Skip to content
Back to skills

Aget Enhance Spec

ASecurity

Guide through the 7-phase specification enhancement lifecycle (Phases 0-6 from L622). Governs creating, updating, wiring, and validating AGET specifications.

  • 666 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added June 5, 2026
ai-agentspythongobash

Security analysis

A100/100

Scanned June 5, 2026

npx -y skills add majiayu000/claude-skill-registry --skill aget-enhance-spec --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Aget Enhance Spec?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Aget Enhance Spec
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/majiayu000-aget-enhance-spec/badge)](https://www.skillsdirectory.com/skills/majiayu000-aget-enhance-spec)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: aget-enhance-spec
description: Guide through the 7-phase specification enhancement lifecycle (Phases 0-6 from L622). Governs creating, updating, wiring, and validating AGET specifications.
version: 1.1.0
---

# /aget-enhance-spec

Guide agents through the core specification enhancement lifecycle. Provides governed workflow for spec creation, updates, cross-reference wiring, and validation.

## Purpose

Replace ad-hoc specification enhancement with a repeatable, governed 7-phase process derived from 5 independent plans across 3 agents (L622). This is the Generator layer (ADR-008 step 3), built on SOP_specification_enhancement.md (Advisory) and SKILL-041 spec (Strict).

## Input

$ARGUMENTS - Enhancement category and optional spec path.

Format: `<CATEGORY> [spec-path] [--version <target-version>]`

Categories: NEW, UPDATE, ADD-REQ, WIRE, CONSOLIDATE, DEPRECATE

Examples:
- `/aget-enhance-spec NEW specs/AGET_FOO_SPEC.md`
- `/aget-enhance-spec UPDATE aget/specs/AGET_SOP_SPEC.md --version 1.2.0`
- `/aget-enhance-spec WIRE aget/specs/SKILL_NAMING_CONVENTION_SPEC.md`
- `/aget-enhance-spec ADD-REQ aget/specs/AGET_SOP_SPEC.md`

## Phase Selection

Not all categories require all phases. Use this table to determine which phases to execute:

| Category | Required Phases |
|----------|:---------------:|
| NEW | 0, 1, 2, 3, 4, 5, 6 (all) |
| UPDATE | 0, 1, 2, 3, 4, 5, 6 (all) |
| ADD-REQ | 0, 1, 2, 3, 4, 5, 6 |
| WIRE | 0, 1, 5, 6 |
| CONSOLIDATE | 0, 1, 2, 3, 4, 5, 6 + reference SOP_specification_consolidation.md |
| DEPRECATE | 0, 1, 3, 6 |

## Execution

### Step 1: Parse Input

```
IF no arguments provided:
  Prompt: "What type of enhancement? (NEW, UPDATE, ADD-REQ, WIRE, CONSOLIDATE, DEPRECATE)"
  Prompt: "Path to spec? (leave blank for NEW)"

IF category not in valid list:
  Error: "Invalid category. Valid: NEW, UPDATE, ADD-REQ, WIRE, CONSOLIDATE, DEPRECATE"

IF category in (UPDATE, ADD-REQ, WIRE) AND no spec path:
  Error: "Spec path required for {category} enhancements"
```

Determine phase list from Phase Selection table above.

### Step 2: Phase 0 — Trigger & Diagnosis

**Objective**: Identify the specification gap and gather evidence.

Execute:
1. If spec path provided, read the spec file and extract current version, status, requirement count
2. Search for related L-docs:
   ```bash
   grep -l "<topic_keywords>" .aget/evolution/L*.md
   ```
3. Search for related specs:
   ```bash
   ls specs/ aget/specs/ 2>/dev/null | grep -i "<topic>"
   ```
4. Document findings as evidence table:
   ```
   | Evidence | Source | Impact |
   |----------|--------|--------|
   | <finding> | <L-doc or spec> | <why this matters> |
   ```

5. **Governance classification** (L624, R-ENHANCE-SPEC-031/032):
   ```
   Determine governance level:
   IF spec path is under aget/ or aget-framework/ (public):
     IF category is UPDATE or NEW:
       REQUIRE: PROJECT_PLAN before proceeding past Phase 0
       Check: does PROJECT_PLAN exist for this enhancement?
       IF not: "PROJECT_PLAN required for public framework UPDATE/NEW. Create one first."
     ELSE (ADD-REQ, WIRE):
       Enhancement_Decision_Log (produced at Step 9) is sufficient
   ELSE (private agent spec):
     Enhancement_Decision_Log is sufficient
   ```

