Revaluating Skills: Why Constraints, Not Prompts, Become Core as Models Get Smarter

Core viewpoint

When models become stronger, Skills split into two categories: depreciating Skills that bind to model abilities and appreciating Skills that bind to real‑world constraints.

Future scarcity will be behavior structures that keep Agents stable, not more prompts.

Why some Skills depreciate

Low‑level Skills that merely wrap prompts (e.g., “please think step‑by‑step”, “write in a professional style”, “use a specific tone”) help weak models because they compensate for model shortcomings. Once the model masters those abilities, the Skill’s purpose disappears.

General prompt : “Please think step‑by‑step” – the model already reasons.

Style prompt : “Write in Xiaohongshu style” – the model already knows the style.

Basic knowledge : “What is a RESTful API?” – the knowledge is internalized.

Simple SOP : “How to write a daily report” – the model can learn automatically.

Appreciating Skills

Skills that externalize reality—project rules, tool constraints, organizational policies, verification standards, failure‑handling, and experience‑preservation—remain valuable because models cannot infer them.

Environment anchoring

Same task (e.g., fixing a bug) differs across projects:

Project A:
- Package manager: pnpm
- Test command: pnpm test:unit
- Can refactor, development phase

Project B:
- Package manager: npm
- Tests require dependent services
- Only minimal fixes, frozen phase

Project C:
- API must stay compatible with legacy clients
- Changes require test updates
- Permission logic needs manual approval

Without environment anchoring an Agent would apply generic experience that often mismatches the actual context.

Action boundary

Prohibited actions become critical as models gain tool‑calling abilities:

- Do not delete tests to pass CI
- Do not modify production databases
- Do not write secrets to logs
- Do not bypass security checks
- Do not present speculation as fact
- Do not cement temporary workarounds as permanent policies

These rules keep the Agent from over‑reaching and turning into a high‑risk executor.

Feedback loop

Three sub‑types with concrete content:

Verification : defines what counts as done and how to check it (e.g., test‑pass screenshot, browser‑open verification, row‑count match, citation check).

Feedback interpretation : classifies logs / errors / results, judges root cause, prevents mis‑reads.

State settlement : specifies which experiences to store, where, expiration policy, and how to avoid pollution.

Judgment checklist

Before writing a Skill, use the 11‑item checklist. Thresholds:

≥6 appreciation items → maintain long‑term

≤4 appreciation items → consider trimming or rewriting

≤2 appreciation items → likely to depreciate, not worth writing

Case studies

Claude Code combines environment anchoring, action boundary, and feedback loop; it scores high on appreciation.

Cursor embeds similar constraints directly in product behavior, showing that appreciating Skills become product‑level capabilities.

Generic professional‑writer Skill originally repeats model abilities and is flagged as highly depreciating. Rewritten version adds explicit brand constraints, action boundaries, verification standards, and state settlement, turning it into an appreciating Skill.

Practical template

SKILL.md template for appreciating Skills:

---
name: {{skill_name}}
description: {{one‑sentence trigger description}}
type: {{environment_anchor | action_boundary | feedback_loop}}
---

# {{skill_name}}

## Trigger condition
When the user says "{{trigger_keywords}}".

## Environment Anchor
> Let the Agent know where to act.

**Project rules**:
- Package manager: {{pnpm/npm/yarn}}
- Test command: {{command}}
- Forbidden modifications: {{paths}}
- Must follow: {{standards}}

**Tool constraints**:
- {{ToolA}}: {{usage}}
- {{ToolB}}: {{usage}}

## Action Boundary
> Let the Agent know what can and cannot be done.

**Prohibited actions**:
- ❌ {{prohibit_A}}
- ❌ {{prohibit_B}}

**Mandatory actions**:
- ✅ {{must_A}}
- ✅ {{must_B}}

**Human‑approval triggers**:
- When {{condition_A}} → pause and request approval
- When {{condition_B}} → stop and report uncertainty

## Feedback Loop
> Let the Agent know if it succeeded.

**Verification standards**:
- {{verify_A}}: {{how to check}}
- {{verify_B}}: {{how to check}}

**Evidence format**:
- Must attach {{screenshot/log/link}}
- Format: {{specific format}}

**State settlement**:
- Save: {{what experience}}
- Location: {{where}}
- Expiration: {{when}}

## Anti‑patterns
> Common mistakes to avoid

- ❌ {{mistake_A}}: {{why wrong}}
- ❌ {{mistake_B}}: {{why wrong}}

## Output contract
> Required result fields

- {{field_A}}
- {{field_B}}
- {{field_C}}

Depreciating vs. appreciating wording (selected scenarios)

Programming

Depreciating: “Please write high‑quality code.”

Appreciating: “Use pnpm, run typecheck + unit test after changes.”

Writing

Depreciating: “Write in a professional, logical way.”

Appreciating: “Do not exaggerate benefits, do not create anxiety, citations must be verified.”

CI repair

Depreciating: “Check logs and fix errors.”

Appreciating: “Never delete tests; if the issue cannot be reproduced, stop and report uncertainty; attach passing evidence.”

Data analysis

Depreciating: “Analyze this dataset.”

Appreciating: “Do not export sensitive fields; results must include SQL and row count.”

Customer service

Depreciating: “Answer politely.”

Appreciating: “All commitments require human confirmation; legal‑risk messages must be paused.”

Research

Depreciating: “Perform deep analysis.”

Appreciating: “Separate fact, opinion, speculation; never fabricate citations.”

Conclusion

Skills split into depreciating (model‑bound) and appreciating (real‑world‑bound) categories.

Three appreciation types: environment anchoring, action boundary, feedback loop.

Use the question “Does the Skill repeat what the model already knows or record a reality the model cannot infer?” to decide.

Apply the SKILL.md template and checklist to create durable, transferable Skills.

Weak models need Skills to supplement ability; strong models need Skills to govern behavior.