AIFC-043: Skill Evolution

Status: Draft 0.1
Standard: AI-First Community Standard
Short name: AIFC
Builds on:

• AIFC-000 Manifest of the AI-first community
• AIFC-001 Core Concepts
• AIFC-002 Community Model
• AIFC-004 Feedback and Change Proposals
• AIFC-010 Knowledge Structure
• AIFC-013 Human and AI Readable Content
• AIFC-020 Human-Managed AI
• AIFC-023 AI as Team Member
• AIFC-024 Human Capability Reserve
• AIFC-040 AI Retrospective
• AIFC-041 AI Waste Backlog
• AIFC-042 Workflow Conversion

Purpose of this document: Define Skill Evolution as the mechanism by which a community converts experience from work, AI outputs, retrospectives, reviews, waste, incidents, and good patterns into improved human skills and AI skills. Skill Evolution ensures that the community learns from work and returns its know-how to the Source of Truth.

1. Purpose of this document

This document defines Skill Evolution.

An AIFC community should not use AI only to create outputs.

It should learn from every significant output, correction, review, incident, or repeated pattern.

If AI creates a good dashboard proposal, the community should not be satisfied that one good output exists.

It should ask:

• Why was it good?
• What rules did it use?
• What layout worked?
• How did it protect human attention?
• What anti-patterns did it avoid?
• How will we write this into a human skill?
• How will we write this into an AI skill?
• How will we get a similar result faster and more reliably next time?

Skill Evolution is the mechanism by which good work becomes better community capability.

2. Core principle

The core principle of this document is:

Every good output should improve future capability.
Every repeated correction should improve the skill.

AIFC says:

Do not let learning remain trapped in individual outputs.
Convert learning into skills.

3. Definition

Skill Evolution is a managed process by which a community extracts experience from work and converts it into updated human skills, AI skills, examples, anti-patterns, checklists, templates, review rules, and Source of Truth artefacts.

Skill Evolution may arise from:

• good output,
• poor output,
• repeated review correction,
• AI Retrospective,
• AI Waste Backlog,
• Workflow Conversion,
• incident,
• human feedback,
• customer feedback,
• agent logs,
• governance review,
• or a change in values, strategy, or workflow.

Minimum requirement

Significant repeated learning from work must be assessed for whether it should lead to a skill update.

4. Human skill

Human skill is a human-readable knowledge artefact that helps a person perform, understand, review, or transfer a capability.

A human skill may include:

• purpose,
• principles,
• procedure,
• inputs,
• outputs,
• quality criteria,
• examples,
• anti-patterns,
• checklist,
• decision rules,
• fallback,
• when to use AI,
• when not to use AI,
• how to review an AI output,
• how to teach a new member.

A human skill is important for Human Capability Reserve.

Minimum requirement

Critical community capabilities must have a human-readable skill or corresponding human-readable knowledge.

5. AI skill

AI skill is a structured instruction, rule, or context that enables an AI agent to perform a specific type of work better, more safely, and more consistently.

An AI skill may include:

• agent role,
• scope,
• input rules,
• output format,
• allowed actions,
• forbidden actions,
• examples,
• anti-patterns,
• quality criteria,
• uncertainty handling,
• approval rules,
• Source of Truth write-back rules,
• AI-NDA Boundary reference,
• escalation rules.

An AI skill can be very powerful.

But it must not be the only place where critical know-how is stored.

Minimum requirement

A critical AI skill must be connected to human-readable knowledge or a human skill.

6. Human skills and AI skills must co-evolve

AIFC does not separate human skills and AI skills into two unrelated worlds.

They should evolve together.

When an AI skill improves, the community should assess:

• whether people understand what the agent does,
• whether a human review skill exists,
• whether the change weakens Human Capability Reserve,
• whether the know-how is portable,
• whether lock-in is emerging.

When a human skill improves, the community should assess:

• whether the AI skill should be updated,
• whether the agent should have better examples,
• whether the output format should change,
• whether a forbidden action should be added,
• whether the review checklist should improve.