**USER CHECKPOINT**: Present evidence table and governance level, then ask:
```
Enhancement Category: <CATEGORY>
Target: <spec path or "new spec">
Evidence: <count> sources found
Governance: <PROJECT_PLAN required | Enhancement_Decision_Log sufficient>

Confirm category and proceed? (yes/adjust category/abort)
```

**Exit Criteria**: Category confirmed. Evidence documented. Gap clearly stated. Governance level determined.

### Step 2.5: Phase 0.5 — Maturity Assessment (L682)

**Objective**: Score the target spec's current maturity level to inform enhancement scope.

Execute:
1. If spec exists (UPDATE, ADD-REQ, WIRE), assess against RUBRIC_specification_maturity_v1.0:
   ```
   Content:     C_ (0=Implicit, 1=Intent, 2=Structured, 3=Grounded, 4=Verified, 5=Wired)
   Enforcement: E_ (0=Unenforceable, 1=Advisory, 2=Detectable, 3=Preventable, 4=Structural)
   Adoption:    A_ (0=Unaware, 1=Aware, 2=Compliant, 3=Internalized, 4=Forward)
   Process:     P_ (0=Ad-hoc, 1=Documented, 2=Procedural, 3=Specified, 4=Automated)
   Governance:  G_ (0=Ungoverned, 1=Tracked, 2=Planned, 3=Governed, 4=Auditable)
   ```
2. Identify the **constraining dimension** (lowest level)
3. For NEW specs, record baseline as C0/E0/A0/P0/G0
4. Note target maturity profile for this enhancement (what level should each dimension reach?)

Present:
```
Current Maturity: C_/E_/A_/P_/G_ (effective: L_)
Constraining Dimension: [dimension] at level [N]
Target After Enhancement: C_/E_/A_/P_/G_
```

**Note**: This phase is informational — it guides effort allocation but does not gate progression. Skip if spec doesn't exist yet (NEW category without prior version).

### Step 3: Phase 1 — Governing Spec Verification

**Objective**: Read governing spec(s) before modifying anything.

Execute:
1. Identify governing spec(s) for the artifact type:
   - Skill specs → read SKILL_NAMING_CONVENTION_SPEC (in `aget/specs/`; note: AGET_SKILLS_SPEC is archived)
   - SOPs → read AGET_SOP_SPEC
   - General specs → read AGET_SPEC_FORMAT (if exists)
2. Read each governing spec — extract specific requirements (CAP-xxx-nnn)
3. Verify cross-reference targets in governing spec are not stale:
   ```bash
   # For each referenced file in the governing spec
   test -f "<referenced_path>" && echo "OK" || echo "STALE: <path>"
   ```
4. If stale targets found, flag for remediation at Phase 5a

**Anti-Pattern Check**: Decorative Spec Basis
- After reading governing spec, verify you can cite specific CAP requirements
- If you declared a spec basis but can't cite specific requirements → STOP and re-read

**Exit Criteria**: Governing spec(s) read and cited. Cross-reference targets verified.

### Step 4: Phase 2 — Census & Classification

**Objective**: Inventory all artifacts that need enhancement.

Execute:
1. Enumerate all specs/artifacts in scope
2. For each artifact, read the file to get current version and status (NEVER assert from memory — L611)
3. Identify dependency order
4. Produce census table:
   ```
   | Artifact | Current Version | Category | Dependencies | Priority |
   |----------|:---------------:|----------|--------------|:--------:|
   | <spec path> | <from file read> | <category> | <specs it references> | <1-N> |
   ```

**USER CHECKPOINT**: Present census table and ask:
```
Census: <count> artifacts in scope
Dependencies: <dependency chain summary>

Approve census and proceed? (yes/add artifact/remove artifact/abort)
```

**Exit Criteria**: All artifacts inventoried. Dependencies mapped. Priority assigned.

### Step 5: Phase 3 — Draft/Update Spec

**Objective**: Write or update the specification content.

Execute:
1. Update spec header (version, status, date)
2. Write requirements in EARS patterns:
   - Ubiquitous: `The SYSTEM shall...`
   - Conditional: `IF condition THEN the SYSTEM shall...`
   - Event-driven: `WHEN event THEN the SYSTEM shall...`
   - Optional: `WHERE condition, the SYSTEM should...`
   - Prohibited: `The SYSTEM shall NOT...`
   - Unwanted: `artifact shall NOT...`
3. Assign unique requirement IDs: R-{DOMAIN}-{NNN}
4. If new vocabulary terms introduced, add SKOS vocabulary section
5. If applicable, use "codify existing behavior" approach (fastest — L622 evidence)

