Skills/pseudo/SKILL.md

name, description

name

description

pseudo

Use when the user wants code explained in human-readable, commented pseudocode rather than prose — invoked as `/pseudo` or when the user asks to "explain this in pseudocode". Every reply is centred on a pseudocode block (prose is allowed, but only in a supporting role to explain what the code cannot show); opens with a source-file-and-approximate-line anchor; tags notable steps with reference markers ([1], [A]) the user can point back at; and sketches the layout of the data structures the code touches alongside the control flow. Strips detail irrelevant to the question and omits low-level mechanics (memory management, error plumbing) unless the question is about them. Defaults to the project's main entry point.

pseudo — explain code as commented pseudocode

Violating the letter of these rules is violating the spirit.

Overview

The user wants to understand how code works without reading the code. The medium of explanation is commented pseudocode — not the real source, and not prose carrying the explanation on its own. Prose has a place, but a supporting one: it explains what the code cannot show and then gets out of the way. The pseudocode is a teaching artefact: it shows the shape of the logic at the altitude the question demands, and nothing more.

This is a conversational behaviour skill. It dispatches no agents and runs no pipeline. While it is active, every answer obeys the rules below.

The Iron Law

PSEUDOCODE IS THE CENTRE OF EVERY ANSWER; PROSE ONLY SERVES IT.
EVERY ANSWER OPENS WITH ITS SOURCE FILE AND APPROXIMATE LINE.
MARK THE NOTABLE POINTS SO THE USER CAN POINT BACK AT THEM.
SHOW ONLY WHAT THE QUESTION IS ABOUT — REDUCE THE REST TO A STUB.
SKETCH THE DATA SHAPES THE CODE TOUCHES, NOT JUST THE CONTROL FLOW.
OMIT LOW-LEVEL MECHANICS UNLESS THE QUESTION IS ABOUT THEM.

Rules

1. Pseudocode carries the answer; prose serves it. The pseudocode block is the centre of every reply. Prose is allowed — but only to explain something the code cannot show on its own, and it stays subordinate: a short note before or after the block, never the main event. Reach first for a # comment inside the block; spend a sentence of prose only when the point needs more room than a comment gives. If a reply is mostly prose with a token snippet, the balance is wrong.

2. Anchor line first. The block opens with a # comment naming the source file and the approximate line the explanation starts at — e.g. # src/server.rs ~line 42. The line is a navigation hint, not a promise of exactness; "~" signals that. If the answer spans several files, each new file's section opens with its own anchor.

3. Mark the notable points. Tag the steps worth discussing with a reference marker — [1], [2] or [A], [B] — at the end of the line, so the user can reply "expand [2]" or "why [B]?" without quoting code. Mark only meaningful points, not every line; numbering noise defeats the purpose.

4. Human-readable, not language-specific. Use plain constructs (if, for each, call, return) and real names from the code where they aid recognition. Do not reproduce the actual source syntax.

5. Reduce the irrelevant. Anything not on the path the user asked about collapses to a one-line stub: validate(input) # details elsewhere. Detail is spent only on the topic.

6. Omit mechanics by default. Memory management, allocation, error-propagation plumbing, logging, and similar are left out — unless the question is precisely about them, in which case they become the topic and rule 5 reduces everything else.

7. Show the data shapes, not just the flow. The control flow is only half the picture; sketch the layout of the data structures the code in focus reads or builds — the record fields, the collection nesting, the variant cases that matter to the question. Put the shape near the code that touches it, in the same reduced spirit: only the fields on the topic path, the rest stubbed (… # other fields elsewhere). When the data layout is the question, it leads the answer and the control flow becomes the stub.

8. Default entry point. When the user names no starting point, begin at the project's main entry (the main function / primary CLI or server bootstrap). Otherwise start where they point.

9. Stay at one altitude; let the user zoom. There is no length cap, but each answer holds a single altitude rather than expanding every branch at once. When a level has more depth than the question needs, show that level and let the markers invite the user to expand one point next — steer the zoom with them rather than dumping every layer unprompted.

Shape of a good answer

# src/server.rs ~line 30 — how a request reaches a handler

Router:                                   # [1] the structure match() reads
    routes: list of { path_pattern, handler }
    …  # fallback handler, middleware chain elsewhere

main():
    config = load_config()                # details elsewhere
    router = build_routes()               # fills Router.routes
    serve(router):
        for each incoming request:
            handler = router.match(request.path)   # [2] scans routes
            response = handler(request)            # [3]
            send(response)

That is a whole answer. It opens with the file-and-line anchor, sketches the Router shape [1] right beside the code that reads it (only the field on the topic path; the rest stubbed), keeps the request path in focus, stubs the config load, omits the socket and memory mechanics, and marks the discussable steps. It holds one altitude: the user can now say "expand [2]" and the next answer zooms into route matching — with its own anchor line.

When to stop

The skill stays active for the conversation. The user leaves it by asking for a plain prose explanation (one not built around a pseudocode block), invoking another skill, or saying so.