Lossless Skills v0.0.0.1: From a Question to a Repo

Why Care?

If you’ve ever wished your AI coding agent showed up to the project already knowing your conventions — your filename rules, your tech taboos, your team’s voice, the exact framework you’d refuse to ship — this repo is one path to that future.

lossless-group/lossless-skills is now a public collection of Agent Skills that codify The Lossless Group’s working conventions. Skills work in Pi, Claude Code, OpenAI Codex, and anything else that follows the standard. Install once, get them everywhere.

For the broader audience: this is the Lossless Group’s “Lost in Public” practice in action — a session that started as a question became a working artifact, and we’re publishing the artifact and the journey in the same breath.

What’s New?

Four skills shipped:

context-vigilance — codifies the context-v/ framework: six folders organized into Planning (specs/ ↔ prompts/), Reflective (blueprints/ ↔ reminders/), and Journey (explorations/, issues/) modes. Four-part epoch.major.minor.patch versioning, Train-Case filenames and tags, “norms not rules” ethos.
astro-knots — the vision-mission and tech conventions for our family of ~10+ Astro sites. The Tech Hierarchy (HTML & CSS first), the approved list (Astro, Svelte, GSAP, Reveal), the hard prohibitions (React, JSX, Angular, “bloat”).
pseudomonorepos — our coined term for parent repos that aggregate children primarily to host parent-level context-v/. Encodes the search-first-before-creating discipline: walk the tree, find prior work, surface findings, link rather than duplicate.
changelog-conventions — what produced this entry. Frontmatter contract (publish: true, lede, ISO dates, authors: for humans only, augmented_with: for AI tooling, files_changed: for the diff trail), filename pattern, and the foundational Changelog First Development theory.

Plus repo infrastructure:

A global ~/.pi/agent/AGENTS.md that loads on every Pi session — encodes the drift policy (“observe inconsistency, surface it, but don’t auto-fix as a side effect”) and the new authorship convention (humans in authors, tooling in augmented_with).
A running CANDIDATES.md backlog of future skills: lfm, lossless-loop, obsidian-integration, submodule-hygiene, astro-component-patterns, lossless-house-style, plus emerging ones from active studies (study-pattern, profiles-doctype).

The Story

The session opened with a cold question: “how is Pi different from Claude Code?” That answer fit in one Markdown file. Then came the actual interesting question — what could a Pi skill do that we couldn’t do before? — and the rest of the session was the answer unfolding.

flowchart TD
    Q["'How is Pi different<br/>from Claude Code?'"]
    A["First doc:<br/>When-Claud-Code-and-When-Pi.md"]
    B["First skill:<br/>context-vigilance"]
    C["Realization:<br/>skills need to compose"]
    D["astro-knots +<br/>pseudomonorepos"]
    E["Realization:<br/>need a memory<br/>that's not a skill"]
    F["AGENTS.md +<br/>drift policy"]
    G["First changelog<br/>entry (this one)"]
    H["Changelog First<br/>Development theory"]

    Q --> A --> B --> C --> D
    D --> E --> F --> G --> H
    H -.feeds back.-> B

A few moments worth marking:

The folder discovery. While building context-vigilance, a search for prior work surfaced unexpected folders in real context-v/ directories: inquiry/, profiles/, plans/. Rather than force them into the canonical six, we wrote the “When you find a seventh folder” section: don’t fight it, identify the cognitive mode, discuss before assimilating, default to keeping it. The framework stays open instead of closed.

The authorship pivot. Late in the session, we hit the question of credit: when Claude Sonnet writes 80% of a doc, is it an “author”? The answer became a small but real philosophical stance: AI agents augment human authorship; they don’t co-author. New frontmatter convention emerged: authors: is humans only, augmented_with: is for tooling, formatted as <tool> on <model name version> (e.g., Pi on Claude Sonnet 4.5). This entry is the first to use it.

The “memory that isn’t a skill” question. Halfway through, we needed somewhere to encode “observe drift, but don’t auto-fix it” — a rule that should apply across all skills, not live inside one. That’s exactly what Pi’s global AGENTS.md is for. Naming the file changed how we use it: it’s the bedrock layer; skills are the loaded-on-demand layer above it.

Changelog First Development. Writing the first changelog entry made us notice we were doing exactly what we’d later articulate as the methodology: framing the work as a story being told shaped what got built. The arc we’re telling here is the same arc we lived. That’s the whole theory in one paragraph.

How It Works

The skills don’t stand alone — they compose. Here’s the architecture as it stands:

~/.pi/agent/
├── AGENTS.md                  ← loads every session
└── skills/                    ← cloned from lossless-skills repo
    ├── context-vigilance/     ← how we document
    ├── astro-knots/           ← how we build (and what we refuse to build)
    ├── pseudomonorepos/       ← how we navigate (and find prior work)
    ├── changelog-conventions/ ← how we ship-log
    └── changelog/             ← this skills repo also eats its own dog food
        └── 2026-05-04_01.md   ← this entry

Each SKILL.md has a tight description field that controls when it auto-loads — so the agent doesn’t pay token cost for skills it isn’t using. References (deeper docs) load on demand inside a skill. Templates exist for the predictable shapes; the skill teaches when to deviate.

The composition flows like this when working in a Lossless tree:

flowchart LR
    AG[AGENTS.md<br/>defaults] --> P[pseudomonorepos<br/>walks the tree<br/>finds prior work]
    P --> C[context-vigilance<br/>writes new docs<br/>or extends old ones]
    C --> A[astro-knots<br/>refuses bad tech<br/>chooses good]
    A --> CL[changelog-conventions<br/>logs what shipped]
    CL -.story feedback.-> P

What’s Next

The shipped four are the start. The backlog is longer than the shipped list — and intentionally so. Top of the queue:

lossless-loop — the dedicated skill for the 5-phase Start → Progress → Reflect → Publish → Market lifecycle. The diagram and spec already live inside pseudomonorepos/references/lifecycle-workflow.md waiting to graduate.
lfm — Lossless Flavored Markdown patterns. The package is in production at @lossless-group/lfm; the skill makes its capabilities loadable for agents working on content.
A comparative blueprint linking context-vigilance to the active studies on open specs and standards (spec-kit, ai-skills, AGENTS.md, OpenSpec) and memory layers for agents (mem0, neo, statebench). Once those studies converge.

Long-running goal: aggregate every changelog across every Lossless repo via the GitHub API into a “Lossless Changelog” umbrella feed. Entries written in this format will render well in that aggregated context.