**USER CHECKPOINT**: Present draft spec content and ask:
```
Spec: <path>
Version: <current> → <target>
Requirements: <count> total (<new_count> new, <modified_count> modified)

Review spec content. Approve and proceed? (yes/edit/abort)
```

**Exit Criteria**: Spec content written. All requirements have IDs and EARS patterns.

### Step 6: Phase 4 — Structural Validation

**Objective**: Verify spec meets structural requirements.

Execute:
1. Check required sections present per governing spec format
2. Check naming conventions
3. Validate EARS pattern compliance for all requirements:
   ```
   For each requirement:
   - Has unique R-{DOMAIN}-{NNN} ID? ✓/✗
   - Uses EARS pattern keyword? ✓/✗
   - Pattern type recorded? ✓/✗
   ```
4. Verify version number follows semantic versioning
5. Run structural validators if available:
   ```bash
   python3 .aget/patterns/sync/validate_spec_compliance.py <spec_path> 2>/dev/null
   ```

Report results:
```
Structural Validation: <PASS/FAIL>
  Sections: <count>/<expected> present
  Requirements: <count> with valid EARS patterns
  Naming: <PASS/FAIL>
  Version: <PASS/FAIL>
  Validator: <PASS/FAIL/not available>
```

**Exit Criteria**: Spec passes structural validation. No format errors.

### Step 7: Phase 5 — Cross-Reference Wiring

**Objective**: Connect spec to related artifacts with bidirectional traceability.

#### Phase 5a: Stale Artifact Remediation

Before wiring, verify all targets are current:
```
For each cross-reference target:
1. File exists? ✓/✗
2. Version/status current? ✓/✗
3. If stale → flag for update before wiring
```

If stale targets found, present remediation plan before proceeding.

#### Phase 5b: Primary Cross-Wiring

For each requirement:
1. Link to evidence L-doc(s) — why does this requirement exist?
2. Link to V-test — how is this requirement enforced?
3. Link to related specs (broader/narrower/related)

Produce cross-reference report:
```
| Requirement | Evidence (L-doc) | Enforcement (V-test) | Related Specs |
|-------------|-----------------|---------------------|---------------|
| R-XXX-001 | L### | V-XXX-001 | SPEC_Y |
```

#### Phase 5c: Bidirectional Verification

For each outgoing reference, verify target acknowledges this spec.
For each incoming reference, verify this spec acknowledges source.

**USER CHECKPOINT**: Present cross-reference report and ask:
```
Cross-References: <count> requirements mapped
Stale targets: <count> (remediated: <count>)
Bidirectional gaps: <count>

Approve wiring and proceed? (yes/fix gaps/abort)
```

**Exit Criteria**: All requirements trace to evidence and enforcement. Bidirectional references verified.

### Step 8: Phase 6 — Self-Compliance Check

**Objective**: Verify the enhancement complies with its own specs.

Execute — these checks are MANDATORY (C3 — no skip option):

1. **Bootstrapping paradox check** (L560):
   - Does the spec enhancement comply with the lessons it references?
   - Does the output comply with its own requirements?
   ```
   For each R-requirement in the spec:
   - Does this enhancement process follow this requirement? ✓/✗
   ```

2. **Spec self-compliance**:
   - Does the spec meet the format requirements of its governing spec?

3. **Vocabulary compliance**:
   - Are all terms used consistently with controlled vocabulary?

Report:
```
Self-Compliance Check:
  Bootstrapping paradox: <PASS/FAIL — details>
  Spec self-compliance: <PASS/FAIL — details>
  Vocabulary compliance: <PASS/FAIL — details>
  Overall: <PASS/FAIL>
```

**USER CHECKPOINT**: Present compliance results and ask:
```
Self-compliance: <PASS/FAIL>
<details of any failures>

Acknowledge results? (acknowledge/fix and re-check/abort)
```