Minimum requirement

A change to a critical AI skill must assess impact on the related human skill and Human Capability Reserve.

7. Why Skill Evolution matters

Without Skill Evolution, the community learns slowly and locally.

Learning remains:

• in one person’s head,
• in a chat,
• in a one-off AI output,
• in a reviewer’s comment,
• in a document correction,
• in informal experience,
• in agent memory,
• or in a lost retrospective note.

This creates knowledge debt.

Skill Evolution converts experience into repeatable capability.

As a result:

• future outputs improve,
• AI agents work more precisely,
• people review better,
• onboarding is faster,
• fallback is easier,
• AI dependency decreases,
• the Source of Truth strengthens.

Minimum requirement

The community must have a mechanism for creating or updating skills from repeated experience.

8. Skill Evolution sources

Skill Evolution may be triggered by different signals.

8.1 Good output signal

An output was very good and its rules are worth extracting.

8.2 Repeated correction signal

Reviewers repeatedly correct the same mistake.

8.3 AI waste signal

AI repeatedly consumes capacity because a skill is missing.

8.4 Workflow conversion signal

Repeated work was converted into a workflow and requires new skills.

8.5 Incident signal

An error or incident revealed a missing skill.

8.6 Human capability signal

People cannot perform, review, or restore the work without AI.

8.7 AI dependency signal

The team is too dependent on AI for routine or critical work.

8.8 Strategy or values signal

A change in values, strategy, or governance requires a change in skills.

Minimum requirement

Skill Evolution must accept inputs from AI Retrospective, AI Waste Backlog, and Workflow Conversion.

9. Skill Evolution lifecycle

AIFC recommends this lifecycle:

signal
-> skill update candidate
-> triage
-> draft
-> review
-> approved
-> published
-> used
-> measured
-> improved

Signal

Learning emerges.

Skill update candidate

The learning is proposed as a possible skill change.

Triage

Impact, owner, priority, and skill type are identified.

Draft

A draft human skill or AI skill is created.

Review

The skill is reviewed by the owner or community.

Approved

The skill is approved.

Published

The skill is written into the Source of Truth or skill repository.

Used

The skill is used in real work.

Measured

Its impact is monitored.

Improved

The skill is further updated based on experience.

Minimum requirement

Critical skills must have lifecycle status, owner, and review mechanism.

10. Skill update candidate

A skill update candidate is a proposal to create or change a skill.

It should describe:

• source of the signal,
• what the community learned,
• which skill is affected,
• whether it is a human skill, AI skill, or both,
• what problem it addresses,
• what output should improve,
• what examples or anti-patterns should be added,
• who owns it,
• how impact will be verified.

Minimum requirement

A significant skill update candidate must be traceable and have an owner or triage owner.

11. Extracting skill from good output

A good output is a source of know-how.

The community should ask:

• What makes this output good?
• What structure worked?
• What tone worked?
• What rules were implicitly followed?
• What decision was correct?
• What information was omitted?
• How was human attention protected?
• What examples should be stored?
• How can this pattern be repeated?

Example

A good dashboard proposal may lead to rules such as:

• use few visual elements,
• do not show unnecessary icons,
• show only the main flow,
• separate decision from detail,
• protect user attention,
• do not overload the first screen,
• explain metrics in context.

Minimum requirement

Significant good outputs in critical areas must be assessed for skill extraction.

12. Extracting skill from repeated corrections

A repeated correction is a signal that a skill is incomplete.

Example:

A reviewer keeps correcting AI outputs because they are too long.

Possible skill update:

Output style:
Prefer short, scannable sections.
Avoid dense paragraphs.
Use hierarchy and whitespace.
Do not repeat obvious context.

Example:

A reviewer keeps adding owner and status.

Possible skill update:

Every critical artefact draft must include owner, status, purpose and review cycle.

Minimum requirement

A repeated review correction must be assessed as a skill update signal.

13. Extracting skill from failure

Errors and incidents are hard but valuable sources of learning.

AI failure may reveal:

• missing forbidden action,
• unclear AI-NDA Boundary,
• weak review rule,
• missing human skill,
• missing fallback,
• poor output format,
• excessive AI Autonomy,
• weak agent scope,
• unclear data classification.

