Skip to content

malishomen/ai-engineering-skills

Repository files navigation

AI Engineering Skills

Production-grade Claude Code skills for AI engineering — operational discipline, bounded execution, verification gates, multi-agent harness patterns.

License: MIT Claude Code Skills PRs Welcome

A growing collection of production-ready Claude Code skills + reference documentation for engineers building complex AI systems. Each skill is self-contained, project-agnostic, and designed to make operational discipline mechanical rather than aspirational.


Why this exists

If you've ever shipped a complex multi-agent system as a solo or small-team engineer, you've probably hit some version of this:

  • A "small fix" became a 3-day debugging session because nobody audited dependencies first.
  • A canonical claim drifted from reality because nobody updated the live-state file.
  • A funding application included a conflict-of-interest reference because nobody had a qualification framework.
  • A production deploy ate the database because nobody documented the rule "never docker compose down."
  • A session started cold because nobody captured the carry-over from yesterday.

These failures aren't moral failures. They're structural — the absence of external scaffolding that a real team would provide via review redundancy.

This repo packages the scaffolding. As Claude Code skills that load on narrow triggers, your AI agent and you share the same operational rules. As reference documentation that survives forks, the rules outlast any one project.


Skills

Skill Status Description
complex-project-discipline ✅ v1.0 Universal operational discipline for solo or small-team execution on complex multi-layer projects. Truth hierarchy, bounded execution, verification gates, append-only logs, AI-agent harness, funding readiness, UX standards, anti-patterns. 28 files, ~5,300 lines.
(more coming) 🚧 Multi-agent design patterns, evaluation harness, deployment automation.

The seven core principles (TLDR)

These cut across every skill. Read once; refer back often.

1. 🏛️ Truth hierarchy is non-negotiable

Single canonical live-state file. Lower levels never override higher. No artifact → UNVERIFIED tag.

2. 🎯 Bounded execution beats heroic execution

One task = one reason. One commit = one logical change. Audit before edit.

3. 📐 No artifact, no claim

PARTIAL is honest. UNVERIFIED is required. SUCCESS requires the full state-sync bundle to close.

4. 📜 Append-only logs are the spine

Old entries never edited. New entries appended. The discipline forces honesty in retrospect.

5. 🛡️ Disclosure-first integrity beats hidden conflicts

Hidden conflict-of-interest is catastrophic. Disclosed conflict reads as integrity signal.

6. ⚠️ Preserved failures are a competitive advantage

Document FAIL canonically. The cleanest empirical demonstration of where your method's boundary actually is.

7. 🔒 Main branch is sacred until explicit unlock

Working branch for active work. Main protected via permissions.deny. Unlock requires explicit user command.


Quick start

Use as Claude Code skills

# Clone into your Claude Code skills directory
cd ~/.claude/skills/  # or your project's .claude/skills/
git clone https://github.com/malishomen/ai-engineering-skills.git

The skill auto-loads on narrow triggers (see each skill's SKILL.md frontmatter description). It does NOT auto-fire on every conversation.

Use as reference documentation

Read skills/complex-project-discipline/README.md for the full skill overview, then drill into specific docs as needed.

Recommended first-pass reading order:

  1. SKILL.md — entry + index
  2. docs/01_TRUTH_HIERARCHY_AND_STATE_FILES.md
  3. docs/02_BOUNDED_EXECUTION.md
  4. docs/05_APPEND_ONLY_LOGS.md
  5. docs/15_ANTI_PATTERNS.md — 41 catalogued anti-patterns

Use as a fork starting point

Fork into your project. Adapt funder names, infra hostnames, branch conventions. Don't modify the canonical version — preserve the upstream so your fork's divergences are auditable.


