What is a Skill?

A skill is a chat-invokable action that orchestrates one of the pipeline primitives already built into AI4Meta. Skills are thin wrappers over existing endpoints, not a parallel execution layer. When you type a slash command in the copilot, the same code path runs that Autopilot would run if you clicked its button. This keeps reviews reproducible: the audit log does not care whether you clicked or typed.

Every skill invocation is recorded with the invoking user, the model, the prompt parameters, the input artifact ids, and the output hash. The record is attached to the project so reviewers can reconstruct the session.

Skill Catalog

The built-in catalog below ships with every project. Plugins may add further skills (see Plugins & Extensions).

/screen

Purpose. Run or rerun title and abstract screening against the project protocol.

Inputs. Paper ids (default: all pending). Optional inclusion or exclusion overrides.

Model class. Classification. Defaults to the fast model configured for the project.

Writes. Screening decisions, conflicts, and provenance records. Does not overwrite human decisions without an explicit flag.

Example. /screen pending strict runs stricter criteria on the pending set and routes ambiguous cases to the conflict queue.

/extract

Purpose. Apply the project codebook to selected papers and write the results into the extraction matrix.

Inputs. Paper ids and optional variable subset.

Model class. Reading and extraction. Defaults to the quality model.

Writes. Extraction cells with source references (page, quote), flagged if confidence is low.

Example. /extract paper:12 variables:sample_size,design fills two specific cells on one paper.

/synthesize

Purpose. Run meta-analysis (MA mode) or narrative synthesis (SR mode) using the compiled dataset.

Inputs. Optional moderator list and estimator override.

Model class. Statistical (no LLM for MA), reasoning for SR narrative.

Writes. Analysis results, forest and funnel plots, heterogeneity and publication-bias diagnostics.

/report

Purpose. Generate the final publication-format report. This is the last step of Autopilot; when invoked explicitly it skips the car scene and opens directly into the journal layout.

Inputs. Optional section list.

Model class. Reasoning for narrative sections.

Writes. A versioned report artifact in the project.

/critique

Purpose. Methodological critique of the current protocol or draft report. Flags missing PICO components, unclear eligibility criteria, overlapping exclusion rules, or weak outcome definitions.

Inputs. Protocol or report artifact id.

Model class. Reasoning. Defaults to the quality model.

Writes. A critique artifact (not attached to the manuscript by default).

/format apa

Purpose. Reformat a passage, table, or reference list to APA 7 style.

Inputs. Text or artifact id.

Model class. Fast.

Writes. A new version of the input artifact. Does not overwrite the original.

/prisma

Purpose. Generate a PRISMA flow diagram from the current project counts (imported, deduplicated, screened, excluded with reason, included).

Inputs. None.

Model class. Deterministic, no LLM.

Writes. A diagram artifact (SVG and PNG) and a counts snapshot.

/codebook ai

Purpose. Generate codebook variables from Protocol PICO and inclusion or exclusion criteria. Enforces the PICO-grounded constraints added as of task 46.

Inputs. Optional user hints (variable names to seed).

Model class. Reasoning.

Writes. Codebook draft for review. Does not activate until confirmed.

Invocation Surfaces

Skills can be invoked from three surfaces, and they all hit the same endpoints:

1. Copilot chat. Type the slash command in the chat input.
2. Workspace buttons. The existing SR Autopilot, Rocket Report, and per-view action buttons wrap the same skills.
3. Keyboard palette. ⌘K opens a quick launcher with the full catalog.

Model Selection

Each skill declares a default model class (fast, quality, or reasoning). The project picker maps each class to a concrete model (Opus, Sonnet, Haiku, or a contracted institutional endpoint). Users may override per invocation, but the override is recorded on the audit entry.

Permissions and Scoping

Skills inherit the project's permission model. A viewer cannot write; a reviewer can queue but not commit; an owner can commit. Skills that modify data check the caller's role before executing.

See Also

• Plugins & Extensions. How to add skills that are not in the built-in catalog.
• AI Features. Underlying primitives that skills wrap.
• Quick Report (Autopilot). Orchestration of several skills into a single pipeline run.