What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
An AI assistant can gain repeatable workflows without turning each one into a new tool: store the procedure in a validated Markdown file, advertise a short description in the system prompt, and load the full instructions only when needed. That is the small skills system Krish Verma describes building for his open-source desktop assistant, Ankita. His “zero code” framing applies to the skill content, not the software that discovers and loads it.
What counts as a skill in this design?
A skill is a named folder under a skills/ directory, containing a SKILL.md file. Its frontmatter holds metadata such as the skill name, description, and optional suggested tools; the rest of the file gives the assistant a procedure to follow.
As an Amazon Associate I earn from qualifying purchases.
Verma says the design grew from repeatable work he had kept as copied notes, including commit reviews, release-note drafting, structured web research, and bug reproduction. Instead of pasting those instructions into every conversation, he made them discoverable as separate skills. The folder name must match the frontmatter name.
How does the assistant load a skill?
The system uses progressive disclosure: it puts a brief listing for each available skill in the system prompt, then loads the full body only when the assistant requests that skill. The listing includes the name, description, suggested tools, and call syntax. In Verma’s implementation, the loaded skill body is capped at 8,000 characters.
#1 Best Overall
This separates prompt discovery from procedural detail. The prompt can tell the assistant what workflows are available without carrying every workflow’s full instructions on every turn. Verma characterizes the short description as the advertisement and the body as the product.
Where do skills stop and tools begin?
A skill describes how to do a task; it does not itself perform the task. When a procedure needs an action—such as inspecting a commit or retrieving information—the assistant discovers and invokes the relevant tool through its normal deferred tool-discovery process, including that tool’s usual approval and validation rules.
That boundary is the central architectural choice: tools do, skills teach, and the system prompt frames. A Markdown instruction file is data read by the assistant, not an execution surface. “Zero code” therefore describes what a skill author writes, not the entire feature: the host application still needs code to find, validate, cache, and load skill files.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →What validation rules keep malformed skills out?
Verma reports that the parser applies these checks before a skill is made available:
- The folder name must match the frontmatter
name. - The name must match
^[a-z0-9-]{1,64}$. - The description must be 10–300 characters.
- The body must be non-empty and under 12,000 characters on disk.
- Suggested tools, if supplied, are limited to 200 characters.
If a skill fails validation, the application logs an error and skips it rather than interrupting startup or adding the malformed entry to the prompt. The 12,000-character disk limit and 8,000-character load-output cap are different constraints: one bounds the stored body, the other the content returned when a skill is loaded.
How are skills disabled and extended?
Users can toggle skills in Plugins > Skills. Disabled skills are omitted from the prompt, and calling one returns a plain error string. A skill folder may also contain an optional plugin.json for palette actions; Verma says those actions are checked against a palette schema.
What did the author build, and what remains uncertain?
In his 2026 account, Verma reports five built-in skills: commit-review, release-notes, web-research, bug-repro, and ankita-dev. He describes the parsing, caching, and tool wrapper as roughly 150 lines. These are the author’s implementation figures, not independently audited measurements, and his article gives no comparative benchmark or quantified performance result.
He also notes that file-stamp cache keys and a reloadSkills() escape hatch handled reloading, but cache invalidation proved more complicated than the feature warranted. His other retrospective lesson was to define the boundary between skills, tools, and prompts earlier. For someone adapting the pattern, the practical design questions are how much summary text to put in the prompt, what limits to validate, what to do when a file fails, and how edits invalidate cached content.
Source: Krish Verma’s first-person article on DEV Community, posted October 1, 2026. The implementation details and counts above reflect his description; the linked code was not independently inspected.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
