AI Interaction Guidelines
This file defines the default rules for AI-assisted work. Project-specific instructions override these defaults when explicit.
Communication
- Be concise, direct, and honest.
- Separate facts, assumptions, recommendations, and unresolved decisions.
- Do not invent answers for
TBDitems. - After 2–3 grounded failed approaches, stop and explain the blocker.
Human Approval
Ask before:
- architecture or dependency changes
- database, auth, payment, secrets, or security-sensitive changes
- destructive commands or file deletion
- Git history rewriting
- commits, merges, releases, or deployments
- adopting prototype code into production
- writes outside the repository, such as tickets in a shared store
- adopting any change to Pathfinder’s skills, roles, context, or templates that a harvest proposed
The human owns judgment, acceptance, merge, and release decisions.
Git and Delivery
Follow context/project-overview.md.
Do not assume branch strategy, commit style, pull requests, versioning, or deployment workflow.
Inspect current Git state before acting. If the workflow is unclear or
TBD, ask.
Delivery Workflow
Use the workflow skills instead of recreating their procedures in chat:
to-tickets— slice one approved Feature into executable tickets./ticket load— load one ticket, its Feature, and relevant context./ticket start— implement that ticket. Optional before review:/ticket adversarychallenges the implementation through reproducible experiments; Tester independently verifies them./ticket review— verify the work and report findings./ticket complete— complete accepted work and durable records, and name the tickets that are now ready.learn-feature— optionally teach what was implemented.reflect— record friction as it appears, harvest periodically, and let the human decide what becomes a Feature.
In human-in-the-loop mode, the human coordinates this loop. In orchestrator
mode, orchestrate coordinates the same ticket actions across dependency-safe
workers in separate worktrees. Successful implementation checkpoints Adversary
pending; a separate bounded Adversary session precedes independent Tester.
Suspected defects reach Tester first. Full confirmed findings and experiments
cross sessions through the transient current-ticket checkpoint at the exact PR
head. Repairs rerun Adversary then Tester; unchanged integration-only refresh
requires fresh Tester/CI. Only current-head reviewed work reaches done and the
integrator. One
worker’s human gate does not stop unrelated workers. The execution-mode rule
lives in skills/ticket/SKILL.md; neither mode changes human authority.
Lifecycle skills assume their responsible role automatically. Explicit
activation with /role is optional and overrides that default for the session.
Status
Ticket status records durable lifecycle state only:
Proposed → Ready → In Progress → Complete
Cancelled and Superseded are terminal alternatives.
Readymeans the human approved execution.- Review and testing are workflow activity, not a status. Human-in-the-loop
coordination remains optional; orchestrated delivery requires its separate
Adversary and Tester phases. A ticket stays
In Progressuntil it is complete. Blockedis not a status. A ticket’s blockers are the edges under its## Blocked by, and anything else that stops work is recorded in current workspace state.- A Feature’s status is derived from its tickets, never maintained by hand.
- The human decides approval, acceptance, cancellation, and supersession.
Context Discipline
- Read only what the current work requires.
- Prefer exact files or sections over broad repository scans.
- Do not load history, roadmap, other tickets, or unrelated context by default.
- Work one active ticket per worker. Human-in-the-loop mode uses one ticket session; orchestrator mode may coordinate several isolated workers.
- If the work can no longer be understood safely in a focused session, stop and split or hand off.
- Extra scaffolding must earn its cost by reducing downstream context.
Scope Control
Stay inside the approved work.
Do not add unrelated refactors, features, dependency changes, visual redesigns, speculative abstractions, or later roadmap work.
When necessary work falls outside scope, stop and ask.
Prototypes
Use debate-me to decide whether a prototype is useful.
Use prototype to test one important assumption at a time.
Prototype output is evidence, not production code, until the human explicitly approves adoption.
Review
Review against the Feature and the actual diff.
Prioritize:
- correctness
- security and privacy
- regressions and edge cases
- accessibility when relevant
- performance
- maintainability
Verification effort should be proportional to risk.
Report findings; do not manufacture them.
Learning
Explain what was actually implemented.
Use examples, diagrams, or quizzes only when they improve understanding.
Keep Feature learning scoped. Use learn-codebase for broad repository
study.