AI with work experience.
A standard agent knows nothing about recruiting, accounting or project work. skillfactor compiles public occupation data, real job ads and expert knowledge into a ready-to-use skill per profession — a git-native library that gets better with every project. Built for computer-based work: the 1,953 white-collar professions come first.
artificial-intelligence-engineer/
├─ SKILL.md
├─ references/
│ ├─ profile · tasks · skills
│ ├─ market.md ← live job-ad evidence
│ ├─ ai-skills.md ← … tiered agent skills
│ ├─ practitioner-qa.md ← Stack Exchange
│ ├─ usecases · intake · quality
│ └─ glossary · literature
├─ evals/
└─ PROVENANCE.md
Every processed occupation gets its skill rankings from real, relevance-checked job ads — click to browse the finished ones.
The crawl started … and has been running for … — computer-based professions first.
Each ad is distilled by a self-hosted model (gemma3 on our GPU — no data leaves the house) into structured facts: skills, tools, seniority. One count = one ad.
Public occupation knowledge becomes a skill
Occupational knowledge belongs to nobody — so nobody maintains it. skillfactor changes that: three sources, one compilation, versioned in git.
Occupation taxonomies
3,000+ occupations from ESCO & O*NET — public, standardized, multilingual. Joined via the official crosswalk.
Job ads
Current skill requirements from global job platforms — weighted market evidence with percentages and an as-of date.
Expert knowledge
Best practices from literature and the web, 900+ proven agent skills from 16 open-source libraries (Anthropic, NVIDIA, Google, community) — and curated practitioner Q&A from six Stack Exchange communities, each entry attributed (CC-BY-SA).
Every item knows where it comes from
After compilation, each package reports its source mix.
Shown here: the AI-engineer package with every source compiled —
every repo carries the same breakdown as PROVENANCE.md with a rendered chart.
- Job boards — market evidence
Full market report from real job ads (JSearch API): ranked requirements with share, seniority distribution, title variants — extracted facts only, aggregated live.
53% - O*NET — tasks & tools
Task statements, work activities and software from the U.S. occupation database.
31% - Wikipedia & AI expert curation
Glossary, literature, use cases, intake questions, quality criteria and evals — AI-curated with cited web sources.
10% - ESCO — occupation & competences
The European profile: essential and optional competences per occupation.
6% - External AI skill packs — mapped
Proven agent skills from the top open-source libraries (anthropics/skills, obra/superpowers, wshobson/agents), matched to the occupation — linked with per-source attribution, never copied.
1% - Stack Exchange — practitioner Q&A
What experienced practitioners actually advise: quality-filtered questions and answers from six professional Stack Exchange communities (Workplace, Project Management, Law, Money, Software Engineering, Data Science), condensed into per-profession insights — each entry attributed to its author, CC-BY-SA 4.0.
0%
Numbers load live from the package manifest. The job-ad rollout for all 3,039 occupations is running — packages gain their market-evidence segment as it lands. See the library-wide chart →
Open a package. Read what the agent knows.
Four occupations, one click each — every file renders directly in the repository.
AI engineer
Every source compiled: taxonomy profile, gated job-ad evidence (US + DACH), expert curation, tiered agent skills and Stack Exchange practitioner Q&A.
Lawyer
Legal profile, case-related tasks and competences — plus mapped compliance & contract agent skills.
Database administrator
The deepest AI-skill mapping in the catalog: 90 agent skills for design, migrations, ops and security.
Web designer
Design competences plus 49 mapped skills — frontend, accessibility, UI patterns and brand systems.
One agent per employee — preloaded with the profession
The agent knows the job from day one — the employee makes it their own. A sparring partner that gets better every day and never quits.
Preloaded
Starts with the skill profile of the profession — from ESCO, O*NET and market data.
Knows the day-to-day
Mails, chats, documents, meeting transcripts — the personal work context, every day.
Grows with you
Every task and every correction improves it; hidden abilities surface in the skill graph.
Stays in the company
Knowledge stays with the role — even when the person leaves. The role never forgets.
Works with Zeiterfassung.CLOUD
The HR API knows every profession a person has — practiced, trained, or just loved. The agent loads them all.
