⚡ AIDD — AI-Driven Development
Stop AI coding agents from grading their own homework.
A methodology with technical teeth for AI coding agents (Claude Code, OpenCode, Codex, Gemini CLI). It forces agents to search before creating, plan out loud before coding, and blocks self-approval via automated hooks — not another AI coding agent itself, a set of habits your existing one follows.
🚀 10-Second Quick Start
Get started in any existing repository without modifying your code:
# 1. Install the CLI
pip install aidd-cli
# 2. Initialize in your project
aidd init .
# 3. Check for architectural violations & spec gaps
aidd check-charter .
aidd search "auth flow"
Using Claude Code? Install the technical enforcement gates with one command:
cp -r skill ~/.claude/skills/aidd && python ~/.claude/skills/aidd/scripts/install_hooks.py
⚖️ Why AIDD vs. traditional prompts / spec files
| Feature | Traditional rules (CLAUDE.md, .cursorrules) |
AIDD methodology |
|---|---|---|
| Compliance | Optional (LLMs tend to skip long instructions) | Strict / blocking (real system hooks) |
| Code review | The same agent grades its own code | Independent subagent audit, mandatory |
| UI mapping | Duplicated, per-screen prose | Unified codes (COMP-nnn, CTL-nnn) |
| Task tracking | Disconnected markdown nobody re-checks | Bidirectional sync with GitHub / Azure DevOps / Bitbucket |
| Context cost | Dumps thousands of tokens into every prompt | Indexed local lookup (find_spec.py) |
The problem
AI coding agents are fast at writing code and bad at knowing what already exists. Left alone, they duplicate specs nobody asked for, re-explain a flow in three paragraphs instead of drawing it, ship a fix that fabricates success instead of actually calling the real backend, and grade their own homework when asked to review it. AIDD is the set of habits that stops that — enforced by hooks where the tool allows it, followed as a written discipline everywhere else.
What it does
- Searches before creating. Every request runs against a compact, auto-built graph of every
existing spec (
specs/index.toon) before anything new gets created — a match means "amend this," not "start over." - Shows the plan before building it. Step 1.5 turns each use case into an interactive
actors × processes flow with generated pseudocode (
aidd flow specs/001-login/visual-flow.toon --open) — the user clicks through it and corrects the flow by node, not in prose. Source is a compact TOON file; layout, routing and pseudocode are derived, never hand-drawn. - Plans out loud. Ambiguity gets resolved as an explicit question before code is written, not discovered after a review.
- Ships in small, reviewable pieces. One code, one file, one PR — a wrong pass costs one edit, not a rewrite.
- Never lets the implementer grade its own work. Step 6's review must come from a separate
agent dispatch, one per domain the change touched. On Claude Code this is an actual technical
gate (
hooks/rule_gate.py+stop_gate.py), not just a rule stated in a doc. - Enforces the rules instead of asking nicely. Hard rules R1–R12: estimates are agent
minutes per wave (human hours only as a reference), skipping a step needs the user's quoted
confirmation (verified against what the user typed or answered), the agent can't answer its own
Align questions (blank answers block planning,
repo —sources must exist and cite a line), plan/tasks can't be written out of order,tasks.mdapproval is the user's recorded "Approve" answer plus a tamper-evident hash, a spec stays open untilaidd rules close/abandon(touching another spec can't switch the gates off), code edits are blocked for every open spec without a recorded approval, every required domain needs its own distinct auditor, and a session can't end on built code without its closing audits (R8 relaxes after 3 blocks); a ✅ screen/API/field needs executed evidence, not a code reading (R10, freshness best-effort), a repeated bug needs a root cause and pattern sweep (R11), and a task that reuses a screen/component must say whether it is a new view or logic only (R12). Hooks write an append-only evidence log (guarded against the agent by R9: the Write/Edit tools and a lexical shell guard, not a sandbox) (per-session in the temp dir, plus a project log only where aspecs/or.aidd/root exists) that the gates read;aidd statuslists every open spec with the ledger