Minify context

Skill designgrappler/agent-os/skills/minify-context

Agent OS is a management layer for complex, multi-agent AI projects. Instead of letting agents get lost in endless conversational loops, it enforces a disciplined workflow: upfront planning, shared context, scoped work tracks, and strict quality gates — so every agent stays focused, coordinated, and aligned from start to finish.

Install

npx -y skills add designgrappler/agent-os --skill minify-context

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Compresses verbose active context files into token-efficient form without losing meaning. Includes opt-in memory compression target governed by the R3 per-category taxonomy.

SKILL.md

9.2 KB, as published. Nobody here has run it

Minify Context

Compresses verbose active context files into token-efficient form without losing meaning. Distinct from clean-context (which archives completed items) — this skill compresses active content that has grown wordy.

When to Run

At sprint end, after clean-context has archived completed tracks
Any time context files feel heavy or agent responses seem unfocused
When tracks.md or plan.md has grown through many edits and contains repetitive or padded prose
When memory files have accumulated verbose prose that could be compressed (opt-in — mention "memory" or "memory files" explicitly)

Rules

Preserve-first: Never delete information. Fewer words for the same meaning — not fewer facts.
Structure intact: All headers, field names, tables, and section order remain unchanged.
No archiving: Do not move or delete any content. That is clean-context's job.
Never remove a memory entry — that is clean-context's domain. Compression is preserve-first.

Targets

Default targets (compress all unless the user specifies a subset):

docs/context/plan.md
docs/context/tracks.md
docs/context/product.md (if present)

Memory target (opt-in — off by default): activated only when the user explicitly mentions "memory", "memory files", or "compress memory" when invoking the skill (e.g. "minify memory", "minify context including memory files"). When the memory target runs, see the Memory Compression section below.

Do NOT touch docs/archive/ or .claude/ files (except .claude/projects/<sanitized-cwd>/memory/ when memory target is explicitly opted in).

PROTOCOL_CANON Hard Exclusion

PROTOCOL_CANON path exclusion: AGENTIC.md, claude/agents/architect.md, and claude/templates/AGENTIC.md are excluded from minification regardless of section content. The skill MUST skip these files entirely.

When any of these files is explicitly named as a minification target, respond with exactly:

PROTOCOL_CANON exclusion — <filename> is not eligible for minification.

No further action on those files.

plan.md Active Sprint Exclusion

plan.md Active Sprint exclusion: when minifying docs/context/plan.md, the topmost ## Current Sprint: section is excluded; only ## Completed Sprint: sections are eligible for compression.

Detect the Active Sprint section as: the first ## Current Sprint: heading and its body, up to but not including the first ## Completed Sprint: heading. Preserve this section verbatim. Apply TRANSACTIONAL_STATE compression only to ## Completed Sprint: sections.

Protocol

For each target file:

1. Audit — Read the file. Identify:

Multi-sentence explanations reducible to one
Redundant information repeated across sections
Filler phrases ("In order to...", "It is important to note that...")
Placeholder text never filled in — flag these, don't compress

2. Categorize — Classify each section under the R3 taxonomy (see below) and determine the appropriate compression ratio. Show section-level category classifications in the preview.

3. Preview — Before any rewrite, show the user the proposed changes section by section, with each section's R3 category and target ratio. The user must confirm before any rewrite is applied.

4. Backup — Before applying any rewrite, write a .minify-backup-<date> copy of the original file alongside it (ISO date format: YYYY-MM-DD). Print the backup path in the Minify Report. Backups are kept until manual cleanup — no auto-purge.

5. Rewrite — Produce a compressed version:

Replace verbose prose with tight sentences
Convert narrative paragraphs to bullets where structure is clearer
Keep all specifics: names, commands, file paths, values, dates

6. Report — After writing, show the delta per file:

## Minify Report: [filename]
Category: [R3 category]
Before: [N] lines
After:  [N] lines
Ratio achieved: [N]:1
Target ratio: [range]
Backup: [path to .minify-backup-<date> file]
Flags: [unfilled placeholders or ambiguous content]

R3 Minification Taxonomy

Five categories govern compression behavior. Every section or file is classified into exactly one category before compression begins.

