Skills

Decodie provides seven Agent Skills compatible with Claude Code, Gemini CLI, Cursor, and 70+ other AI coding agents.

npx skills add owenbush/decodie-skill --all

decodie-observe

Activate at the start of a coding session. The agent documents decisions, patterns, and concepts as it writes code in real-time.

• Entries are written after each meaningful decision, interleaved with coding
• Tracks session start/end with automatic session IDs
• Detects duplicates and cross-references related entries
• Marks entries as superseded when referenced code is rewritten or removed
• Links to external documentation (MDN, php.net, React docs, etc.)

decodie-analyze

Generate learning entries from existing code — even code that wasn’t written with an AI agent. Read-only: never modifies source files.

• Target a single file, a directory, or an entire project
• Selective mode (default): picks the 3–5 most significant patterns per file
• Exhaustive mode: documents every meaningful pattern
• Analysis sessions are prefixed analyze- for easy identification
• Same duplicate detection and cross-referencing as observe

Source annotations

Place annotation markers in source code comments to control what gets analyzed. These work in both selective and exhaustive modes.

Marker	Scope
@decodie-include:file	Entire file
@decodie-include:class	Next class/interface/enum
@decodie-include:function	Next function/method
@decodie-include:start/end	Block region
@decodie-ignore:file	Entire file
@decodie-ignore:class	Next class/interface/enum
@decodie-ignore:function	Next function/method
@decodie-ignore:start/end	Block region

@decodie-ignore always takes precedence over @decodie-include when scopes overlap. @decodie-include entries don’t count against the 3–5 limit in selective mode.

decodie-explain

Get an immediate, detailed explanation of a specific piece of code — what it does, how it works, potential issues, and improvement suggestions.

• Summary — a high-level overview of what the code does
• Detailed breakdowns — complex sections explained individually with code excerpts
• Issues — potential bugs, security concerns, and edge cases flagged with severity
• Improvements — refactoring suggestions and modern alternatives with rationale
• Ephemeral by default — nothing is saved unless you explicitly ask
• Saved explanations use the explain- session prefix

decodie-ask

Ask questions about existing learning entries and get deeper explanations using the entry content and live source code as context.

• Resolves the target entry by ID, keyword match, or most recent in session
• Loads the full entry content and reads current source files for context
• Goes beyond the entry — explains underlying concepts, provides examples, clarifies trade-offs
• Read-only — does not create or modify entries

decodie-overview

Generate a high-level overview of a file, directory, or project — purpose, structure, entry points, and dependencies. The onboarding skill: run this first when opening unfamiliar code.

• Target a single file, a directory, or the whole project
• Persists by default — overviews are saved as entries so they can be browsed and shared
• Re-running on the same target overwrites the existing overview in place
• Overview sessions are prefixed overview-
• Captures purpose, structure, optional entry_points, and optional dependencies

decodie-verify

Walk every entry’s references, confirm the anchored code still resolves, and stamp confirmed entries with the current commit SHA.

• Re-checks each anchor_hash against the current file contents
• Backfills the sources field on older entries
• Sets verified_sha to the current HEAD when all anchors resolve
• Flips stale: true when an anchor is missing or a file has been removed
• Optional path argument scopes the check to a subdirectory

decodie-flag-stale

The cheap counterpart to decodie-verify, designed for CI on every PR. Diffs source files against each entry’s verified_sha and flips stale: true on entries whose files have changed.

• Uses git diff --name-only — fast enough for every PR
• Only sets the flag; clearing it requires decodie-verify
• Skips entries that have never been verified
• Output is machine-friendly so wrapping tools decide PR blocking policy
• Used under the hood by the GitHub Action