Many professions, one person
relation_type: practiced · trained · secondary · hobby. The payroll answer stays
unambiguous (is_primary, Tätigkeitsschlüssel/KldB); the agent still sees the whole human.
Hobby skills count
Skills attach to the person, not the diploma. Self-assessed abilities flow in with provenance
(self · supervisor · imported · skillfactor) — hidden abilities get a structured data
source, not just observation.
Right person, right slot
skill_requirements per project, team or shift; the matching engine
(find_matching_employees) returns who qualifies, with match percentage.
practiced · primary · KldB 24522hobby · self-declared[
{ "profession": "Industriemechaniker", "slug": "industrial-machinery-mechanic",
"relation_type": "practiced", "is_primary": true, "kldb_code": "24522" },
{ "profession": "Fotograf", "slug": "photographer",
"relation_type": "hobby", "is_primary": false, "source": "self-declared" }
]
Mapped via ESCO URI · O*NET-SOC · KldB 2010 (Tätigkeitsschlüssel digits 1–5) — no proprietary IDs.
Private stays private. Public knowledge grows.
Knowledge rises by proving itself
- Same insight in ≥ 2 projects → suggested promotion into the org overlay
- Proven org knowledge → anonymized second distillate as a community PR
- Hard rule: nothing leaves
projects/automatically — the path outward is always an explicitly confirmed, generalized, anonymized distillate - Agents load the cascade Community → Organisation → Project; more specific beats more general
Your profession's skill — in Claude, ChatGPT and Gemini
The git repo is the canonical format: plain Markdown, YAML, JSON — no vendor syntax in the content. Thin adapters build the same knowledge into an Anthropic agent skill, a Custom-GPT bundle and a Gemini bundle; one MCP gateway handles reading and writing for all three.
Claude
Agent-skill package per profession: trigger description, contribution behaviour, competence files as lazy references — installable via plugin marketplace.
ChatGPT
Custom-GPT bundle: generated system instruction + competence files as knowledge upload; writing goes through an Action onto the same gateway.
Gemini
Instruction + grounding files; writing via function calling against the same gateway. One content, three thin adapters.
Enrichment path A — chat distillate
Solved something real in a conversation? The skill offers to save it: a project distillate into your private org repo, and — if a transferable core exists — an anonymized distillate as a community pull request. You confirm the full text; nothing commits directly to main.
Enrichment path B — M365 mining
A connector mirrors mails, Teams chats and meeting transcripts into SQL Server. A batch job extracts lesson candidates (consent-gated, roles instead of names, raw data never leaves the database) and files them as a review queue of pull requests — no auto-merge.
One sentence saves the lesson
A project manager — SkillFactor skill installed, connector linked — just learned in a Claude conversation that customer A ignores her slide-deck status reports, but responds within hours to a three-bullet email. She keeps that experience:
Save the key insight from this chat as a lesson learned in my SkillFactor repo.
Happy to — I suggest two placements. Please review both full texts; exactly what you confirm gets committed.
# Customer A: status updates as a three-bullet email **Situation:** Weekly slide-deck status reports to customer A went unanswered for weeks. **Insight:** Customer A's project lead reads email on the go and never opens attachments. A three-bullet email — done / next / blocked — gets a response within hours. **Action for this project:** Send Friday status as a three-bullet email; keep the deck for the monthly review only.
# Match the report format to how the reader actually reads **Pattern:** Status reports fail when the format suits the sender, not the reader — decks go unread by people who live in their inbox. **Transferable practice:** Ask (or observe) how your counterpart consumes information, then default to the shortest format that fits it — e.g. a three-bullet email: done / next / blocked. Keep long formats for milestones.
Both, just one, or neither?
both
Done — both filed as pull requests (never a direct commit):
✅ Org repo: PR #41 · customer-a/rollout-2026 — "Status as three-bullet email"
✅ Community: PR #587 · project-manager — "Match the report format to the reader"
Both await maintainer review. I checked the target folders for similar lessons first — no duplicates found.
Works the same in ChatGPT and Gemini (adapters on the same gateway). Requirements: the profession skill installed + the SkillFactor connector linked — setup guide →
Refreshed monthly — skills follow the market
The role never forgets.
3,039 professions are ready — from the production-depth recruiter to the ICT architect, enriched with the best open-source agent skills. Every skill is a repo, every release a reviewed pull request.