boaz-hwang/skilloom

# Consolidate

Reduce duplicated skills without losing useful behavior. Compare intended results
and quality requirements, not just names or similar prose. Discuss in the user's language.

## 1. Compare the supplied skills

Inspect the skills or skill directory the user identifies. Use names and descriptions
to find candidates, then read their instructions, linked resources, evaluation
criteria, and available accepted examples. Identify differences in outcomes, inputs,
templates, constraints, permissions, and required tools or environments.

Group skills only when one coherent skill can serve their purposes more clearly.
Represent incidental input or template differences as inputs/resources when suitable.

# Distill

Preserve the minimum needed to reproduce a result the user considers good.
Assume a capable agent: specify outcomes, essential constraints, and evidence
of quality; leave implementation choices open. Discuss in the user's language.

## 1. Confirm the destination

Read the available session and accepted deliverable. Infer the quality bar from
requests, corrections, and final acceptance. Ask for missing acceptance or essential
evidence specifically; do not invent history or search unrelated sessions.

Propose three skill names, recommend the strongest, and describe the intended
deliverable in one line. Ask whether the name and outcome match the user's intent;

# Evolve

Improve observed behavior with the smallest sufficient change. Review deletion
first; adopt changes for their demonstrated effect, not their brevity. Keep the
skill's purpose and essential quality requirements. Discuss in the user's language.

## 1. Establish the evidence

Identify one target skill and relevant records from this session or sources the
user provides. Inspect its instructions, linked resources, inputs, actual results,
and user feedback. Confirm that the skill was used and identify the version used;
distinguish differences in model, tools, and input conditions across runs.
Ask only for missing evidence that materially affects the decision. One run can
support a hypothesis, but does not establish a repeatable improvement. Do not

# Refine

Make one skill clear and sufficient for a capable agent. Review deletion first;
preserve what quality depends on. Whole-skill restructuring is in scope, combining
multiple skills is not. Discuss in the user's language.

## 1. Establish the intended result

Read the target's instructions and linked templates, references, and scripts.
Infer its intended output, essential constraints, and quality criteria from the
skill, available accepted examples, and user feedback. Execution records can help
but are not required. Do not confuse today's inputs with reusable requirements.

When intent or conflicting requirements would change the rewrite, propose an