Compress Prompt

Compress a KB Prompt

Mental model

• The artifact you compress is loaded into a coding agent that runs in ANOTHER repo, on ANOTHER user's task. Every token becomes that agent's context → scaffolding = distraction that dilutes focus.
• Compress ≠ shrink words. Compress = strip scaffolding, keep 100% signal, sharpen focus, replace with meaningful unicode characters, like arrows.
• Reader is AI like you, capable: it already knows the domain AND the KB grammar. Use terms and acronyms. Don't explain — nudge.
• CAPS = importance. Word count is an OUTCOME, never a target.
• INVARIANT ≠ STEP. Invariants (always-on constraints) → declare ONCE, flag always-on. Steps → ordered, run once. NEVER re-assert an invariant as per-step reminders — that's the echo authors reflexively add; compress hoists + cuts it.
• GOLDEN RULE: NEVER trade a high-value token for a few saved words (unless it is repeatedly used and overall gives high results, we can loose 2% of value overall).
• BULLETS vs ORDERED: ALWAYS convert bullets to ordered lists, if work is sequential or can be sequential. Reason: aligns with AI sequential token generation.
• DENSIFY EVERY rule.
• Take time to think during reasoning, take different options, iterate multiple times TRANSFORM.
• Your task is MAXIMUM compression, not just low hanging fruits!
• NO rush, TAKE time
• Don't bring this skill terms or meta-thinking
• Identify what is load bearing

Allowed reads — read-only, NEVER adjust

Read ONLY: the target artifact + its type schema + the grammar below. Nothing else. Stay focused.

Schemas (learn which XML scopes are MANDATORY vs optional, and each scope's role):

• docs/schemas/skill.md
• docs/schemas/workflow.md
• docs/schemas/phase.md
• docs/schemas/rule.md
• docs/schemas/agent.md
• docs/schemas/template.md
• docs/schemas/generic.md

Grammar — directive commands the system ACTS ON; protect verbatim + their args:

KEEP verbatim (never shrink / drop)

• MANDATORY scope tags + nesting — structure is signal. OPTIONAL scopes EARN keep (load-bearing audit).
• Grammar commands above + their args: file / skill / tool / model names, paths, section anchors.
• CAPS importance markers: MUST · NEVER · DO NOT · HALT · WAIT · SELF-CHECK · HITL …
• Per-scope / per-step instructions, kept IN their scope (e.g. update-state, gate notes).
• Semantic distinctions: required vs recommended, blocking vs optional, default vs conditional.

CUT — where real reduction lives

• Tautology → rule stated >1× across scopes; keep ONE authoritative copy, kill the echoes.
• Pointer-echo → info already reachable via a named cite (invariant · ## scope · file · skill) → NEVER re-assert or re-summarize it inline. Invariants → hoist to ONE always-on block; other echoes → cut. The pointer IS the content.
• Meta-commentary explaining the prompt's own notation / convention to a reader.
• Stale / orphaned items → reference a scheme, attribute, or value no longer present.
• WHOLE-SCOPE echo → CUT the scope, not just its lines.
• LOAD-BEARING test: delete scope → agent acts differently? No ⇒ cut.
• Audit EACH scope, esp. references · best_practices · validation · pitfalls.
• Keep ONLY signal unreachable elsewhere in-file / via cite.
• Mandatory-but-echo scope → shrink to minimum unique nugget.
• Lone nugget → hoist to load-bearing home, drop wrapper.
• Repeated literals → define once as a short alias (e.g. OUT/ = <long/path>), reuse everywhere.
• Cut the fluff.
• Restated the same thing in different ways.
• Everything obvious or already known by AI (keep only terms/nudges!).

COMPRESS

• HARD CAP: every rule / bullet line < 10 words.
• Group same-topic rules, merge, rephrase clearly, output as separate.
• NO new lines as escape hatch.
• Whole file: MAY add ≤ 10 lines total.
• NEVER drop signal to hit the cap.
• Verbose prose / step-narration the agent already infers → terse cue. e.g. "ONE PHASE AT A TIME: read file, execute, update state, advance" → "ONE PHASE AT A TIME. READ JIT."
• Favor unicode connectives for density: → · ⇒ ≠ ± … (English words only otherwise).
• DENSE, USE TERMS, ACRONYMS, TERSE-phrases (not sentences!)
• NUDGE using single words for ACTIONS, ASPECTS, THINKING, GOALS, REASONS, etc.