**Exit Criteria**: Self-compliance confirmed. No bootstrapping paradox detected.

### Step 9: Produce Outputs

Generate final deliverables:
1. **Updated/New Spec**: The spec file (already written in Phase 3, validated in Phase 4)
2. **Cross-Reference Report**: From Phase 5b (formatted table)
3. **Enhancement Decision Log**:
   ```
   Enhancement Decision Log:
     Category: <category>
     Governing Spec: <spec> (CAP citations: <list>)
     Census: <count> artifacts
     Requirements: <count> (EARS: <breakdown>)
     Self-Compliance: <PASS/FAIL>
     Completed Phases: <list>
     Skipped Phases: <list with justification>
   ```
4. **Stale Artifact Report**: From Phase 5a (if any)

Present summary:
```
=== /aget-enhance-spec Complete ===

Category: <CATEGORY>
Target: <spec path>
Version: <previous> → <new>
Requirements: <count>
Cross-References: <count> mapped
Self-Compliance: <PASS/FAIL>
Phases Executed: <list>

Outputs:
  - Spec: <path>
  - Cross-Ref Report: <inline or path>
  - Decision Log: <inline>
```

## Constraints

These are INVIOLABLE — you MUST NOT violate these constraints:

1. **C1**: NEVER skip Phase 1 (Governing Spec Verification)
2. **C2**: NEVER assert spec state from memory — always read files (L611)
3. **C3**: NEVER skip Phase 6 (Self-Compliance Check) — no skip option exists
4. **C4**: NEVER execute deployment phases 7-10 (template, tooling, fleet, post-deploy)
5. **C5**: ALWAYS present user checkpoint before proceeding past Phases 0, 2, 3, 5, 6
6. **C6**: ALWAYS produce cross-reference report as output
7. **C7**: ALWAYS cite governing spec requirements (CAP-xxx-nnn) in deliverables
8. **C8**: ALWAYS use EARS patterns for requirements
9. **C9**: ALWAYS assign unique R-{DOMAIN}-{NNN} IDs to requirements

## Anti-Pattern Detection

The skill actively checks for these anti-patterns:

| Anti-Pattern | Phase | Detection |
|--------------|:-----:|-----------|
| Decorative Spec Basis | 1 | Spec basis declared but no CAP citations in deliverables |
| Memory-Based Assertions (L611) | 2 | Version/status claimed without file read |
| Bootstrapping Paradox (L560) | 6 | Output violates its own requirements |
| Certainty Bias (L555) | 6 | Self-compliance check skipped |
| ADR-008 Progression Skip | 0 | Generating skill artifacts without SOP/spec layers |

## Related Skills

- `/aget-create-project` — Create project plan for enhancement work
- `/aget-propose-skill` — Propose new skills (used in this project at G3)
- `/aget-record-lesson` — Capture lessons discovered during enhancement
- `/aget-review-project` — Review enhancement project progress
- `/aget-study-topic` — Research topic before enhancing


### Step 10: Skill Completion Signal (D71 Layer 3)

**Purpose**: Make skill execution completeness structurally visible. If this signal is absent, the skill execution is incomplete.

The skill MUST output this signal as the LAST section of the report:

```markdown
## Skill Completion Signal
**Self-Compliance**: PASS | FAIL (items: [list])
**Cross-References**: PASS (N mapped) | SKIP (justification)
**Phases Completed**: 0→1→2→3→4→5→6→9→10
```

If Phase 6 (Self-Compliance Check) was not executed, the signal MUST report FAIL. Per C3, Phase 6 has no skip option.

**Enforcement**: Strict (ADR-008, D71). Signal absence = incomplete execution.

## Traceability

| Link | Reference |
|------|-----------|
| Spec | SKILL-041 (`aget/.aget/specs/skills/SKILL-041_aget-enhance-spec.yaml`) |
| SOP | `sops/SOP_specification_enhancement.md` v1.0.0 |
| L-docs | L622 (lifecycle), L623 (meta-governance gap), L560, L557, L555, L611 |
| Proposal | SP-002 (`planning/skill-proposals/PROPOSAL_aget-enhance-spec.md`) |
| Project | `planning/PROJECT_PLAN_aget_enhance_spec_skill_v1.0.md` |
| ADR-008 | L-docs → SOP (Advisory) → Spec (Strict) → Skill (Generator) |
| Verb | "enhance" — 25th approved verb, 4th domain innovation |

---

*aget-enhance-spec v1.1.0*
*Category: Governance*
*ADR-008 Position: Generator (step 3 of Advisory → Strict → Generator)*
*"Five plans, three agents, one skill."*

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.