and WHY blocked, andaidd rules checkprintsPASS|FAIL Rn … → fix. Honest limits: Claude Code only; Bash and PowerShell can still edit code and, through the CLI or libraries, try to forge evidence (the guard is lexical, not a sandbox); the gates check that a subagent ran and a question was answered, not that they were good; the user's click cannot be proven; hooks fail open on a crash, launch failure or timeout (recorded ashook_errorwhen possible); session ids/resume behaviour are not verified. The rules stop accidental and self-justified skipping, not a determined agent.AIDD_RULES=off|0|false|no|warnis the owner's escape hatch. - Remembers the why. Decisions, rejected options, bug root causes and constraints go into
.aidd/memory/as curated, code-anchored AIDD-TOON entries, committed with the code — no daemon, no LLM observer, no vectors, stdlib only. Search is progressive:aidd mem searchreturns one line per hit,aidd mem show <id>the full row. On Claude Codememory_context.pysurfaces the relevant entries at session start (installed by default; the per-filememory_file_context.pyis opt-in).STATE.mdstays "where am I",specs/index.toonstays "what exists"; memory is "why". A one-off, read-only importer can seed it from a claude-mem database (secrets and IPs scrubbed). Memory is committed content that gets injected into sessions: review it like code. - Works the same across tools. One methodology (
skill/AIDD.md), thin adapters per tool — no relearning the process when the assistant changes. For Claude Code, OpenCode, and Codex, that's a native skill/AGENTS.md pointer; for Gemini CLI, Cursor, Windsurf, Cline, and GitHub Copilot,generate_adapters.pyrenders each one's own native command-file format straight fromcommands/aidd-*.md— one source, no per-agent copy to keep in sync. Both the issue-tracker providers and these adapter targets are auto-discovered extensions (seedocs/EXTENDING.md), so a project can add its own without forking this repo. - Has a charter that's actually checked. Project-wide rules split into judgment calls
(prose) and checkable ones (a pattern + glob table
check_charter.pyruns for real) — not a principles doc trusted to memory. Run against a real production project, it caught two stale claims in that project's own hand-written notes about specific libraries being unused (both false, verified in seconds) and a third that turned out far worse than self-rated (340 real violations of a rule logged as low-severity). - Verifies its own map before planning against it. The spec graph is parsed mechanically
(fast, cheap), which also means it can misparse a renamed code or a stale relationship — so a
rebuild dispatches an independent Graph Coherence Auditor before
plan.md/tasks.mdbuild on the new edges. Enforced by a hook pair, not a step an agent could skip under time pressure. - Syncs to GitHub, Azure DevOps, or Bitbucket — not just markdown. An approved
tasks.mdbecomes real, trackable issues viatasks_to_issues.py --provider {github,azure_devops, bitbucket}— dry-run by default, never duplicates on re-run. It stays in sync from there:aidd tracker syncdiffs (or writes, with--apply)tasks.md's Status column against live tracker status, andaidd tracker link-prattaches an opened PR's URL back to its task's tracker issue andtasks.mdrow. CI templates for both trackers ship underskill/templates/ci/, installed withaidd ci install {github,azure-devops}. - Has a package-install mechanism, kept to the load-bearing parts — not yet a third-party
ecosystem.
catalog/is the marketplace mechanism: packages (templates, agent adapters, issue-tracker providers, hooks) install into your project withaidd marketplace install <id>, content-hashed soremovenever discards a file you hand-edited since. Aprovider/adapter/hookpackage lands under.aidd/extensions/<id>/and is auto-discovered from there — no code change to this repo needed to use it. Structural validation runs in CI on every submission; a maintainer still reviews and merges each one by hand — seecatalog/README.md. Today that catalog holds exactly one package, first-party (hotfix-report); zero third-party submissions have been merged yet — the mechanism is built and tested, the ecosystem it's built for doesn't exist yet. Manage what's discovered (first-party or project-local) withaidd extensions list|info <id>|enable <id>|disable <id>— seedocs/EXTENDING.md.
