Skills and agents
Canonical files: .agents/skills/<name>/SKILL.md. agents list reads that tree. init copies it. Editing a copy inside .claude or .cursor and not the canonical file will be overwritten on the next forced install, and will not ship.
The skills are instructions. They are not models, not processes, and not a guarantee. A host that does not load SKILL.md will not follow them. Live loading has not been smoke-tested in each IDE. See Roadmap.
Shared contract#
Every skill in this repository is required by test to:
- start with YAML frontmatter
- set
nameequal to the directory name - include a description
- include a
## Safe workflowsection that is not a copy of another skill's workflow - tell the model not to fabricate, claim, or approve results it does not have
Each workflow is a different procedure: prime routes one lane, strategy picks the lane, unit reads focused-test markers, flow reads a written journey, api reads a contract, visual reads an attached image note, access reads source text, load reads a written budget, data reads a named fixture, secure reads a pasted diff for literals, regression reads a --changed report, flake reads a pasted rerun log, defect quotes a pasted error, pipeline reads the workflow files the scan listed, release runs release-check only when asked, evidence reads the JSON, verify compares verify output to saved stdout. None of them run a suite, open a browser, or sign a file.
If an assistant using these skills claims "all tests passed" and cannot point at a command and an output, it is breaking the skill. Prefer the CLI report.
The roster#
| Skill | Job |
|---|---|
pramanasutra-prime | Coordinates inspection, QA strategy, evidence review and a release recommendation. Does not approve a deploy |
pramanasutra-strategy | Risk-based quality strategy and traceability from requirement to test |
pramanasutra-unit | Focused, deterministic unit and component tests with real assertions |
pramanasutra-flow | Cross-system and end-to-end scenarios that can be repeated |
pramanasutra-api | REST, GraphQL, gRPC and event contracts, and whether tests cover them |
pramanasutra-visual | Screenshot baselines and rendering gaps. Does not take screenshots by itself |
pramanasutra-access | Accessibility and cross-device planning, including manual checks |
pramanasutra-load | Performance test design, thresholds and abort rules. Not a license to load-test production |
pramanasutra-data | Migration, backup, rollback and fixture review |
pramanasutra-secure | Security-test gaps from evidence. Not a penetration test |
pramanasutra-regression | What changed, and which tests are relevant |
pramanasutra-flake | Nondeterministic tests, using a reproduction, not a hunch |
pramanasutra-defect | A failure turned into a minimal reproduction |
pramanasutra-pipeline | Whether automation and gates mean what the team thinks |
pramanasutra-release | Criteria, rollout and rollback. Human approval stays human |
pramanasutra-evidence | Reports and evidence that stay traceable and avoid leaking secrets |
pramanasutra-verify | Challenges the gate before someone approves |
Install paths#
| Host | Path | Verified live in that host? |
|---|---|---|
| Claude Code | .claude/skills | No |
| Cursor | .cursor/skills | No |
| OpenAI Codex | .agents/skills | No |
| GitHub Copilot | .github/skills | No |
| Gemini CLI | .gemini/skills | No |
| Generic | .agents/skills | No |
Product names belong to their owners. Listing a path is compatibility information, not an endorsement and not a partnership. See Legal.
How Prime is meant to delegate#
The skill text tells the model to delegate to a specialist only if the host supports delegation, and otherwise to do the stages in order. PramanaSutra does not implement an orchestrator process. There is no message bus. "Agent team" in the README means seventeen instruction files and one chief file, not seventeen running programs.
What to put in a prompt#
Name the skill, name the repository, and restate the two rules that matter most: cite file and line, and do not claim a test ran unless a tool returned that result. Ask it to read report.md if you have already produced one. Do not paste production secrets into the prompt. The skill tells the model not to collect secrets; the prompt should not hand them over anyway.
Changing a skill#
Edit .agents/skills/<name>/SKILL.md. Keep the frontmatter name equal to the folder. Keep the safe workflow. Run npm test. The test reads every skill. Copies in a target project are not updated until someone runs init again with --force after review.