Minimum requirement

A significant AI failure or incident must be assessed for skill update.

14. Skill quality criteria

A good skill must be:

• clear,
• usable,
• concrete,
• testable,
• connected to the Source of Truth,
• connected to values,
• reasonably concise,
• equipped with examples,
• equipped with anti-patterns,
• owned,
• reviewed,
• portable.

A human skill must help a person. An AI skill must reduce ambiguity for an agent. Both should protect community capability.

Minimum requirement

Critical skills must have quality criteria or a review checklist.

15. Skill examples

Examples are important for people and AI.

A good example shows:

• what a good output looks like,
• why it is good,
• where the boundary is,
• what format is expected,
• what tone works,
• what metadata is required,
• what context is needed.

Examples should be concrete.

Minimum requirement

Critical skills with ambiguity risk must contain examples.

16. Skill anti-patterns

Anti-patterns show what to avoid.

They matter because AI and people often need to see not only the positive pattern, but also the boundary.

Example for an AI documentation skill:

Anti-pattern:
Long dense Markdown document with many headings but no clear priority, owner, status or next action.

Example for a UX dashboard skill:

Anti-pattern:
Too many badges, icons, colors and buttons on the first screen, forcing the user to decode the interface instead of understanding the message.

Minimum requirement

Critical skills must contain anti-patterns when incorrect use creates significant risk or attention debt.

17. Skill and values

Skills are not only technical instructions.

They should reflect community values.

For example, the value attention protection may appear in skills as:

• write concisely,
• structure content,
• do not overload,
• show the decision,
• separate detail,
• use summaries,
• protect the first screen.

The value resilience may appear in skills as:

• define fallback,
• keep a human-readable procedure,
• do not store know-how only in AI,
• limit dependency,
• maintain review competence.

Minimum requirement

Critical skills must align with community values.

18. Skill and Source of Truth

Skills must be written in the Source of Truth or managed skill repository.

A skill stored only in chat or agent memory is not sufficient.

The Source of Truth enables:

• review,
• versioning,
• audit,
• reuse,
• onboarding,
• export,
• vendor exit,
• Human Capability Reserve.

Minimum requirement

Critical skills must be versioned or traceable in the Source of Truth.

19. Skill and Human Cockpit Layer

The Human Cockpit Layer can make skills visible and easier to use.

It may show:

• relevant skill for the current task,
• checklist,
• examples,
• anti-patterns,
• skill maturity,
• review status,
• suggested skill update,
• AI-generated skill draft,
• approval request,
• skill usage metrics.

The Human Cockpit Layer helps skills become working aids, not just documents.

Minimum requirement

Critical skills must be human-accessible in the work context where they are used.

20. Skill and AI team members

An AI team member must use approved AI skills.

Its role, output style, allowed actions, forbidden actions, and approval rules must be connected to the Source of Truth.

If an agent is often wrong, Skill Evolution should assess:

• whether its role is unclear,
• whether its scope is too broad,
• whether examples are missing,
• whether forbidden actions are missing,
• whether the output format is poor,
• whether the agent lacks the right context,
• whether AI Autonomy is too high.

Minimum requirement

AI team members must have skills that are owned, reviewed, and updatable.

21. Skill and Human Capability Reserve

Skill Evolution is one of the main mechanisms of Human Capability Reserve.

If an AI skill improves an agent’s work but no human skill exists, AI dependency may increase.

When a human skill is created, the community strengthens people’s ability to:

• understand the work,
• review AI,
• continue without AI,
• teach new members,
• restore the workflow.

Minimum requirement

A critical AI-assisted capability must have a human skill or human-readable explanation that supports Human Capability Reserve.

22. Skill and Workflow Conversion

Workflow Conversion often creates new skills.

For example:

• a new workflow needs a human skill,
• a new validator needs an explanation,
• a new Human Cockpit action needs instructions,
• a new AI agent needs an AI skill,
• a new template needs examples.