What's in this repo
skill/— the Claude Code skill:SKILL.md, enforcement hooks (hooks/: ten by default, centred on the singlerule_gate.py), and every script/template (scripts/,templates/).skill/AIDD.mdis the tool-agnostic methodology core every other adapter — including this repo's own CLI — points back to. Includes AIDD Memory (scripts/aidd_memory.py,aidd_memory_import.py; hooksmemory_context.py,memory_file_context.py, opt-in viainstall_hooks.py --with-memory-file-hook).aidd/— the CLI (pip install -e .→ theaiddcommand): a thin dispatcher overskill/scripts/, for a human at a terminal or a CI job, no AI agent required.commands/— the eight/aidd-*pipeline-stage commands for Claude Code.adapters/— drop-in adapters for other agent tools (OpenCode skill + plugin, anAGENTS.mdsnippet for Codex and others, aGEMINI.mdpointer) plusdot-aidd/, the portable.aidd/bundle (methodology + scripts + templates, no Claude-specific pieces) any project installs once and every adapter reads from. Seeadapters/README.mdfor the install steps.catalog/— the marketplace mechanism:schema.json, two registries (catalog.jsonfirst-party, currently one package;catalog.community.jsonthird-party, currently empty), andpackages/<id>/folders. Installed/searched viaskill/scripts/marketplace.py(oraidd marketplace ...), validated in CI byskill/scripts/validate_catalog_entry.py. Seecatalog/README.mdandcatalog/CONTRIBUTING.md.tests/— the stdlibunittestsuite (zero dependencies,python -m unittest discover -s tests), run on every push/PR by.github/workflows/test.yml(andcatalog-validate.ymlfor anything touchingcatalog/).
Install (Claude Code)
cp -r skill ~/.claude/skills/aidd
python ~/.claude/skills/aidd/scripts/install_hooks.py
Install (the CLI — for a terminal or CI, no AI agent needed)
pip install aidd-cli
aidd init /path/to/your/project # installs the .aidd/ bundle there
aidd search "login" # search the spec graph
aidd check specs/001-login/ # mechanical gap-check
aidd check-charter . # run charter.md's checkable rules
aidd status # hard-rules ledger of all open specs + WHY blocked (--json; exit 0 always)
aidd rules check specs/001-login/ # PASS|FAIL Rn message -> fix (exit 1 on a violation)
aidd rules approve specs/001-login/ # write the tasks.md Approved: line (needs the user's recorded "Approve" answer)
aidd rules close 001-login # complete the spec (needs approval + qa-audit + auditors + the user's "Yes, close" answer)
aidd rules abandon 001-login # drop it (needs the user's recorded "Abandon" answer)
aidd flow specs/001-login/visual-flow.toon --open # Flowmap: interactive flow + pseudocode
aidd mem add --type decision --title "Login uses magic links" --why "No password storage" --codes US-001,API-003 --scope 001-login
aidd mem search login # one line per hit: id, date, type, codes, title
aidd mem show m-3fa91c02 # full entry (why, files, supersedes)
aidd mem timeline m-3fa91c02 # neighbouring entries of the same scope, by date
aidd mem file skill/scripts/flowmap.py # entries that mention a file
aidd mem inject # the digest the SessionStart hook injects (also: inject --file PATH)
aidd mem stats # counts by type/scope, bytes
aidd mem compact --before 2026-07-01 --apply # archive superseded/old entries (dry run without --apply)
aidd mem import-claude-mem <claude-mem.db> --project myproj --since 2026-01-01 --dry-run # one-off, read-only import (also --scope, --include, --no-summaries)
# AIDD_MEMORY_DIR overrides .aidd/memory/; `aidd mem --root X <cmd>` works before or after the command; `add --date` backdates
aidd tasks-to-issues specs/001-login/tasks.md --apply # sync tasks (github by default)
aidd tasks-to-issues specs/001-login/tasks.md --provider azure_devops --org ... --project ... --apply
aidd adapters generate all /path/to/your/project # native commands for Gemini CLI, Cursor, Windsurf, Cline, Copilot
aidd marketplace list # browse installable third-party packages
aidd marketplace install hotfix-report /path/to/your/project
aidd extensions list # discovered provider/adapter/hook extensions
aidd tracker sync specs/001-login/tasks.md --provider github --repo owner/repo # dry-run status diff
aidd tracker link-pr specs/001-login/tasks.md --pr-url ... --branch ... --apply
aidd ci install github /path/to/your/project # install the GitHub Actions CI template
Published on PyPI as aidd-cli. To work on AIDD itself
instead of just using it, clone and install in editable mode:
git clone https://github.com/quimhen/aidd.git && cd aidd
pip install -e .
