prompt-compress

Codex compatibility note:

• Invoke repository skills with $skill-name in Codex; this mirrored copy rewrites legacy Claude /skill-name references.
• Task tracker mandate: BEFORE executing any workflow or skill step, create/update task tracking for all steps and keep it synchronized as progress changes.
• User-question prompts mean to ask the user directly in Codex.
• Ignore Claude-specific mode-switch instructions when they appear.
• Strict execution contract: when a user explicitly invokes a skill, execute that skill protocol as written.
• Subagent authorization: when a skill is user-invoked or AI-detected and its protocol requires subagents, that skill activation authorizes use of the required spawn_agent subagent(s) for that task.
• Do not skip, reorder, or merge protocol steps unless the user explicitly approves the deviation first.
• For workflow skills, execute each listed child-skill step explicitly and report step-by-step evidence.
• If a required step/tool cannot run in this environment, stop and ask the user before adapting.

Codex Project-Reference Loading (No Hooks)

Codex does not receive Claude hook-based doc injection. When coding, planning, debugging, testing, or reviewing, open project docs explicitly using this routing.

Always read:

• docs/project-config.json (project-specific paths, commands, modules, and workflow/test settings)
• docs/project-reference/docs-index-reference.md (routes to the full docs/project-reference/* catalog)
• docs/project-reference/lessons.md (always-on guardrails and anti-patterns)

Situation-based docs:

• Backend/CQRS/API/domain/entity changes: backend-patterns-reference.md, domain-entities-reference.md, project-structure-reference.md
• Frontend/UI/styling/design-system: frontend-patterns-reference.md, scss-styling-guide.md, design-system/README.md
• Spec/test-case planning or TC mapping: feature-docs-reference.md
• Integration test implementation/review: integration-test-reference.md
• E2E test implementation/review: e2e-test-reference.md
• Code review/audit work: code-review-rules.md plus domain docs above based on changed files

Do not read all docs blindly. Start from docs-index-reference.md, then open only relevant files for the task.

Quick Summary

Goal: Two-phase optimization of any markdown prompt file: (1) Caveman Compression — aggressively strip stop words and grammatical scaffolding while preserving semantic meaning; (2) Prompt Enhancement — apply AI attention anchoring so AI actually reads and follows all instructions.

Workflow:

1. Read — Read the target file completely
2. Compress — Apply caveman compression pass (Phase 1)
3. Enhance — Apply prompt enhancement transforms (Phase 2)
4. Verify — No content loss, correct structure, rule density ≥ pre-optimization

Key Rules:

• Compress FIRST, enhance SECOND — compression removes noise; enhancement then structures signal
• Preserve meaningful rules, constraints, code examples, and file:line evidence — never remove them
• Post-optimization rule density (MUST ATTENTION/NEVER/ALWAYS per 100 lines) must be ≥ pre-optimization
• Caveman compression applies to prose only — never compress code blocks, YAML, or structured tables
• Prompt quality > token count, but verbose prompts degrade quality — optimize clarity-per-token

---

Target File

Compress and enhance this file: $ARGUMENTS

If no file specified, ask via a direct user question. If text passed instead of a file path, apply caveman compression to the text directly and output the result.

---

Phase 1: Caveman Compression

Aggressively remove stop words and grammatical scaffolding while preserving meaning. Think like a caveman — use only content words that carry semantic weight.

