intelligence-dev-packs

June 23, 2026 · View on GitHub

Shared AI-first engineering practices, packaged as portable rules, agents, and skills. Author the workflow once, reuse it in every repository, and let both your team and your AI coding agents read the same source.

Built and maintained by Ainova Systems. Extracted from running multiple AI-coded enterprise systems in daily production, where AI writes the majority of the code.

Packs and domains

Two decoupled ideas (see packs/README.md):

  • A pack is what you install together. A domain prefix (dev-/git-/spec-) is the namespace on each artifact.
PackDomainsScopeDepends on
coredev-, git-Universal - engineering discipline, session hygiene, git/PR/release. Install everywhere.-
specspec-Opt-in - spec-driven development. Not every project adopts it, so it ships separately.core

The owner flow (spec pack)

When a project adopts spec-driven development, the owner touches the work at three gates, and everything between runs autonomously with reviews, checks, and fallbacks built in.

StepWhoSkill
1. State the taskOwner-
2. Spec written (requirements, plan, tasks)AIspec-create
3. Gate 1 - review and approve the specOwnerspec-approve
4. Implementation: branch, parallel subagents, tests, milestone commitsAIspec-execute
5. PR opened, CI driven to green, review comments handled, outcome labelAIgit-finalize-pr
6. Gate 2 - accept the PROwner-
7. Squash-merge, close the spec, cleanupAIgit-merge-pr -> spec-close

The spec lifecycle is machine-tracked and AI-driven end to end - the owner's only touchpoints are the three gates. A spec carries a status: spec-create writes proposed; the owner reviews and runs spec-approve (approved); spec-execute runs it (in-progress) to a PR; after the owner accepts, git-merge-pr merges and spec-close finalizes it (completed). spec-cancel retires a dropped or superseded spec with a recorded reason. Every transition keeps the spec status and the docs substrate (feature docs, rules, model, glossary) in sync automatically - nothing is updated by hand.

Fully autonomous mode: spec-approve fills a queue the owner reviewed once, and spec-execute-next drains it - it picks the highest-value approved spec (or finishes an in-progress one), drives it to an outcome-labeled PR, closes any merged spec, and resets the workspace. Suitable for scheduled and looped runs. Every autonomous run ends with exactly one outcome label: ai:ready-to-merge, ai:manual, or ai:failed.

Before the first spec, spec-init is a one-time bootstrap: it scaffolds the in-repo docs substrate (domain model, glossary, rules-as-contracts, feature baselines, decisions, the specs/ tree) and safely migrates existing documentation into it, drafting from code for owner review. It learns and adopts a project's existing docs structure rather than imposing one.

A project that has not adopted spec-driven development still uses the core pack on its own: the git/PR/review/release skills and the engineering-discipline rules work standalone.

core pack

dev- domain - engineering discipline and session hygiene

ArtifactKindRole
dev-skill-firstruleCheck the skill catalog before improvising a workflow
dev-context-engineeringruleConventions, decisions, and domain knowledge live in the repo
dev-verification-gatesruleTypecheck, lint, tests pass before every commit; gates never weakened
dev-rollback-safetyruleReversible migrations, feature flags, expand-contract sequencing
dev-code-revieweragentReviews changes for correctness, conventions, tests, security (read-only)
dev-test-engineeragentTest strategy and coverage across all levels
dev-run-testsskillTypecheck, lint, tests with scope detection and failure analysis
dev-review-changesskillRead-only diff review with severity verdict
dev-handoffskillSelf-contained continuation prompt for a fresh session

git- domain - git / PR / release

ArtifactKindRole
git-commit-conventionsruleCommit message format, push discipline, forbidden trailers
git-workflowruleBranch model, protected branches, feature-branch flow
git-commit-pushskillVerified milestone commit and fast-forward push
git-open-prskillOpen a PR for the branch, using the repo template and profile knobs
git-resolve-conflictsskillSemantic conflict resolution, full gates after
git-review-pr-commentsskillTriage reviewer feedback: fix, discuss, or decline with reason
git-finalize-prskillCI to green plus every review comment handled - PR ready to merge
git-merge-prskillAfter owner accept: guard-checked squash-merge, base sync, cleanup
git-create-releaseskillVersion, changelog, tag per the project's release flow
git-scan-secretsskillCredential scan over diff, tree, or history

spec pack

Domain spec-. Depends on core.

