PROJECTSKILLSCONNECTING…

SYS.00 / OVERVIEW

What this is

ProjectSkills reads a codebase, separates what it can prove from what it inferred, and writes AI Skills grounded in that specific project.

Start here4 SECTIONS

The claim

One sentence, and the part of it that is unusual.

THE LOOP
PROJECT  →  INTELLIGENCE  →  SKILLS  →  AI TEAM

An AI agent working in your repository knows the language and the framework. It does not know that your API is the only writer of authoritative rows, that one package must never be transformed by a particular bundler, or that a relative path in one environment variable resolves to two different directories. That knowledge exists — in the code, in the configuration, in the shape of the repository — and no general model has read it.

ProjectSkills reads it, and writes it down in the form an agent loads: a Skill per role, plus a Constitution the whole team works under.

Parsers first, models second

Five of the pipeline steps need no AI provider at all.

Everything a parser can determine is determined by a parser. Languages, frameworks, dependency versions, import edges, route files, module boundaries — none of that is a judgement, so none of it is asked of a model.

The practical consequence: an account with no model credentials still gets a complete structural map of its project, its dependency graph, its findings and its constraints. What it loses is the reasoning layer on top — not the product.

  • Reading, mapping, technology detection, architecture and dependencies are deterministic.
  • Intelligence, Constitution, Skill generation and Skill review are the reasoning layer.
  • The knowledge graph is deterministic too, and costs zero tokens.

The full breakdown is in The pipeline.

Why the output can be checked

Every claim carries a confidence and the files it came from.

ConfidenceWhat it meansWho can produce it
VERIFIEDRead directly from your files, and re-checked against the real file indexA parser, or a model whose citations resolved
INFERREDReasoned from what was read, not stated anywhereA model, or a deterministic heuristic like architecture style
UNKNOWNNobody established it. A result, not a gapAnything that could not conclude

A claim marked VERIFIED that cites a file which is not in the snapshot is demoted to UNKNOWN — not flagged, not footnoted, demoted, before it reaches a screen. That gate is code. It is not a sentence in a prompt asking the model to be careful.

Read Grounding and evidence for how the gate works and what it deliberately cannot catch.

What it is not

The four things people reasonably expect that this does not do.

  • It does not write code into your repository. Nothing here has a write path to a git host — every analysis reads a snapshot.
  • It does not execute your project. No install, no build, no test run, no script from a manifest. A check that would need one is reported as NOT_SUPPORTED rather than guessed at.
  • It does not watch your repository. An analysis happens when you ask for one.
  • It does not chat about your codebase. The graph answers questions by returning a subgraph of real files; nothing in that path is generated.
DOCS23 CHAPTERS