Reusable skills for evidence-backed research, approval-gated operations, structured records, and durable agent handoffs.
These packages extract the operating patterns behind real agent-assisted products without publishing their private prompts, source code, credentials, or user data.
| Skill | Use it when | Produces |
|---|---|---|
evidence-backed-research |
Claims need traceable sources and honest confidence labels | A structured evidence packet with conflicts and gaps |
human-approved-publishing |
An agent may prepare content but a person must authorize publication | A reviewed draft and explicit publication decision |
schema-validated-records |
Research must become machine-readable without losing provenance | A schema-valid record plus validation report |
context-packet-builder |
A fresh agent needs the smallest useful view of file-backed state | A compact, prioritized handoff packet |
flowchart LR
A["Source material"] --> B["Bounded skill"]
B --> C["Structured artifact"]
C --> D["Deterministic validation"]
D --> E["Human decision"]
E --> F["External action, if approved"]
The common design principle is simple: agents may gather, transform, check, and recommend. External actions remain visible and approval-gated.
Copy a skill directory into the skill location used by your agent environment, or point
your agent at its SKILL.md. Each skill is self-contained and links only to files inside
its own directory.
The instruction format is compatible with the common SKILL.md convention. The included
scripts are tested with Python 3.11+.
python3 scripts/validate_skills.py
python3 -m unittest discover -s tests
python3 skills/context-packet-builder/scripts/build_context_packet.py \
--input skills/context-packet-builder/examples/input/project-state.json \
--output /tmp/context-packet.mdCI runs the same checks on every pull request.
- Synthetic examples only.
- No private URLs, personal identifiers, customer records, or environment-specific paths.
- No claim that a model verified something unless a deterministic check or human review did.
- No external send, submit, merge, or publish action without an explicit approval step.
- Small core instructions; supporting examples live beside the skill.
These skills are sanitized extractions from patterns used in TheTechStack and a file-backed agent operations system. The implementations here are standalone public artifacts, not copies of either private system.