What to Remove

Category	Examples
Articles	a, an, the
Auxiliary verbs	is, are, was, were, am, be, been, being, have, has, had, do, does, did
Redundant prepositions	of, for, to, in, on, at (when meaning stays clear without them)
Pronouns (when context clear)	it, this, that, these, those
Pure intensifiers	very, quite, rather, somewhat, really, extremely

What to Keep (Always)

Category	Reason
All nouns	Core semantic units
All main verbs (not auxiliaries)	Actions carry meaning
All meaningful adjectives	Add semantic signal
Numbers and quantifiers	at least, approximately, more than, 15, many
Uncertainty qualifiers	appears to be, seems, might, what sounded like
Critical prepositions	from, with, without, stuck to — change meaning
Time/frequency words	every Tuesday, weekly, always, never
Names and titles	Dr., Mr., Senator
Technical/domain terms	Never simplify domain language
Negations	not, no, never, without

Preposition Decision Rule

• Keep when defining relationship: made from wood (keep from), stuck to wall (keep to)
• Remove when purely grammatical: system for processing data → system processing data
• Keep in/on/at for location/position: file in /src (keep) vs written in prose (remove)

Compression Examples

Original	Compressed	Removed
"The system was designed to process data efficiently"	"System designed process data efficiently."	The, was, to
"It removes predictable grammar while preserving the unpredictable content"	"Removes predictable grammar preserving unpredictable content."	It, the, while
"There were at least 20 people"	"At least 20 people."	There, were
"Made from wood and metal"	"Made from wood and metal."	nothing — from kept
"This is a method for compressing LLM contexts"	"Method compressing LLM contexts."	This, is, a, for

Compression Scope

Apply to:

• Prose paragraphs and explanatory text
• Bullet point descriptions
• Rule statements (keep imperative verbs)
• Section intros and transitions

Do NOT compress:

• Code blocks (any language)
• YAML frontmatter
• Structured tables (column values may be fragmented — keep as-is)
• file:line references and paths
• <!-- SYNC --> tags and their content
• Frontmatter fields

---

Phase 2: Prompt Enhancement

Applies after compression. Source: Anthropic prompt engineering guide, Stanford "lost-in-the-middle" research, 2025-2026 LLM context optimization studies.

Transform 4: Token Optimization (Conciseness Pass)

After caveman compression, apply structural cleanup:

Cut:

• Filler phrases — "It is important to note that", "Please make sure to" → state the rule directly
• Redundant explanations — if heading says it, body need not re-explain
• Duplicate content — merge sections saying same thing (except intentional top/bottom anchoring)
• Overly verbose examples — trim to minimum lines demonstrating pattern

Keep:

• Code examples with actual file paths/patterns
• Decision tables and lookup references
• Anti-pattern before/after pairs
• All file:line evidence and concrete paths
• Top/bottom anchoring (intentional duplication)

---

Process

Step 1: Read and Analyze

1. Read target file completely
2. Record: current line count, rule density (MUST ATTENTION/NEVER/ALWAYS count)
3. List all READ references → classify as .claude/ (needs inline summary) or docs/ (skip)
4. Identify: missing Quick Summary, missing Closing Reminders, prose-heavy sections

Step 2: Caveman Compression Pass

1. Identify all prose paragraphs and bullet descriptions
2. Apply compression rules — remove stop words, keep semantic content
3. Skip code blocks, YAML, tables, SYNC tags, file paths
4. Verify meaning preserved after each paragraph

Step 3: Create Inline Summaries

For each .claude/ protocol reference:

1. Read the referenced file
2. Extract 2-3 key rules
3. Write blockquote inline summary
4. Keep MUST ATTENTION READ instruction on next line

Step 4: Add/Fix Top Section

• If Quick Summary missing → create from file content
• If present but weak → strengthen with Goal, Workflow, Key Rules
• Protocol summaries appear before Quick Summary

Step 5: Add/Fix Bottom Section

• If Closing Reminders missing → add standard section
• Pick rules AI most commonly skips
• Remove old "IMPORTANT Task Planning Notes" if superseded

Step 6: Verify

Check	Pass Condition
No YAML corruption	Frontmatter intact
No content loss	All rules, code, paths present
Rule density	Post ≥ pre (count MUST ATTENTION/NEVER/ALWAYS)
Line count	Reduced (compression worked)
Formatting	Blank lines between sections, headers correct
READ classification	.claude/ → inline summary, docs/ → skipped

---

[IMPORTANT] Use task tracking to break ALL work into small tasks BEFORE starting.