TRANSACTIONAL_STATE — 4:1 to 6:1

Sprint plans, completed-sprint summaries, plan-doc files.

Compress: strip narrative prose, reduce verification checklists to commands only, collapse multi-step Work sections in DONE tracks to one-line outcomes.

Preserve verbatim: track ID, status, files in scope, final commit SHA, key decisions.

Examples: completed sprint entries in plan.md, done tracks in tracks.md, bridge summaries for merged work.

FEEDBACK_RULES — 2:1 maximum

Feedback rule files and any file/section whose primary content is binary constraints.

Compress: drop explanatory framing only.

Preserve verbatim: imperative phrasing ("Never X. Always Y."), canonical example, source/trigger reference.

Hard constraint: never compress feedback rules into prose. The imperative form must survive compression unchanged.

REFERENCE_DOCS — 1.1:1 to 1.3:1

Reference documentation, how-to sections, external resource pointers.

Compress: remove duplicates, outdated links, formatting noise only.

Preserve verbatim: all concrete details — file paths, commands, role names, repo names, URLs.

If compression would force a choice between brevity and a concrete file path, keep the path.

PROTOCOL_CANON — 1:1

Protocol rules, agent definitions, AGENTIC.md content, architecture invariants.

Compress: remove redundancy only per the "one rule, one place" discipline.

Everything else is preserved verbatim. Compression ratio target is 1:1 (no change).

Path exclusion: AGENTIC.md, claude/agents/architect.md, and claude/templates/AGENTIC.md are excluded from minification entirely — see PROTOCOL_CANON Hard Exclusion above.

NARRATIVE_CONTEXT — 8:1 to ∞

Historical narrative, sprint retrospectives, meeting notes, contextual background that is not actively referenced.

Compress: strip to a single sentence summary, or excise entirely.

Move to archive: if the narrative has historical value (e.g. explains a key architectural decision), move the detailed content to docs/archive/ before stripping.

Memory Compression

Opt-in only. Activated when the user explicitly mentions memory when invoking the skill.

Absent-path handling: if ~/.claude/projects/<sanitized-cwd>/memory/ does not exist, print:

memory target requested but no memory directory found at <path> — skipping and continue cleanly.

Classification by filename prefix:

Prefix	R3 Category	Target Ratio
`feedback_*.md`	FEEDBACK_RULES	2:1 max — preserve imperative phrasing + canonical example + source
`reference_*.md`	REFERENCE_DOCS	1.1:1 — preserve all concrete details and file paths
`project_*.md`	TRANSACTIONAL_STATE (if sprint archived) or REFERENCE_DOCS (if sprint active)	4:1–6:1 or 1.1:1 accordingly
No matching prefix	PROTOCOL_CANON	1:1 — do not compress; flag for manual review

project_*.md sprint-archive test: a sprint is considered archived if it appears in docs/context/plan.md's ## Completed Sprint: sections OR in docs/archive/plan-docs/. If the sprint is still active (appears in ## Current Sprint:), treat the file as REFERENCE_DOCS (1.1:1) to preserve current state.

MEMORY.md index compression: when the memory target runs, compress MEMORY.md index lines to the canonical one-line hook form: - [Title](file.md) — one-line hook (max ~120 chars). Verbose multi-line index entries are reduced to this form.

Hard constraints:

Never remove a memory entry — that is clean-context's domain. Every entry present before compression must be present after.
The backup-before-rewrite rule applies to every memory file modified.
Files classified as PROTOCOL_CANON (no reference_ / feedback_ / project_ prefix) are not compressed — they are flagged as "manual review required" in the report.