AKM Eval — Evidence-based self-assessment for agentic knowledge management

Assess how a person + knowledge repository + agent runtime + operating loops work together. AKM Eval applies the AKM Index v1.2 rubric: five pillars, 25 behaviourally anchored criteria, and a weighted score out of 100.

Skill release: 1.2.2 · Rubric: 1.2 — these version numbers describe different artefacts. The v1.2 specification adds descriptive dimensions; its scoring anchors remain those of v1.1.

Website · Full English rubric · English source · Participation guide

What it assesses

The target is the operating system for knowledge work—not the quality of a single note, a person's intelligence, or the number of AI tools installed.

Each pillar has five criteria rated 0–4: absent, ad hoc, documented, systematised, and self-improving. The evidence-based audit requires artefact-linked judgements; unsupported claims are capped at level one. Level four requires recorded feedback leading to a system change, not simply more automation. Read the complete anchors and their exceptions before scoring.

The M0–M4 labels are interpretive bands, not validated population norms. The weighted total, ordinal-level arithmetic, N/A handling and boundary conventions should be interpreted with the full specification. A high score is not a safety certification, independent external audit, or proof of improved productivity.

Outputs

• A structured akm-report.json: declared scope and version, environment, criterion ratings, evidence notes, domain totals, strengths, improvement priorities and optional history.
• A Markdown summary and, for submitted reports, an HTML report following the current skill workflow.
• With separate public-display permission, an assessment-board card and—if detailed publication is also permitted—a public report.
• An optional descriptive S / A / O profile: scale, recurring-operation autonomy and output coupling. These describe context rather than add points. See the English dimensions specification.

The website supports Korean and English. Participant names, source material and community comments may remain in their original language; the site does not automatically translate or reinterpret a participant's evidence.

Requirements

• Claude Code, or another coding agent able to read the operating instructions and inspect the relevant environment.
• A persistent knowledge repository and an agent runtime connected to it.
• Intended repeated use, with activity traces within the preceding 30 days.

Install

Clone into your development directory and link the skill. The commands do not overwrite an existing skill installation:

git clone https://github.com/johnfkoo951/akm-eval.git ~/DEV/akm-eval
mkdir -p ~/.claude/skills
ln -s ~/DEV/akm-eval ~/.claude/skills/akm-eval

If the destination already exists, inspect your existing installation instead of replacing it blindly. Alternatively, download the skill archive, inspect its contents and extract it to a new directory.

Run in English

Start a fresh Claude Code session, then ask:

Run the AKM Index. Read SKILL.md and the English companion guide SKILL.en.md.
Use references/rubric-en.md for scoring. Write the assessment and explanations in English.
Declare the system boundary, distinguish unavailable evidence from absence, and do not change my system while assessing it.
Show the upload notice and respect my separate choices for submission and public display.

The process is: check eligibility → inspect evidence → score conservatively → challenge the judgements → explain strengths and improvements → create the report → show the submission notice. Typical duration in the instructions is 10–60 minutes, not a guaranteed runtime.

SKILL.md remains the executable skill entry point. SKILL.en.md is its English reading companion, not a separate installed skill or a changed scoring protocol.

Submission, research use and privacy

Read this before running the skill. The current workflow informs you before sending the final report to CMDSPACE and proceeds unless you decline. If you do not want a submission, say so explicitly; the assessment can remain in your agent environment.

• Submission and public display are different choices. submit governs transmission. publicCard and publicDetail govern public display; showRealName and showEmail govern those identity fields.
• The collection notice states research, rubric improvement, and community return of anonymised patterns/statistics as purposes. Raw reports are retained privately and are not an open dataset.
• A successful submission returns a receipt identifier. Use it when requesting correction or withdrawal through [email protected].
• Assessment runs in your agent environment. If you use a hosted AI provider, that provider's data handling still applies. “Not submitted to CMDSPACE” does not mean no data are processed by your model provider.
• Public-display flags do not, by themselves, certify the ethics or permissions of every possible secondary research use.

The paired self-report study is a separate measurement task: perceived ratings are collected before an evidence-based audit under the study form's research notice and consent. Do not mistake its evidence-free 0–4 ratings for completed audit results.

Repository map

akm-eval/
├── README.md                     # English overview (default)
├── README.ko.md                  # Korean overview
├── SKILL.md                      # Executable operating instructions
├── SKILL.en.md                   # English reading companion
└── references/
    ├── rubric.md                 # Korean rubric v1.2
    ├── rubric-en.md              # English rubric v1.2
    ├── dimensions.md             # Descriptive dimensions
    ├── dimensions-en.md          # English dimensions companion
    ├── report-schema.md          # Report schema and submission contract
    ├── report-schema-en.md       # English schema companion
    ├── report-template.html      # Local HTML report template
    └── runtime-paths.md          # Runtime evidence-discovery reference

Research status and citation

AKM is an implemented, evolving diagnostic rubric. Independent inter-rater reliability, construct validity, and user outcomes should not be assumed from operational use or from model-assisted checks. When citing an assessment, identify the rubric version, mode, evidence boundary and applicable scoring policy.

Suggested attribution: “Based on AKM Index v1.2 (https://akm.cmdspace.work).” No arXiv identifier or DOI for a new manuscript is claimed here.

License and author

MIT for the software and associated repository documentation. A research manuscript may carry a separate publication license.

Created by Yohan Koo (CMDSPACE). By CMDSPACE.