What's inside complex-project-discipline

SKILL.md                                          Entry + index
docs/
  01 Truth hierarchy and state files              9-level source-of-truth model
  02 Bounded execution                            Task pack format + 10-step flow
  03 Verification discipline                      "No artifact, no claim"
  04 Branch and push discipline                   Main protection via mechanical denylist
  05 Append-only logs                             Pipe-separated rows + ## sections
  06 Change class classification                  R0 / R1 / R2 — when in doubt, classify higher
  07 UX standards canonical                       10 international standards distilled
  08 Production infra discipline                  Deploy + DB volumes + tunnel
  09 Corpus-driven development                    7-step pattern + DoD thresholds
  10 Funding readiness discipline                 Owner-operated + reference qualification
  11 Communication patterns                       Apology + opt-in paths
  12 Claude Code harness pattern                  8-step layered rollout
  13 Memory structure                             Repository + state + top-of-stack
  14 Session close discipline                     Close gate + 6-condition final answer
  15 Anti-patterns catalogue                      41 named AP-NN with counter-patterns
  16 Polish discipline                            Adversarial re-read

templates/
  task_pack.md, bounded_packet.json,
  agents_log_row.md, weekly_status_section.md,
  reference_prep_brief.md, apology_email_pattern.md,
  paste_ready_field_map.md, claude_settings.json

checklists/
  pre_commit_check.md, pre_push_check.md,
  session_close_check.md, final_answer_gate.md,
  ux_per_surface_check.md, deploy_verification_check.md

Provenance

Distilled from a 2026 production project running, in parallel:

  • A research line with measured gates and preserved failures.
  • A production web service with single-node Docker Compose + Cloudflare Tunnel ingress.
  • A funding-readiness application cycle across the alignment-funder ecosystem.
  • A layered Claude Code harness (settings.json + SessionStart hook + corpus runner + sub-agent).

All identifying information about the source project (organization names, partner relationships, specific gates, contact details, server addresses, SHA hashes) has been stripped from this canonical version. The patterns remain; the specifics do not.


Contributing

PRs welcome for:

  • Universal-pattern improvements distilled from your own project experience.
  • New anti-patterns observed in real solo / small-team execution (see docs/15_ANTI_PATTERNS.md format).
  • Per-funder profile updates in docs/10_FUNDING_READINESS_DISCIPLINE.md (alignment-funder ecosystem evolves; per-funder notes age).
  • New skills in the skills/ directory, following the same structure (SKILL.md + docs/ + templates/ + checklists/).

PRs not welcome for:

  • ❌ Adding project-specific identifiers (the canonical version is intentionally generic).
  • ❌ Soft-pedalling the integrity discipline — preserved failures, append-only logs, disclosure-first rules are load-bearing.
  • ❌ Adding dependencies on tooling beyond Claude Code, git, Docker.

See CONTRIBUTING.md for details.


Related work

This collection complements but does not replace:


Roadmap

Skills planned for future inclusion (alphabetical, not prioritized):

  • claude-multi-agent-design — patterns for designing multi-agent assemblies with explicit capability envelopes.
  • corpus-driven-evaluation — extending the corpus-driven pattern from dev to evaluation harness.
  • deployment-automation — universal CI/CD patterns for solo founders.
  • frontend-state-coverage — corpus-encoded coverage of state-conditional UI rendering.
  • funding-application-craft — extending the paste-ready field map pattern to additional funder types.

If you'd like to author one, see CONTRIBUTING.md.


Star ⭐ if useful

Stars help other engineers find this. If any pattern in here saved you a debugging session or a botched deploy, a star is the cheapest possible thank-you.

If you fork, tag me so I can see how others adapt the patterns.


License

MIT License — free to use, modify, redistribute. No attribution required, but if you cite this work somewhere I can see, I'll learn from how it gets used.


One-line summary

External structure for solo / small-team AI engineering. Truth hierarchy, bounded execution, append-only logs, verification discipline, disclosure-first integrity, preserved failures, mechanical main-branch protection. Make the agent and the operator share the same rules.

About

Claude Code skills for production-grade AI engineering — operational discipline, bounded execution, verification gates, multi-agent harness patterns.

Topics

Resources

License

Contributing

Security policy

Stars

Watchers

Forks

Packages

 
 
 

Contributors