ArtifactKindRole
spec-disciplineruleWhen a change needs a spec, and the docs chain
spec-orchestrationruleMulti-agent doctrine: consistency, delegation by pointers, outcome labels
spec-architectagentAuthors specs, ADRs, module boundaries
spec-docs-writeragentDocumentation and decision records in sync with code
spec-initskillOne-time bootstrap: scaffold the in-repo docs substrate and migrate existing docs into it, drafting core docs from code
spec-createskillTask to spec (or update one): grills the owner on ambiguity, writes requirements + plan + tasks, status: proposed
spec-approveskillRecord the owner's gate-1 decision: status: approved, entering the autonomous queue
spec-executeskillApproved spec to PR: parallel subagents, consistency gates, docs reconciliation
spec-continueskillResume a mid-flight spec: inherited-work drift audit, then execution
spec-execute-nextskillDrain the approved queue: pick the highest-value spec, drive it end-to-end, close merged specs, reset
spec-closeskillPost-merge: status: completed, confirm docs reconciled, archive per convention
spec-cancelskillRetire a spec (cancelled / superseded) with a recorded reason; reconcile its draft docs
spec-documentskillWrite/update one docs artifact (feature, rule, glossary, model, architecture) from code
spec-decisionskillNumbered ADR behind a three-condition gate
spec-audit-docsskillDocs claims audited against code: drift vs violation

Quick start (paste prompt)

Paste this into your AI coding agent, working in your project. It sets up the packs via intelligence-sync remote sources, adapts them to your repo, and reports - without committing.

Set up the intelligence-dev-packs shared engineering packs in THIS project, via
intelligence-sync remote sources. Do it safely and report; do not commit or push.

1. PREREQUISITE - intelligence-sync. Find the intelligence umbrella: the directory holding
   `config.yaml` and a `sync/` engine (usually `intelligence/`, but detect it - it may be
   named or capitalized differently). If none exists, this project has no intelligence-sync:
   STOP and tell the owner to install it first from
   https://github.com/ainova-systems/intelligence-sync (paste its README Quick Start prompt),
   then re-run this. Do not proceed without it. If it exists, confirm the engine supports
   remote `git+` sources by grepping `<umbrella>/sync/scripts/lib/common.sh` for
   `resolve_source_dir`; if missing, update the engine first via `<umbrella>/sync/scripts/update.sh`.

2. INSPECT. Read the umbrella `config.yaml`, the project's rules, and any existing
   documentation conventions. Note the branch model, build/test/lint commands, PR platform,
   and any existing rule that overlaps a pack rule.

3. ADD THE PACKS (remote sources - no submodule, no copy). Append to the existing `sources`:
   - core (always - universal git/PR/review/discipline):
     - rules:  "git+https://github.com/ainova-systems/intelligence-dev-packs.git@main#packs/core/rules"
     - agents: "git+https://github.com/ainova-systems/intelligence-dev-packs.git@main#packs/core/agents"
     - skills: "git+https://github.com/ainova-systems/intelligence-dev-packs.git@main#packs/core/skills"
   - spec (ONLY if this project wants spec-driven development AND has no existing docs/spec
     workflow - the spec pack assumes the ai-first-docs tree, docs/specs/ + model.md +
     glossary.md, and would conflict with an existing one): add the same three lines with
     `#packs/spec/...`. If the project already has its own docs/spec approach, SKIP spec and say so.
   Use `@main` for now, or pin to a tag/SHA for reproducibility.

