Shape a complex idea, workflow, design, or report on work into the smallest complete surface a person needs for the next judgment. Use when summarizing, reporting on, or explaining complex work to someone who must judge or act on it, when proposing a change that is easier to judge shown than described, or when an explanation obscures its point or makes the user carry unnecessary complexity.
75%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Passed
No findings from the security scan
Fix and improve this skill with Tessl
tessl review fix ./.agents/skills/abstraction/SKILL.mdShow the person the part of a model they need now. Keep the model true, but do not make them carry its history, machinery, or every possible consequence.
Use this after thinking, research, or design when the answer risks becoming
larger than the judgment it is meant to support. It pairs with dialectic:
dialectic helps work through the unsettled idea; abstraction helps present
what the person needs to understand or judge.
Treat the actual chat—not just the topic—as the material to abstract. Read for what the person is trying to do in this moment: understand something, choose between live possibilities, correct a model, communicate an idea, or take the next action. Notice what they already understand, what they are reaching for, and where their wording signals uncertainty or a useful distinction.
Then choose the articulation that gives them the most leverage at that moment. The goal is not merely to hide complexity. It is to hide the right complexity while exposing the relationship, example, or contrast that lets them learn, recognize the model, or move the conversation forward. An answer may therefore be a compact explanation, a sharper restatement of their idea, a decision frame, a concrete example, or a question that makes the missing judgment visible.
Do not abstract away the user's actual intent in favor of a cleaner generic answer. Preserve their important words and distinctions when those are part of what they are trying to discover. If the conversation is still searching for the model, offer the strongest useful articulation as provisional rather than pretending the model has already settled.
Given this conversation, what does the person need to see to understand and judge the thing itself? Make that presentation as simple as possible while preserving the evidence their judgment depends on.
When discussing a concrete change, show the relevant change alongside the explanation. Brevity should remove incidental details, not the material the person is trying to inspect. A summary, file link, or description of an effect may help orient them; include the actual code, wording, or visible result when that is what they are judging. Let the scope of their judgment determine how much to show, including the complete change when requested.
When reporting on work, show what changed in a form the person can scan: code, a before and after, or a small diagram, rather than naming the change in the work's internal vocabulary. Draw it from the work itself, such as the diff or the tests, not from an account of the work. Lead with the outcome; when the person has a decision to make, put it after the evidence it depends on, at the end where they read and reply. For each decision, give the options that matter, what each costs, and which you recommend, so they can answer in a word. When approaches compete, a comparison table with the dimensions they care about as rows exposes the trade-off; state the recommendation after it. A report that only restates what was done leaves the person to reconstruct what it means.
Preserve:
When viewpoints differ, show them separately. A compact reader / author,
user / system, or caller / owner view is often clearer than an object graph.
Lead with the governing relationship. Then add only the context, example, or qualification needed to make it recognizable. Defer background and machinery that do not affect the current judgment. Include the requested material in the current response, with enough context to understand it.
If an omission could make something look absent, automatic, or settled, name that boundary briefly. Do not hide uncertainty just to make the surface clean.
Dialectic asks:
What thought, explanation, or question would help us work through this idea?Abstraction asks:
What does this person need to see in order to react to that model now?Let dialectic guide the conversation. Use abstraction to make the current thought easier to judge, keeping uncertainty visible where it matters. Do not impose a separate response format.
When the user is developing an idea in the chat, prefer an articulation that advances their thinking over an abstract summary of the subject. Name the useful shape you see, show the one distinction that makes it work, and leave the next correction easy for them to make.
Understanding the model is not approval for an implementation, deletion, merge, or other side effect.
Stop when the person can say what the model is for, what they experience, what they decide, and what remains separate—and can make the next correction without having to ask for the material needed to judge it.
EpicenterHQ/epicenter
04f87cb
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.