Repository navigation
Meta Tools
Meta-tools implement the factory-of-factories principle: instead of creating artifacts, they create systems that create artifacts. One command produces complete production capability.
The department factory. Given a domain description, MetaBuild bootstraps an entire Streeling University department in a single command:
Input: domain description (e.g., "cybernetics — feedback systems and control theory")
Output:
├── grammars/sci-cybernetics.ebnf (EBNF research grammar)
├── state/streeling/departments/cybernetics.department.json (department state)
├── schemas/research-weights for the department
├── tests/behavioral/cybernetics-cases.md (behavioral test suite)
└── Persona mapping to existing catalog
What MetaBuild validates before completing:
- Grammar has at least 6 sections (consistent with core grammars)
- Department JSON validates against department schema
- Behavioral test has at least 5 test cases
- Persona mapping exists in the persona catalog
Validated results (Batch B): 8 departments bootstrapped via MetaBuild — Audio Engineering, Data Visualization, Cybernetics, MetaQA, Semiotics, Network Science, Information Theory, Systems Engineering.
The pattern generalizer. Don't just fix the problem — fix the system that allowed the problem.
When a bug or governance gap is found, MetaFix:
- Fixes the immediate instance
- Identifies the root cause pattern
- Searches for other instances of the same pattern
- Proposes a policy or constitutional note to prevent recurrence
- Creates a behavioral test case for the pattern
- Updates the evolution log with the fix
Bug: README count says 17 policies, actually 30
MetaFix:
→ Fix: update README count
→ Pattern: artifact count drift
→ Policy: readme-sync-policy.yaml (create or update)
→ Test: metasync-cases.md (add test case)
→ Prevention: README-SYNC-ON-MERGE IxQL pipeline
This is the meta-fix principle — fix the category, not just the instance.
The drift detector. Continuously verifies that documentation reflects reality:
- README counts match actual file counts
- CLAUDE.md descriptions match actual behavior
- Cross-repo references point to existing artifacts
- Grammar section counts match actual grammar content
- Policy catalog lists match actual policy files
MetaSync runs:
- Weekly (Monday 07:00 UTC) via
pipelines/metasync.ixql - After every push to master
- On demand via
/demerzel-metasync
When drift is detected, MetaSync either auto-fixes (low risk) or creates a GitHub issue with a drift report.
The pattern spotter. Detects when a recurring observation is becoming a pattern worth promoting.
Design spec: docs/superpowers/specs/2026-03-24-meta-recognition-engine-design.md
Meta-recognition monitors:
- PDCA cycle outcomes for repeated patterns (3+ occurrences → flag for promotion)
- Session observations for anti-patterns (3+ occurrences → flag for policy)
- Evolution log for artifacts with zero citations after 30 days (flag for deprecation)
- Behavioral tests for gaps in coverage (completeness instinct signal)
The health check. Before any complex autonomous operation, run self-diagnostic to detect:
| Check | What It Detects |
|---|---|
| Context exhaustion | Token budget too low to complete task |
| Action loops | Same tool being called repeatedly without progress |
| Wrong repo | Operating on wrong directory/branch |
| Git failures | Uncommitted changes blocking operations |
| Task drift | Current work diverging from original goal |
| Confidence miscalibration | Claiming T for something that is U |
Self-diagnostic runs automatically before:
- Starting a
/demerzel-drivecycle - Spawning a
/demerzel-team - Running any
--autonomousflag operation
From the user's feedback that shaped Demerzel's architecture:
"Don't create things, create systems that create things."
This principle means:
- MetaBuild doesn't create one department — it's the system that creates any department
- MetaFix doesn't fix one bug — it's the system that fixes categories of bugs
- MetaSync doesn't check one README — it's the system that ensures all docs stay current
- The Seldon Plan doesn't run one research cycle — it's the system that runs all research
The leverage is always in the system, not the artifact.
MetaBuild is powerful — it can produce a department in 60 seconds. This makes the discipline to NOT build even more critical. Before running MetaBuild:
- Does this department have a clear consumer? (Which repo will use it?)
- Is there an existing department that covers this domain? (Extend, don't duplicate)
- Who will commission the first research cycle? (Has a real question to answer?)
- What is the ERGOL metric for this department? (How will it deliver value to a real user?)
See Manifesto principle 3, and the Anti-LOLLI policy.
- Architecture — meta-tools role in the system design
- Streeling-University — MetaBuild output
- Policies — MetaFix generates policies; MetaSync enforces readme-sync
- IxQL — MetaSync expressed as IxQL pipeline
- Manifesto — factory-of-factories as a core principle