4. GENERATE THE PROFILE so skills adapt without re-scanning each run. Read the schema (raw:
   https://raw.githubusercontent.com/ainova-systems/intelligence-dev-packs/main/packs/core/templates/dev-project-profile.md),
   fill it from this repo (branch model, branch prefixes, typecheck/lint/test commands, PR
   platform and merge method, release flow, and - if spec was added - the docs structure), and
   save it as `<umbrella>/rules/dev-project-profile.md`. Leave any value you can't detect for the owner.

5. FLAG CONFLICTS, don't resolve silently. For each pack rule, report whether the project
   already has a rule that overlaps or contradicts it. The project's own rule wins on conflict -
   recommend keep / drop / scope; the owner decides.

6. SYNC AND REPORT. Run `bash <umbrella>/sync/scripts/sync.sh`. Report: sources added, whether
   spec was included and why, the generated profile values, every rule conflict + recommendation,
   and which pack skills are now available. Do NOT commit or push - the owner reviews first.

The manual equivalents (and submodule / global / copy alternatives) are below.

Install

For an intelligence-sync that supports git+ remote sources, no submodule or copy is needed: point intelligence/config.yaml straight at the packs. The #subpath selects the pack and type; the @ref pins the version. Add core always; add spec only for spec-driven projects.

sources:
  rules:
    - "intelligence/rules"
    - "git+https://github.com/ainova-systems/intelligence-dev-packs.git@v0.1.0#packs/core/rules"
    - "git+https://github.com/ainova-systems/intelligence-dev-packs.git@v0.1.0#packs/spec/rules"
  agents:
    - "intelligence/agents"
    - "git+https://github.com/ainova-systems/intelligence-dev-packs.git@v0.1.0#packs/core/agents"
    - "git+https://github.com/ainova-systems/intelligence-dev-packs.git@v0.1.0#packs/spec/agents"
  skills:
    - "intelligence/skills"
    - "intelligence/sync/skills"
    - "git+https://github.com/ainova-systems/intelligence-dev-packs.git@v0.1.0#packs/core/skills"
    - "git+https://github.com/ainova-systems/intelligence-dev-packs.git@v0.1.0#packs/spec/skills"

Then bash intelligence/sync/scripts/sync.sh. The engine shallow-clones each remote spec fresh every run (the same repo@ref is cloned once even across the six entries above), reads the #subpath directory, and feeds it through the same pipeline as a local source. Pin the @ref to a tag or SHA - an unpinned ref tracks the default branch and changes under you; the ref must be slashless (tag, SHA, or slashless branch). A clone failure (offline, bad URL) warns and skips just that source; the rest of the sync still succeeds. URL rules and the full reference are in docs/INTEGRATION.md.

Mode B - git submodule (vendored)

For offline / air-gapped CI or teams that want the packs checked into their tree. Works with intelligence-sync 0.3.1 or later.

git submodule add https://github.com/ainova-systems/intelligence-dev-packs intelligence/dev-packs
sources:
  rules:
    - "intelligence/rules"
    - "intelligence/dev-packs/packs/core/rules"
    - "intelligence/dev-packs/packs/spec/rules"
  agents:
    - "intelligence/agents"
    - "intelligence/dev-packs/packs/core/agents"
    - "intelligence/dev-packs/packs/spec/agents"
  skills:
    - "intelligence/skills"
    - "intelligence/sync/skills"
    - "intelligence/dev-packs/packs/core/skills"
    - "intelligence/dev-packs/packs/spec/skills"

submodules:
  - "intelligence/dev-packs"

Then bash intelligence/sync/scripts/sync.sh. Update with git submodule update --remote intelligence/dev-packs.

Mode C - global skills (Claude Code)

git clone https://github.com/ainova-systems/intelligence-dev-packs
bash intelligence-dev-packs/scripts/claude-install-global.sh          # every pack
bash intelligence-dev-packs/scripts/claude-install-global.sh core     # core only

Installs the skills for your user, available in every project. Without a project profile, skills fall back to auto-detection and ask when a value is ambiguous.

