fabricatio-skill

MIT

An extension of fabricatio.

Installation

This package is part of the fabricatio monorepo and can be installed as an optional dependency:

pip install fabricatio[skill]

# or with uv
# uv pip install fabricatio[skill]

Or install fabricatio-skill along with all other components of fabricatio:

pip install fabricatio[full]

# or with uv
# uv pip install fabricatio[full]

Overview

A text-based skill system for fabricatio agents. Skills are plain markdown files (YAML frontmatter name/description/tags + markdown body) that inject just-in-time, task-relevant context into LLM prompts through progressive disclosure: cheap metadata first, LLM-powered selection next, distilled essence last — never the full corpus.

Exclusion is just as progressive: a deterministic Rust keyword pre-filter thins oversized pools before any LLM sees them, selection admits only catalog-validated candidates (capped), and distillation drops everything the question does not need. Content that fails an earlier stage never reaches a later prompt.

Key Features

• Markdown-native skills — author skills as .md files with YAML frontmatter; no schema lock-in beyond three metadata keys.
• Two library layouts — per root, skills resolve either by the agent-skills convention (<name>/SKILL.md) or as flat files (<name>.md); the library loads the cross-client .agents/skills locations (project + ~), plus any extra_skill_dirs you configure.
• By-name gathering — gather_skills(names, dirs=None) resolves skills straight from the lookup roots by name (direct path reads, no corpus scan), loads them into the library, and tracks them on the role.
• Progressive pipeline — Level 0 (library): the process-wide SkillRegistry — recursive scanning (load_scanned), by-name resolution (load_by_name), keyword search (SkillRegistry.search, also the deterministic pre-filter for huge libraries); Level 1 (Python): LLM-powered relevance selection (select_skills, delegated to the framework UseLLM.achoose chooser over skill briefings — set-validated, auto-retried, SMOL tier, capped) and essence distillation (distill_skills); Level 2: the composed consult_skills pipeline (select -> distill -> consulted knowledge). Consultation only — answering stays in your Action.
• Progressive disclosure dial — every call trades fidelity for tokens via select= / distill= / forced names=.
• Lightweight composition — heavy Skill objects live in the process-wide library (get_skill_registry(), reachable from a role as role.skill_library); your roles/actions carry only a list of name handles, and the loaders (scan_skills, gather_skills, add_skills) chain on the role.
• Rust-backed performance — parsing, lookup, and keyword matching are PyO3 (fabricatio_skill.rust).

Usage

1. Author skill files

Drop markdown files into a skill directory. The default roots are the cross-client ones only — .agents/skills (project) and ~/.agents/skills (user), the location the Agent Skills standard tells every host to read. Each root supports both layouts — the agent-skills directory convention and flat files:

.agents/skills/
├── rust-async/SKILL.md      # <name>/SKILL.md layout
└── code_review.md           # <name>.md layout

---
name: rust-async
description: "How to write async Rust with tokio correctly"
tags: [rust, async]
---
Markdown body with the actual instructions...

Frontmatter notes:

• name is the unique catalog id (the LLM selects by it); description is the whole selection surface — keep it a dense one-liner; tags help the Rust keyword pre-filter.
• YAML rules apply: quote string values that contain : (e.g. descriptions), or the field silently parses as empty.

2. Consult — zero-config (canonical path)

Nothing to load. The process-wide library is created already loaded from the cross-client directories (.agents/skills under the process working directory, plus the user-level ~/.agents/skills) and from extra_skill_dirs; a role that loaded no names of its own consults the whole library. Authoring files and consulting is the whole loop:

from fabricatio import Action, Event, Role, Task, WorkFlow
from fabricatio_skill import UseSkill


class AnswerWithSkills(Action, UseSkill):
    """Answer the task briefing grounded in consulted skill knowledge."""

    output_key: str = "task_output"

    async def _execute(self, task_input: Task[str], **_) -> str:
        # select relevant skills -> distill what they say (both SMOL tier);
        # "" when nothing is relevant. Answering stays YOUR job.
        knowledge = await self.consult_skills(task_input.briefing)
        if not knowledge:
            return task_input.briefing
        return await self.aask(f"{knowledge}\n\n---\n\n{task_input.briefing}")


role = (
    Role.with_bio(name="skilled", description="answers using the skill library")
    .subscribe(
        Event.quick_instantiate("ask"),
        WorkFlow(name="skill_qa", steps=(AnswerWithSkills,)),
    )
    .dispatch()
)
# dispatch a Task whose briefing is the question to Event "ask" — done

Need the library without consulting anything? A role that tracked nothing resolves the whole library, and the handle is always at hand:

self.skills                    # the role's pool as Skill objects (the library, when nothing was tracked)
self.skill_library.names()     # every name the library holds right now
self.skill_library.get("rust-async")   # one skill by exact name, or None

The library comes from get_skill_registry(), created once per process: it loads the cross-client roots — fixed on the Rust side, with ~ expanded and absent roots skipped — plus extra_skill_dirs.

Client-specific locations (.claude/skills, a bundled skills/ dir, …) are deliberately not part of the standard roots. Load them explicitly — scan_skills(".claude/skills"), gather_skills([...], dirs=[...]), or list them in extra_skill_dirs to search them process-wide.