Install (other tools)
See adapters/README.md.
Documentation
- Why AIDD — what changes for you, in plain language, with the measured numbers.
- Pipeline reference — the full technical breakdown: the code system, the spec graph, every step, the enforcement hooks.
- FAQ — common questions. Have one that's not there? Open an issue or a discussion.
- Extending AIDD — the three extension seams (issue tracker
providers and multi-agent adapter targets, both auto-discovered from
skill/extensions/**/manifest.jsonor.aidd/extensions/**/manifest.json, plus charter checkable rules), and how they relate to the marketplace (catalog/). - Consulting — if your team wants help adopting AIDD.
Status
Early — the core pipeline, the search/graph tools, the charter checker, the graph-coherence
gate, the issue-tracker sync (GitHub/Azure DevOps/Bitbucket, plus bidirectional status sync
via aidd tracker sync and PR<->task linking via aidd tracker link-pr), CI templates and
aidd ci install, the Claude Code hooks (ten by default, plus an opt-in per-file memory hook; one rule_gate.py enforcing hard rules R1–R12
including the independent-audit and graph-coherence gates, and a stop_gate.py), the evidence log, aidd status/aidd rules, the multi-agent adapter generator (generate_adapters.py, rendering
Gemini CLI/Cursor/Windsurf/Cline/Copilot commands from one source), the real auto-discovery
extension registry (extension_registry.py, loading provider/adapter/hook extensions from
skill/extensions/**/manifest.json and .aidd/extensions/**/manifest.json — see
docs/EXTENDING.md), and the marketplace (catalog/ + marketplace.py +
validate_catalog_entry.py, one first-party package shipped so far: hotfix-report) are built,
tested (stdlib unittest, run in CI), and have been run against a real production
codebase. OpenCode, Codex, and the five generated adapter targets follow the same methodology;
only the Claude Code hooks currently give it technical teeth. Published on PyPI as aidd-cli.
The marketplace is intentionally not a fully automated accept path: CI validates structure and
flags anything worth a human's attention, but a maintainer reviews and merges every catalog
submission by hand — see catalog/README.md.
Support this project
If AIDD saves your agent sessions real time, sponsor it on GitHub — it funds the time spent maintaining the pipeline, the hooks, and the multi-agent adapters.
License
MIT — see LICENSE.
Metadata
Release files for aidd-cli 0.5.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| aidd_cli-0.5.0.tar.gz | 1.4 MB | Without provenance | Details for aidd_cli-0.5.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| aidd_cli-0.5.0-py3-none-any.whl | Python 3 | none | any | Details for aidd_cli-0.5.0-py3-none-any.whl |
Total release size: 3.0 MB
Release files / aidd_cli-0.5.0.tar.gz
| Download URL | aidd_cli-0.5.0.tar.gz |
|---|---|
| Size | 1.4 MB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
bab948187df3250ccfc52f5cc281190bdb8c8750210ec9929dea41c349dfda18
|
|
BLAKE2b-256 checksum How to use checksums |
e0ae62be28c1468e8c5b7118ccceda62263eac23c343d7aeb9e12a85a64c7877
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.12.10
|
Release files / aidd_cli-0.5.0-py3-none-any.whl
| Download URL | aidd_cli-0.5.0-py3-none-any.whl |
|---|---|
| Size | 1.6 MB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
b3cd3925a5886b546e0437335073100045a32e43871e81db6bf36b73af1bb575
|
|
BLAKE2b-256 checksum How to use checksums |
06f46ba27261674521a58f1e2f553d6a5f131afcc15d10a66729d1667312da12
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.12.10
|
Security for aidd-cli 0.5.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.