Mode D - plain copy

cp -r intelligence-dev-packs/packs/core/rules/*  my-project/intelligence/rules/
cp -r intelligence-dev-packs/packs/core/agents/* my-project/intelligence/agents/
cp -r intelligence-dev-packs/packs/core/skills/* my-project/intelligence/skills/
# add packs/spec/* the same way for spec-driven projects

See docs/INTEGRATION.md for the full reference, including uninstall.

Configure for your project

Nothing is wired up by hand. The only manual step is the sources: entries you already added (Mode A/B) plus sync; every skill then adapts to the repo with zero further config:

  1. Auto-detect (default). Skills read the repo - default branch from git, an origin/develop as the integration branch, typecheck/lint/test commands from the manifests, PR platform from the remote - and ask once only when something is genuinely ambiguous.

  2. Optional profile, generated for you (not copied). To pin those answers so nothing is re-detected or re-asked, have your AI agent write the profile once. It saves a filled dev-project-profile.md into your rules source, where it rides as an always-on rule the agent reads each task. Paste:

    Read packs/core/templates/dev-project-profile.md from the intelligence-dev-packs source as the schema, inspect this repository, and write a filled-in dev-project-profile.md into my intelligence rules source. Change nothing else.

The profile (when present) declares the branch model, verification commands, PR platform and merge method, release flow, and the docs structure (specs_dir, features_dir, rules_dir, decisions_dir, spec_grouping). The greenfield documentation default follows the ai-first-docs tree: feature docs as the behavior baseline, change specs as docs/specs/NNN-<slug>/ (requirements.md EARS, plan.md, tasks.md), business rules as contracts, numbered ADRs.

Design principles

  1. Context engineering over prompt engineering. Output quality is governed by the information environment in the repository, not by prompt tricks.
  2. Maximum context in the repository. Conventions, decisions, and domain knowledge are durable repo artifacts that humans and agents both read.
  3. Skill-first. Repeatable work starts from a skill, so output stops depending on who prompted it, and capability compounds with the catalog.
  4. Spec discipline at the right threshold. Small clear tasks ship directly; boundary-crossing work gets requirements + plan + tasks first.
  5. Consistency over creativity in execution. The spec is the contract; subagents receive pointers to the same sources, never re-explained instructions.
  6. Verifiability and rollback over trust. Strict gates catch AI errors before production; reversible changes make the remaining errors cheap.

Compatibility

  • intelligence-sync: remote git+ sources (Mode A) need a build with that feature (it lands after 0.4.2 - check the engine's CHANGELOG for "Remote git sources"); submodule / copy (Modes B-D) work with 0.3.1 or later. The packs also work without the sync engine, consumed directly by Claude Code or any tool that reads SKILL.md folders.
  • Tools: anything intelligence-sync targets (Claude Code, Cursor, GitHub Copilot, Codex, Pi, opencode, AGENTS.md readers), plus direct use.
  • Naming: every artifact carries a domain prefix (dev-/git-/spec-), so pack content never collides with project artifacts or the reserved intelligence- meta-skills.

Versioning

Semantic versioning, history in CHANGELOG.md. Pin the @ref (Mode A) or the submodule (Mode B) to a tag for reproducible setups; track main for the latest practices.

  • intelligence-sync - the open-source engine that delivers these packs into every IDE your team uses (Claude Code, Cursor, GitHub Copilot, Codex, and more). intelligence-dev-packs is the shared content. intelligence-sync is how it reaches each tool.

License

MIT. Copyright 2026 Ainova Systems.


Created by Dmitrij Zykovic - Fractional CTO at Ainova Systems.

Helping engineering teams adopt AI-First SDLC and build autonomous AI engineering pipelines.

LinkedIn · Advisory & Consulting