Minimum requirement

Workflow Conversion must assess the need for human skill and AI skill updates.

23. Skill and AI Waste Backlog

AI waste often reveals a missing skill.

Examples:

• repeated poor outputs -> AI skill update,
• repeated poor prompts -> human prompting skill,
• high rejection rate -> output quality skill,
• review overload -> review checklist,
• dependency waste -> human skill and fallback.

Minimum requirement

AI Waste Backlog items must be assessed for skill evolution.

24. Skill and AI Retrospective

AI Retrospective must look for skill evolution opportunities.

Questions:

• What did we learn?
• Which skill should capture it?
• Is it a human skill, AI skill, or both?
• What example should be added?
• What anti-pattern should be added?
• What checklist should be created?
• How should agent instructions change?
• How can repeated review corrections be reduced?

Minimum requirement

AI Retrospective must have the output category skill update proposal.

25. Skill maturity

AIFC may distinguish skill maturity levels.

Level 0 - implicit

The skill exists only in people’s heads or in chat.

Level 1 - described

The skill is described for humans, but has no examples or review.

Level 2 - usable

The skill has a procedure, examples, and owner.

Level 3 - governed

The skill has lifecycle, review, versioning, and connection to the Source of Truth.

Level 4 - integrated

The skill is connected to the Human Cockpit Layer, AI agents, validation, workflow, and metrics.

Minimum requirement

Critical skills must be at least usable and should move toward governed.

26. Skill review

Skills may become stale.

Skill review should verify:

• is the skill still current?
• does it match values?
• does it match the workflow?
• does it match the AI-NDA Boundary?
• does it support Human Capability Reserve?
• does it have the right examples?
• does it have current anti-patterns?
• is it too long?
• is it too vague?
• is it used?
• does it have an owner?

Minimum requirement

Critical skills must have a review cycle.

27. Skill retirement

Some skills should be retired.

Reasons:

• the workflow changed,
• the skill was replaced by a better one,
• the AI agent was retired,
• the process is no longer performed,
• the skill is harmful,
• values changed,
• compliance changed,
• the skill creates dependency.

Skill retirement must be managed so old skills are not used as current Source of Truth.

Minimum requirement

Critical skills must support deprecated or retired status.

28. Measuring skill impact

Skill Evolution should be measurable.

Possible measures include:

• reduced rejection rate,
• fewer review corrections,
• improved output quality,
• reduced AI waste,
• reduced dependency,
• faster onboarding,
• better Source of Truth write-back,
• fewer incidents,
• better audit,
• higher reuse.

Minimum requirement

A significant skill update must have expected impact or a verification method.

29. Skill ownership

A skill must have an owner.

The owner is responsible for:

• freshness,
• review,
• examples,
• anti-patterns,
• connection to values,
• connection to workflow,
• connection to AI agents,
• impact measurement,
• retirement.

The owner does not have to be the author.

The owner is responsible for the skill being usable and true.

Minimum requirement

Critical skills must have an owner.

30. Skill portability

Skills must be portable.

Human skills must be human-readable. AI skills must be exportable or represented outside a proprietary tool. Critical skills must not be locked into a single vendor.

Skill portability supports:

• exit strategy,
• Human Capability Reserve,
• onboarding,
• vendor replacement,
• audit,
• Source of Truth continuity.

Minimum requirement

Critical skills must be exportable or stored in a portable format.

31. Skill update governance

Not every skill change must be approved by high-level governance.

AIFC recommends risk-based governance.

Low-risk skill update

For example:

• adding an example,
• clarifying wording,
• adding a common anti-pattern.

The skill owner may approve it.

Medium-risk skill update

For example:

• changing output format,
• changing review checklist,
• changing an agent role.

Requires owner review.

High-risk skill update

For example:

• changing allowed actions,
• changing forbidden actions,
• changing AI Autonomy,
• changing how restricted data is handled,
• changing Operational DNA handling.

Requires governance approval.

Minimum requirement

Skill update governance must match the impact of the change.

