Entry point for any AI agent: OpenCode, Claude Code, Cursor, GitHub Copilot, or any LLM-based tool. Load this file first. Follow the mode that matches your current task.
aidd/memory/ # Persistent project context — read at session start
aidd/work/ # Runtime artifacts — AUDIT, INTAKE, PLAN, REVIEW, SUMMARY, HANDOFF
aidd/review/ # Domain-specific review checklists (web3, ml, general)
rules/ # Plain Markdown rules — load relevant files per stack
rules/INDEX.md # Rules catalog with always-apply and deprecated flags
prompts/ # Workflow prompts — use as instructions per phase
scripts/ # Validation gate scripts — mandatory checkpoints
docs/workflow/README.md # Complete workflow reference
docs/design/architecture.md # Export guide, architecture decisions
When: Starting a new task, feature, or fix. Before writing any code.
Steps:
- Read
aidd/memory/activeContext.mdandaidd/memory/projectbrief.md - Run audit if repository is non-trivial and no
aidd/work/AUDIT.mdexists - Capture requirements using
prompts/intake.md→ producesaidd/work/INTAKE.md - Produce technical plan using
prompts/plan.md→ producesaidd/work/PLAN.md - Gate:
bash scripts/validate-plan.sh— do not proceed if it fails
When: aidd/work/PLAN.md exists and has passed validate-plan.sh.
Steps:
- Load always-apply rules from
rules/INDEX.md - Load only stack/task-relevant rules (just-in-time loading)
- Use
bash scripts/c7-docs.sh ...before coding against external libraries - Implement following
aidd/work/PLAN.mdsteps exactly — no scope creep - Run:
bash scripts/aidd-check.sh - If UI changes:
bash scripts/aidd-verify-ui.sh - Update
aidd/memory/activeContext.mdwith completed work and next steps
When: Implementation is complete and aidd-check.sh passes.
Steps:
- Load checklist based on domain:
web3→aidd/review/review-checklist-web3.mdml→aidd/review/review-checklist-ml.mdmixed→ both checklistsotheror unknown →aidd/review/review-checklist-general.md
- Produce review using
prompts/review.md→aidd/work/REVIEW.md - Gate:
bash scripts/review-check.sh— task is not done until Verdict isAPPROVE - Optional closeout:
bash scripts/aidd-finish.sh(guided commit/push/PR + cleanup + handoff)
| Gate | Script | Blocks |
|---|---|---|
| Plan validation | bash scripts/validate-plan.sh |
Implementation (ARCHITECT → EDITOR) |
| Implementation check | bash scripts/aidd-check.sh |
Review (EDITOR → REVIEWER) |
| Review verdict | bash scripts/review-check.sh |
Done (REVIEWER → complete) |
Definition of Done: PLAN validated + implementation complete + activeContext.md updated + REVIEW verdict is APPROVE.
- GitHub signals: use
bash scripts/gh-context.sh ...(JSON output) - Library docs: use
bash scripts/c7-docs.sh --library <id> [--topic <topic>] - Context7 usage is mandatory and enforced by
aidd-check.sh
- Use
rtkwrappers for shell operations when supported. - Use
CONTEXT_BUDGET(low|medium|high, defaultlow) to control context loading. - Keep
aidd/work/SUMMARY.mdupdated as the short in-cycle digest. - Generate
aidd/work/HANDOFF.mdat cycle closeout for fast restart. - Prefer findings-first concise outputs; avoid repeating unchanged context.
Use obsidian_alyra_* tools only for complex or ambiguous blockchain / AI concepts requiring confirmation of definitions, formulas, or nuanced technical details.
Protocol:
obsidian_alyra_search_notes— short query, limit 3–5 resultsobsidian_alyra_read_note— read at most 1 note (the most relevant)- Summarize findings in your own words — do not paste large note blocks
Never use MCP Obsidian for general coding tasks. Never attempt to write to the vault.