3. Tune the disclosure level per call

consult_skills() is the progressive-disclosure dial; it returns what the relevant skills say (or "" when nothing is relevant — a "[]" selection from the LLM, an exhausted selection retry, or an empty library all mean the same thing: there is nothing to consult):

For finer control, step through the levels manually:

picked = await agent.select_skills(question)              # metadata-only chooser
if picked:                                                # None = LLM never answered validly
    essence = await agent.distill_skills(question, picked)  # compress only selected bodies
    answer = await agent.aask(f"{essence}\n\n---\n\n{question}")
else:
    answer = question                                     # nothing relevant: no skill context

What the LLM sees at each stage (progressive exclusion in action):

• SELECT — the framework UseLLM.achoose chooser lists every candidate by its briefing (name: description; bodies never enter this prompt) and asks for a JSON array of catalog names. The reply is parsed as a set: unknown names are ignored, duplicates collapse, unparseable replies are retried automatically (up to 3 attempts). A deterministic keyword search (SkillRegistry.search(..., names=<the role's pool>)) pre-filters the pool first when it exceeds prefilter_threshold, and the result is trimmed to max_selected_skills (or the per-call k).
• DISTILL — only the selected skills' bodies enter this prompt, with the instruction to extract just the parts relevant to the question and discard everything else.

4. Explicit loading — custom directories & the registry

Skip the auto-load by loading skills yourself. The loaders chain, and re-running one is safe: a name already in the library keeps its first copy, and the loader reports it again rather than re-reading the file, so the role tracks it either way:

from fabricatio import Action, Task
from fabricatio_skill import UseSkill


class AnswerWithTeamSkills(Action, UseSkill):
    """Answer using skills from a non-default directory."""

    output_key: str = "task_output"

    async def _execute(self, task_input: Task[str], **_) -> str:
        # scan the custom root, then resolve two names against it
        self.scan_skills("team-skills").gather_skills(["security", "review"], dirs=["team-skills"])
        knowledge = await self.consult_skills(task_input.briefing)
        ...

add_skills(skills) registers Skill objects you parsed yourself.

5. Gather directly by name

When you already know which skills you want, skip scanning entirely: gather_skills resolves each name through the lookup roots — trying <dir>/<name>/SKILL.md first, then <dir>/<name>.md (direct path reads, no directory walk) — registers the hits, and tracks them on the role. Unresolvable names are skipped, not raised: the library reports them in one warning naming every root it searched, so an all-miss call simply tracks nothing. With no dirs=, the lookup roots are the cross-client ones plus extra_skill_dirs, so the agent-skills library is available with zero configuration:

from fabricatio import Action, Task
from fabricatio_skill import UseSkill


class AnswerWithNamedSkills(Action, UseSkill):
    """Answer using two explicitly gathered skills."""

    output_key: str = "task_output"

    async def _execute(self, task_input: Task[str], **_) -> str:
        self.gather_skills(["rust-async", "code_review"])   # by name, no scan
        knowledge = await self.consult_skills(task_input.briefing)
        return await self.aask(f"{knowledge}\n\n---\n\n{task_input.briefing}")

Pass dirs=[...] to resolve against custom libraries instead of the standard roots (~ is expanded by the library). Without a role, the same resolution goes straight through the library: load_by_name returns the names it made available.

On a role or action, the library is the skill_library property, so role.skill_library.search("async", names=role.skill_names) searches exactly the pool the role consults.

The library is also exposed for framework-free pipelines: SkillRegistry() is a bare handle over the same process-wide store, and get_skill_registry() is the instance loaded from the standard roots plus the extra ones:

from fabricatio_skill import SkillRegistry

library = SkillRegistry()                                # process-wide: one copy per name
scanned = library.load_scanned([".agents/skills"])       # parse every .md file under a root
library.load_skill_dirs(["~/my-skills"])                 # the standard roots + these extras
wanted = library.load_by_name(["rust-async"], [".agents/skills"])  # resolve by name (no walk)
library.search("async", names=wanted, in_content=True)   # keyword search over that pool
library.get("rust-async")                                # one loaded skill, or None
library.remove(["rust-async"])                           # free the name; next load re-reads it

Both loaders return the names they made available — a name already in the library is reported too, never read a second time — which is how the role-level scan_skills/gather_skills track what they loaded. load_scanned is strict (a missing root raises FileNotFoundError), while load_by_name and load_skill_dirs are lenient: unresolved names and absent roots are skipped, and load_by_name reports the names nothing resolved in one warning naming the roots it searched; add, remove, and clear return the library for chaining.

Configuration

All options below are read through the fabricatio configuration chain (see the Configuration Guide at ../../docs/source/configuration.rst). Set them under the [ext.skill] table in fabricatio.toml, equivalently under [tool.fabricatio.ext.skill] in pyproject.toml, or via FABRICATIO_EXT__SKILL__<FIELD_UPPER> environment variables.

# fabricatio.toml
[ext.skill]
distill_skills_template = "built-in/distill_skills"
max_selected_skills = 8
prefilter_threshold = 100
extra_skill_dirs = []

Access at runtime: from fabricatio_skill.config import skill_config.

Dependencies

Core dependencies:

• fabricatio-core - Core interfaces and utilities ...

License

This project is licensed under the MIT License.