ScribIA

Scriba + IA — the AI scribe that keeps your documentation in sync with your code.

---

Scribia is an incremental documentation system for AI-assisted development. When you run /scribia in Claude Code (or scribia run in your terminal), it detects what changed since the last run, extracts semantic entities (APIs, classes, models, config), updates only the impacted documentation sections, and optionally syncs a knowledge graph (Graphify) or LLM wiki.

It also captures conversational context — the architectural decisions, motivations, and design rationale that you explain during development — and writes them to docs/decisions/. Code alone doesn't explain why; Scribia does.

It never regenerates everything from scratch. It never deletes existing content. It only patches what changed.

---

Why Scribia

Every AI-assisted workflow has the same problem: code evolves fast, documentation doesn't. Scribia sits between your git history and your docs directory. Run it when you're ready, get only the diff applied to your docs.

• Manual only — you decide when to update. No background daemons, no hooks.
• Incremental — compares to the last checkpoint, touches only affected sections.
• Backend-agnostic — Markdown, Graphify, LLM Wiki, or your own plugin.
• Claude Code skill — invoke with /scribia directly from your AI session.
• Standalone CLI — works without Claude too: scribia run.
• Context-aware — captures architectural decisions from your conversations, not just code changes.

---

Installation

Requirements

• Python 3.11+
• git

Quick install

git clone https://github.com/marcolettieri/scribia.git
cd scribia
./install.sh

The installer will ask three questions:

1. Primary documentation backend — Markdown (default), LLM Wiki, Graphify, or custom
2. Knowledge backends — optional graph/wiki layers to run alongside the primary
3. Confirms manual-only update mode

It then installs the Python package, creates scribia.yaml in your project, and links the Claude Code skill.

Manual install

pip install -e .

To use as a Claude Code skill:

ln -s "$(pwd)" ~/.claude/skills/scribia

Initialize in your project

Navigate to your project and run:

scribia init

This scaffolds the docs/ directory, creates .scribia/state.json with the current HEAD as checkpoint, and copies scribia.yaml from the template.

---

Usage

In Claude Code

/scribia                          # detect changes, update docs
/scribia init                     # scaffold docs/ + set initial checkpoint
/scribia --dry-run                # preview what would be documented
/scribia --force                  # reprocess even if already at HEAD
/scribia --from-commit abc1234    # override start commit
/scribia state show               # inspect last checkpoint
/scribia state reset              # move checkpoint to current HEAD
/scribia config                   # show effective configuration
/scribia backends                 # list available backends

# Conversational context capture
/scribia note "why we chose X over Y"   # queue an architectural note
/scribia session show                   # review session-captured context
/scribia session apply                  # push session context into docs
/scribia session clear                  # discard session log

In terminal

scribia run                       # same as /scribia
scribia run --dry-run
scribia run --from-commit abc1234
scribia init
scribia state show
scribia state reset
scribia config
scribia backends

# Context notes
scribia note "rationale text"     # queue a note for next run
scribia session show              # review session log
scribia session apply             # apply session log to docs
scribia session clear             # clear session log
scribia session capture           # read Stop hook stdin → session log

---

How it works

git diff (from checkpoint → HEAD)
        │
        ▼
   Diff Engine          — identifies changed files, additions/deletions
        │
        ▼
 Semantic Analyzer      — extracts APIs, classes, models, config entries
        │                  filters formatting noise
        ▼
Documentation Updater   — writes only impacted sections using safe markers
        │                  never deletes, never overwrites unmanaged content
        ▼
  Backend Plugin(s)     — Markdown / Graphify / LLM Wiki / custom
        │
        ▼
 Checkpoint saved       — .scribia/state.json updated to current HEAD

Safe write contract

Every managed section is bounded by HTML comments:

<!-- scribia:start:api-my_module -->
| Name | Type | Change | File |
| ... |
<!-- scribia:end:api-my_module -->

Content outside markers is never touched.

Checkpoint

State is stored in .scribia/state.json:

{
  "last_commit": "abc1234...",
  "last_run": "2026-06-01T10:30:00Z",
  "last_summary": "12 files, 8 entities",
  "backend_config": { "backends": ["markdown"] },
  "run_count": 7
}

---

Conversational Context Capture

Code captures what was built. Scribia also captures why.

During an AI-assisted session, you often explain design decisions, trade-offs, and architectural intent that never makes it into the code. Scribia provides two ways to preserve that context in docs/decisions/.

Approach A — Inline notes

Queue a note manually (or the /scribia skill does it automatically from the conversation):

scribia note "chose event-sourcing over CRUD because we need full audit trail"
scribia note "the plugin loader uses importlib instead of exec to preserve traceback info"

Notes are staged in .scribia/context_queue.jsonl and flushed into docs on the next scribia run. The /scribia Claude Code skill automatically extracts architectural insights from the current conversation before running the pipeline.

