A compact autonomous software-engineering skill library for evidence-first, intent-aligned coding agents.
Core idea: agents earn trust with evidence, small diffs, falsifiable goals, and honest stops — not confidence.
Start with using-evidence-driven-engineering.
| Phase / risk | Skill |
|---|---|
| Routing | using-evidence-driven-engineering |
| Intent, ambiguity, vocabulary | intent-alignment |
| Turning requests into pass/fail loops | falsifiable-goals |
| Specs and vertical-slice plans | design-planning |
| Small scoped diffs | surgical-changes |
| RED/GREEN/REFACTOR | test-driven-development |
| Bugs and failed builds | diagnostic-debugging |
| Refactors and seams | architecture-deepening |
| Completion claims | verification-honesty |
| Probing frontier capability / capability spikes | frontier-probing |
| Delegation and handoffs | subagent-orchestration |
| Review | two-stage-code-review |
| Git safety | git-safety-and-guardrails |
| Issue queues | triage-queue-management |
| Skill writing | skill-authoring-tdd |
| Anti-sycophancy | communication-integrity |
| Uncertainty / capability stops | failure-vocabulary |
- If a skill plausibly applies, load it before responding.
- If intent is ambiguous, recommend a default or stop before editing.
- Convert vague work into a failing test, reproducer, metric, or explicit verification loop.
- Every changed line must trace to the user's request.
- No speculative abstraction, config, seam, or impossible-case error handling.
- Production code starts with a failing test when feasible.
- Debugging starts with a feedback loop, not a hypothesis.
- Completion claims require fresh evidence in the same turn.
- Review spec compliance before code quality.
- Destructive git needs hooks or explicit typed confirmation.
- Skills:
skills/software-development/*/SKILL.md - Git guard script:
skills/software-development/git-safety-and-guardrails/scripts/block-dangerous-git.sh - Manifest:
manifest.json