32. Suggested metadata

Example metadata for a skill:

skill:
  id:
  title:
  skill_type: human | ai | hybrid
  status: draft | proposed | active | deprecated | retired | archived
  owner:
  purpose:
  related_values:
  related_workflows:
  related_ai_team_members:
  related_human_roles:
  criticality: low | medium | high | critical
  source_of_truth_location:
  maturity_level: 0 | 1 | 2 | 3 | 4
  ai_dependency_risk: low | medium | high | critical
  human_capability_support: true | false
  examples:
  anti_patterns:
  review_cycle:
  last_reviewed:
  version:

Example metadata for a skill update proposal:

skill_update_proposal:
  id:
  title:
  status: observed | proposed | under_review | approved | rejected | implemented | verified
  owner:
  source:
    - good_output
    - repeated_correction
    - ai_retrospective
    - ai_waste_backlog
    - workflow_conversion
    - incident
    - human_feedback
    - customer_feedback
  affected_skill:
  update_type:
    - new_skill
    - example
    - anti_pattern
    - checklist
    - output_format
    - allowed_action
    - forbidden_action
    - review_rule
    - fallback
    - retirement
  human_skill_impact:
  ai_skill_impact:
  human_capability_impact:
  ai_dependency_impact:
  governance_level:
  expected_impact:
  verification_method:
  related_change_proposal:
  created_at:
  completed_at:

These structures are illustrative.

The final schema should be defined in the agent-actionable layer of the standard.

33. Anti-patterns

AIFC rejects the following anti-patterns.

33.1 Learning trapped in output

The community creates a good output but does not extract rules from it for the future.

33.2 Repeated correction without skill update

Reviewers keep correcting the same mistake, but the skill does not change.

33.3 AI skill without human skill

The agent knows how to do the work, but people do not understand it.

33.4 Skill trapped in vendor platform

A critical skill is stored only in a proprietary tool.

33.5 Skill without owner

A skill exists, but nobody maintains it.

33.6 Skill without examples

The skill is general, but neither people nor AI have a concrete pattern.

33.7 Skill without anti-patterns

The boundary of poor use is unclear.

33.8 Skill drift

The skill gradually diverges from values, workflow, or reality.

33.9 Overgrown skill

The skill is so long and complex that nobody uses it.

33.10 Skill update without governance

A critical skill changes without review even though it changes AI or human behavior.

34. Minimal requirements

In the area of Skill Evolution, an AIFC community must at minimum:

1. Distinguish human skills from AI skills.
2. Ensure critical capabilities have a human-readable skill or corresponding knowledge.
3. Connect critical AI skills to human-readable knowledge or a human skill.
4. Evolve human skills and AI skills together.
5. Accept inputs from AI Retrospective.
6. Accept inputs from the AI Waste Backlog.
7. Accept inputs from Workflow Conversion.
8. Assess significant good outputs for skill extraction.
9. Assess repeated review corrections as skill update signals.
10. Assess significant AI failures or incidents for skill update.
11. Ensure critical skills have quality criteria or a review checklist.
12. Ensure critical ambiguous skills contain examples.
13. Ensure critical risky skills contain anti-patterns.
14. Align critical skills with community values.
15. Store critical skills in the Source of Truth or a managed skill repository.
16. Make critical skills available in the work context where they are used.
17. Ensure AI team members have owned, reviewed, and updatable skills.
18. Ensure critical skills have a review cycle.
19. Support deprecated or retired status for critical skills.
20. Provide expected impact or a verification method for significant skill updates.
21. Assign owners to critical skills.
22. Store or export critical skills in a portable format.
23. Match skill update governance to the impact of the change.

35. Summary

Skill Evolution is the mechanism by which a community learns from work.

Without Skill Evolution, good outputs remain one-off. Corrections repeat. AI errors return. Human learning disappears. Agentic know-how is closed inside a tool. The community consumes experience but does not own it.

AIFC therefore says:

Turn good outputs into examples.
Turn repeated corrections into rules.
Turn failures into anti-patterns.
Turn experience into skills.

Skill Evolution ensures that every significant experience strengthens the community’s future capability.

Skill Evolution turns experience into reusable human and AI capability.