Agentic Engineering Playbook

A me-first operating system for keeping top-tier coding agents bounded, useful, and trusted.

Optimized for a solo engineer running multiple agent sessions in parallel, where the dominant pain is smart-model overengineering: models that are capable enough to solve the problem, but too unconstrained to stop at the simple solution.

This starts as a personal workflow. Team sharing comes later, after the rules have earned their keep in real work.

Read Order

1. 00-quick-start — day-1 setup checklist
2. 01-workflow-loop — /work default, lower-level gates, review, merge
3. 02-context-engineering — layered context: AGENTS.md/CLAUDE.md, ADRs, golden examples
4. 03-spec-format — Given/When/Then + constraints + examples
5. 04-task-decomposition — tracer bullets, 1-3hr slices
6. 05-quality-gates — preferred Biome, Vitest, Commitlint, Lefthook, pnpm, and CI setup
7. 06-tooling-stack — skills, subagents, hooks, CLIs, and local settings
8. 07-parallelization — git worktrees, 2-3 concurrent slots, conflict mgmt
9. 08-trust-and-verification — gates, self-QA, recoverability
10. 09-anti-overengineering — canonical constraints list, drop-in for progressive-disclosure agent docs
11. 10-greenfield-bootstrap — hand-shape 2-3 reference features before agents take over
12. 11-isolated-db-branches — Neon branch databases for parallel Drizzle schema work
13. 12-metrics-and-telemetry — local success metrics, report sidecars, telemetry stance
14. 13-sharing-with-team — onboarding ramp, adoption pattern
15. 14-philosophy — trust, bounded autonomy, and the five laws

Drop-In Artifacts

• templates/ — AGENTS/CLAUDE, progressive agent docs, spec, ADR, PRD, PR body, metrics, quality-gate snippets
• subagents/ — planner, anti-overeng-reviewer, decomposer, pr-author
• skills/ — installable skill directories for /work, /setup-quality, /isolated-db-branches, and the supporting workflow commands
• hooks/ — stop-hook checks, dangerous-command guard, pre-tool safety examples
• scripts/ — local report-sidecar summarizers

Install Skills

Install a skill with:

npx skills add wiseiodev/agentic-playbook/skills/<skill>

Common starting set:

npx skills add wiseiodev/agentic-playbook/skills/work
npx skills add wiseiodev/agentic-playbook/skills/setup-quality
npx skills add wiseiodev/agentic-playbook/skills/isolated-db-branches

See skills/README.md for the full skill catalog and install commands.

Give This To Your Agent

Use this prompt to incorporate the playbook into a repo:

Please incorporate wiseiodev/agentic-playbook into this repo.

Read the playbook README first, then adapt the artifacts to this repo instead of copying blindly.

Do the following:
- Copy or adapt templates/agent-docs into AGENTS.md/CLAUDE.md and docs/agent/.
- Install only the relevant skills from skills/ using `npx skills add wiseiodev/agentic-playbook/skills/<skill>`.
- Set up or tighten quality gates using the repo's actual package manager, scripts, task runner, and CI shape.
- Add hooks only when they fit this repo's workflow and are not surprising.
- Preserve existing conventions unless the playbook explicitly fixes a gap.
- Report exactly what changed, which gates passed, and which recommendations were intentionally skipped.

Do not add a custom installer, dashboard, or extra process docs.

Core Thesis

Quality at speed comes from upstream constraints, not downstream rework. Smart models with high thinking overengineer in inverse proportion to how constrained the task is. The cure is: behavior specs + small slices + golden examples + full ceremony by default + anti-overeng review.

Use the best model. Bound its cleverness.

For agent-owned implementation, /work is the default operating loop. Invoking /work authorizes the agent to branch, implement, run gates, self-QA, review, write the report and metrics sidecar, commit, push, and open a Ready PR after the required checks pass. Humans can waive gates. Agents cannot self-waive.

Success is measured from the audit trail the workflow already creates: reports, metrics sidecars, PRs, reviews, and follow-up fixes. Do not build a dashboard until the local JSON proves the metrics are worth keeping.