Approach B — Session log

Install a Stop hook to capture context automatically at the end of each Claude Code session:

scribia hook install --trigger stop           # project-local
scribia hook install --trigger stop --global  # all projects

When the session ends, the hook writes conversation context to .scribia/session_log.jsonl. Review and apply it when you're ready:

scribia session show    # inspect what was captured
scribia session apply   # push into docs/decisions/ and clear log
scribia session clear   # discard without applying

Where decisions are written

Both approaches write to the same destination:

docs/
  decisions/
    2026-06-01.md    # one file per day, one section per commit
wiki/
  decisions.md       # if llm_wiki backend is active

Sections are bounded by scribia markers so they are never overwritten by subsequent runs — only appended.

---

Configuration

scribia.yaml in your project root (or ~/.scribia.yaml for user-global defaults):

# Primary documentation backend
# Built-ins: markdown | graphify | llm_wiki | custom (plugins/)
backend: markdown

# Additional knowledge backends (run after primary)
# knowledge_backends:
#   - llm_wiki
#   - graphify

# Where to write docs
docs_dir: docs
changelog: CHANGELOG.md

# Output language (auto = detect from CLAUDE.md or system locale)
language: auto

# Files to ignore during diff analysis
exclude_patterns:
  - "test_*"
  - "*.test.ts"
  - "migrations/*"

# Graphify backend options
graphify:
  flags: "--update --wiki --no-viz"
  target_path: "."

# LLM Wiki backend options
llm_wiki:
  output_dir: wiki

---

Backends

Markdown (default)

No dependencies. Writes plain .md files:

docs/
  index.md
  api/<module>.md
  features/
  architecture/data-models.md
  configuration.md
CHANGELOG.md

Graphify

Triggers a /graphify . --update --wiki run after documentation is updated. Requires the graphify Claude Code skill. When used via /scribia in Claude Code, Scribia invokes graphify automatically. In CLI mode it prints the command to run manually.

LLM Wiki

Writes per-module structured summaries to wiki/ — optimized for LLM context retrieval. Each file is a flat, machine-readable summary of what exists in that module and what changed.

Custom backend

1. Copy templates/custom_backend.py.template to plugins/my_backend.py
2. Set name = "my_backend" on the class
3. Implement init(), update(), persist()
4. Set backend: my_backend in scribia.yaml

No core code changes required. The plugin is loaded dynamically at runtime.

from scribia.backends.base import BackendPlugin
from scribia.engine.models import ChangeSet

class MyBackend(BackendPlugin):
    name = "my_backend"

    def init(self, config: dict) -> None: ...
    def update(self, changeset: ChangeSet) -> list[str]: ...
    def persist(self) -> None: ...

---

Docs structure generated

After first run on a typical Python project:

docs/
  index.md                     # navigation index
  api/
    my_module.md               # per-module API table
    service.md
  architecture/
    data-models.md             # data model registry
  decisions/
    2026-06-01.md              # architectural context per day/commit
  configuration.md             # config reference
CHANGELOG.md                   # prepend-only changelog
wiki/                          # only if llm_wiki backend enabled
  index.md
  my_module.md
  decisions.md                 # all session context notes

---

Project structure

scribia/
├── SKILL.md                   # Claude Code skill (/scribia)
├── scribia.yaml.example       # configuration template
├── install.sh                 # interactive installer
├── pyproject.toml
├── plugins/                   # custom backend plugins go here
│   └── example_custom_backend.py
├── templates/
│   └── custom_backend.py.template
└── src/
    └── scribia/
        ├── cli.py             # CLI entry point
        ├── config.py          # configuration loader
        ├── engine/
        │   ├── models.py      # ChangeSet, FileChange, SemanticEntity
        │   ├── diff.py        # git diff engine
        │   └── analyzer.py    # semantic entity extraction
        ├── updater/
        │   └── safe_write.py  # section-aware file writer
        ├── backends/
        │   ├── base.py        # BackendPlugin ABC
        │   ├── loader.py      # dynamic backend loader
        │   ├── markdown.py    # Markdown backend
        │   ├── graphify.py    # Graphify backend
        │   └── llm_wiki.py    # LLM Wiki backend
        └── state/
            ├── manager.py     # .scribia/state.json checkpoint
            ├── context_queue.py  # .scribia/context_queue.jsonl (Approach A)
            └── session_log.py    # .scribia/session_log.jsonl (Approach B)

---

Contributing

Contributions welcome. The cleanest way to extend Scribia is to write a new backend plugin — see templates/custom_backend.py.template. If your backend is general enough to be useful to others, open a PR to add it to src/scribia/backends/.

---

License

MIT with Attribution Requirement — free for personal use, attribution required for organizational or commercial use. See LICENSE for the full terms.

© 2026 M. Lettieri