Skills
Perch carries a pack of operator skills that load only when they are relevant. Each one is a set of working rules, preferred tools, and a verification standard for a kind of work. Here is what ships, how one gets picked, and how to add your own.
A skill is a body of working knowledge for one kind of work: how to build a spreadsheet that a reviewer can trust, how to operate a logged-in browser, how to inspect a repository, how to make a deck that holds up in a board meeting. Perch ships a pack of them, and the right ones load for the turn in front of you.
The reason this exists is context. Loading every instruction for every kind of work on every turn would spend the window on rules that do not apply. Skill metadata stays cheap to scan, and a skill's full body loads only when the turn matches it.
The short version
SKILL.md files.What ships
| Skill | What it covers |
|---|---|
| workspace | Navigating, searching, summarizing, and safely editing local folders and codebases with file-grounded evidence. |
| spreadsheets | XLSX, CSV, TSV, and Google Sheets. Formulas, pivots, charts, reconciliation and exception tabs. |
| documents | DOCX, Google Docs, memos, letters, reports, redlines, and tracked changes. |
| presentations | PPTX decks, board and pitch decks, speaker notes, chart slides. |
| Extract, cite by page, split, merge, create, and inspect PDFs, including scans and forms. | |
| diagrams | Mermaid and draw.io. Workflows, org charts, entity networks, control graphs, architecture maps. |
| data-quality | Profiling messy data, finding anomalies, reconciling systems, and scoring risk with deterministic math before narrative. |
| research | Gathering, comparing, ranking, and citing evidence from the web, scholarly sources, and your own files. |
| github | Repository and git state, commit and pull request review, debugging failing checks, and publishing changes behind permission gates. |
| browser-operator | Driving the logged-in browser using visual state and semantic page targets. |
| wait-operator | Watching bounded conditions, polling deploys and jobs, and resuming with verified state rather than a guess. |
Two more are about Perch itself: about-perch answers detailed product questions, and release ships Perch's own builds. A skill-creator skill authors and validates new skills, with frontmatter checks and trigger collision detection.
How one gets picked
A skill gets loaded two ways.
First, Perch matches your message and the state of your workspace against each built-in skill's metadata: its trigger hints, the surfaces it supports, the tools it prefers, and the other skills it works well alongside. It admits a small number of skills, not the whole pack.
Second, the model can load a skill itself. Every skill available on the turn is listed for the model by name and description, and the model loads the full instructions of the one a task needs. That is how a skill you wrote gets used even though it has no trigger hints. A skill that has been loaded stays active for the rest of the chat.
Three things influence the first choice:
- What you asked for. The wording of the request against each skill's trigger hints.
- What is available. A skill whose preferred tools are not available on your surface is not a useful thing to load. The browser skill is fully supported on Desktop and limited in the CLI, and it is declared that way rather than discovered at failure time.
- What is in front of you. The folder and files in scope.
Skills also name companion skills, so a deck that needs chart data pulls in the spreadsheet rules, and a page-cited source packet pulls in the PDF rules.
Why each one carries a verification standard
The part that matters most is the least visible. A skill is not only instructions for producing something. It also defines what has to be true before the result is handed to you.
- A workbook is checked for the file existing, formulas being present where they belong, totals reconciling, and CSV exports keeping exact values.
- A deck is checked for slide count, readable text, layout that does not overflow, and charts that actually rendered.
- Identifiers such as invoice numbers, account numbers, and tickers are preserved as text, because silently coercing them into numbers is how a reconciliation quietly goes wrong.
This is the same discipline as citation verification, applied to files instead of claims.
Your own skills
A skill is a folder with a SKILL.md file in it: a name, a description of when to use it, and the steps. Perch reads your skills from these places, in the folder you are working in and in your home folder:
| Path | What it is |
|---|---|
.perch/skills/<name>/SKILL.md | Perch's own skills folder. |
.agents/skills/<name>/SKILL.md | The shared Agent Skills folder that other tools read too. |
.perch/plugins/<plugin>/ | An Agent Plugins package. Its skills get the id <plugin>:<skill>. |
A skill in the folder wins over a personal skill with the same name, and either one replaces a built-in skill of that name. Files bundled next to SKILL.md, such as references or scripts, are read only when the model asks for one.
In Perch Desktop, with a folder open, the Skills pill above the message box shows how many skills are on in that folder. Its menu lists each skill with where it comes from (Built-in, This folder, Personal, or Plugin) and a switch to turn it off for that folder. New skill creates one under .perch/skills/, and Add from a folder copies in a skill you already have.
In the CLI:
perch skills # list skills, or `perch skills <name>` to print one
perch skills new <name> # start a new skill in .perch/skills
perch skills eject <name> # copy a built-in skill into .perch/skills to edit
perch skills disable <name> # turn a skill off; `enable` turns it back onAdd --personal to any of the last three to target your home folder instead of the project.
How the built-in set is made
The skills that ship with Perch are built in. They are generated from source at build time into a committed registry, so the set you get is reproducible for a given release rather than assembled at runtime.
The agent harness
Perch runs a full agent harness on Desktop and CLI: shell, search, file editing, background jobs, a sealed sandbox, plan mode, and risk-tiered permissions. Here is the complete toolset and the controls that govern it.
Retrieval
How Perch grounds an answer in your documents, what it has learned about you, and the public web when a question needs it.
