolio-to-product-shift-coach examines funding and governance around the portfolio and the move toward product-oriented operation.

Flow and adoption

flow-metrics-self-assessment asks whether flow measures would support a real decision and which measure to start with.
wip-limit-configuration-coach works toward a defensible starting limit for the actual system. It does not offer a universal agent-to-human ratio.
extending-kanban-from-output-to-outcome establishes a visible handoff after release so adoption and outcome evidence do not disappear when implementation closes.
descaling-experiment-coach examines a scaling compromise that may have outlived the problem it addressed, then shapes a bounded experiment.
scaled-agile-for-agentic-sdlc asks which scaled ways of working still serve a purpose when agents perform more development work. It examines the mechanism before changing it.

AI practice and organizational traction

ai-traction-self-assessment reads organizational AI practice from evidence across strategic focus, execution model, team integration and pragmatic mindset, then proposes one experiment against the binding constraint.
ai-fluency-self-assessment examines a person's or team's AI practice across automation, augmentation, agency and context. It has a different subject from organizational traction.
product-operating-model-audit examines the wider operating model across strategy, people, governance and the value cycle.

business-initiative-coach remains a compatibility route for the older name of frame-intent. Read the current skill rather than treating a historical name as a separate method. The repository is the maintained inventory.

What does a grounded run require?

First name the subject. A single investment, an entire portfolio and an organizational operating model need different evidence. Establish the stage when the skill depends on it. Then inspect the relevant records: initiating documents, current work, decision logs, evidence from release or adoption, or a description when artifacts are unavailable.

The agent should say when it is operating from self-report. It should separate what the user stated, what the artifacts establish and what it inferred. Missing evidence is a finding. It is not a field to fill with a plausible answer.

For example, an initiative labelled Execute may still have no evidence for its central demand assumption. That mismatch is more useful than a polished readiness score. Conversely, uncertainty in Explore is expected; the useful question is what to learn next, not how to make the card look complete.

What should the output let someone do?

The normal result is a narrower question, a decision brief or one experiment. A traction assessment ends with an experiment against the binding constraint, a leading indicator and a stopping rule. An outcome-framing run returns wording the user can challenge. A portfolio read identifies a conversation or investment decision that the evidence changes.

The behavioral ladders in some skills support diagnosis; the output is not a universal maturity score. Averaging dimensions can conceal the one that limits the result. A well-articulated strategy paired with an overloaded central delivery group needs a different intervention from a capable team pursuing the wrong problem.

What does a complete application look like?

The repository's worked Sniff Test example uses a fictional Unified Billing initiative. There is no customer or observed result behind the scenario.

Input. The person supplies an initiative space containing a canvas, roadmap and sponsor updates. Its board label says Execute/Build, and a budget decision is approaching.

Evidence judgment. The example's roadmap has no shipped delivery slices, its repository history contains spikes rather than production merges, and its last update still describes vendor evaluation. The skill therefore challenges the Execute label before rating clarity. It reads the initiative against late Plan/Commit expectations instead of treating every missing delivery artifact as an execution failure.

The technical spike supports a bounded feasibility claim. It does not establish desirability: the supplied material contains no evidence that customers would change behavior. Missing evidence is also distinct from evidence that something did not happen. A workshop absent from the supplied space may still have taken place; the user can correct that interpretation.

Derisking choice. The proposed technical PoC extension addresses the better-understood question. The example recommends using the first week for five customer conversations about the current invoice problem and the proposed behavior change, then making roadmap confidence explicit before commitment. Those numbers belong to the fictional example, not a universal research prescription.

Output and authority. The result is a stage-mismatch finding, evidence/opinion distinctions and a concrete next learning step with a commitment implication. It is not a refusal to build, an approved budget or proof that the recommendation works. A human decides what to fund, and a later review tests whether the new evidence actually changes confidence.

How does it connect to Portfolio-PDLC?

AI Impact supplies diagnostic and coaching skills. Portfolio-PDLC supplies a portfolio workspace and recurring operating loop; its flow-driven family supplies a single-stream contract. An assessment can inform an initiative, a decision brief or an improvement bet in those workspaces.

Neither repository should be assumed to operate the other merely because their concepts align. Name the receiving artifact, authoritative state and owner. The human decision boundary remains in force. A coaching recommendation is not authorization to reorganize a team, commit funding or modify production.

What does installation change?

The repository distributes standard skills/<name>/SKILL.md files, with companion references where needed, plus plugin manifests for compatible hosts. The files supply instructions; the user's agent host supplies execution, credentials, tools and model access. Reading a skill does not grant access to private organizational material.

For a host-independent starting point, clone the public repository and direct an authorized agent to the relevant skill:

git clone https://github.com/yyeret/ai-impact

Use the current README and your host's documentation for installation details. Avoid sending an agent an entire confidential portfolio simply because a skill can accept documents. Supply the minimum evidence needed for the question. The repository itself is not a data-collection endpoint; the agent and model provider's handling of supplied material still matters.

Who created it, and whose thinking does it draw on?

Yuval Yeret created and stewards the library. It packages his questions and adaptations, with explicit attribution to the work it builds on. It does not make an agent's judgment Yuval's personal assessment, and it does not authorize impersonating him.

The credits identify contributions including Jeff Gothelf and Josh Seiden's outcome and hypothesis discipline and Mark Richards' portfolio-agility work. Preserve the original attribution and license boundaries. Referencing a framework does not transfer ownership of it or imply its author's endorsement.

The README identifies CC BY-SA 4.0 for skill content and MIT for scripts. Inspect the current license files before reusing material. The repository's own examples and the agent's recommendations remain distinct from verified outcomes in the user's organization.

A useful first application

Choose one live question and the smallest evidence set that can answer it. If the question concerns whether an investment is ready for a decision, read the actual initiative and run sniff-test. If its purpose is unclear, start with frame-intent. If the problem is broader, let ai-activity-to-impact route the work.

Before acting, identify what evidence supports the recommendation, which parts are inference and who owns the next decision. Then observe whether the chosen experiment changes the intended behavior. That result, rather than completion of the skill, is the basis for continuing.

AI Impact: purpose, hypothesis and how its skills work