Output Quality — Token efficiency without sacrificing quality.

1. No inventories/counts — AI can grep | wc -l. Counts go stale instantly
2. No directory trees — AI can glob/ls. Use 1-line path conventions
3. No TOCs — AI reads linearly. TOC wastes tokens
4. No examples that repeat what rules say — one example only if non-obvious
5. Lead with answer, not reasoning. Skip filler words and preamble
6. Sacrifice grammar for concision in reports
7. Unresolved questions at end, if any

Context Engineering Principles — Research-backed principles for prompt quality. Source: Anthropic prompt engineering guide, Stanford "lost-in-the-middle" research, 2025-2026 LLM context optimization studies.

1. Primacy-Recency Effect — LLM performance drops 15-47% for middle-context information (Stanford). AI attention peaks at first/last 10% of text. Action: Place the 3 most critical rules in both the first 5 lines AND the last 5 lines of every prompt. Queries at end improve quality by up to 30% (Anthropic).
2. High-Signal Density — Anthropic: "Identify the smallest collection of high-signal tokens that maximize the probability of the desired outcome." Action: Every line should change AI behavior. If removing a line doesn't change output → cut it. Target ≥8 rules (MUST ATTENTION/NEVER/ALWAYS) per 100 lines.
3. Context Rot — LLM performance degrades as context length grows — even when all content is relevant. Compression (5-20x) maintains or improves accuracy while saving 70-94% tokens. Action: Compress aggressively. Shorter, denser prompts outperform longer, diluted ones.
4. Structured > Prose — Tables, bullets, XML/markdown parse faster than paragraphs. Constrained formats reduce error rates vs free-text. Action: Convert narrative to tables/bullets. Use markdown headers for semantic sections.
5. RCCF Framework — Modern LLMs (2025+) already know how to reason. What they need: Role (personality), Context (grounding), Constraints (guardrails), Format (structure). Constraints and format matter more than verbose instructions.
6. Checkbox Avoidance — [ ] syntax triggers mechanical compliance — AI ticks boxes without reasoning. Bullet rules force reading and evaluation. Action: Replace - [ ] Check X with - MUST ATTENTION verify X.
7. Example Economy — 3-5 examples optimal for few-shot; diminishing returns after. Action: 1 best example per pattern. Use BAD→GOOD pairs (2-3 lines each) for anti-patterns.
8. Deferred Tool Loading — Claude Code delays loading tool definitions when they exceed 10% of context window. Action: Keep injected docs well under 10% of context budget. Docs exceeding ~3,000 lines are too large for injection — split or compress.
9. Rule Density Verification — Post-optimization rule count (MUST ATTENTION/NEVER/ALWAYS) must be ≥ pre-optimization count. Compression should preserve or increase density, never decrease it. Action: Count before and after every optimization pass.
10. Affirmative Directives — Models comply with affirmative directives more reliably than prohibitions; a bare "don't X" leaves the correct action unspecified, so the model substitutes an arbitrary alternative. Action: State the action to take, not only the action to avoid. Keep NEVER/forbidden guardrails for hard invariants — but pair each with the right path ("Do X" not just "Don't do Y").
11. Rationale-Carrying Instructions — A rule shipped with its reason generalizes to edge cases the rule never enumerated and survives compression; a bare imperative gets misapplied or silently dropped. Action: Append a terse — why: … clause to every non-obvious rule. The reason names the failure prevented or outcome wanted — never restates the rule.

Prompt Enhancement Transforms (Base) — Transforms 1-3 are identical across prompt-enhance / prompt-compress / prompt-expand. Transform 4 is per-skill (conciseness pass for enhance/compress; structural clarity pass for expand) and stays local to each skill.

Transform 1: Inline Summaries for READ References

Problem: AI sees MUST ATTENTION READ file.md and skips it. Solution: Add a 2-3 line summary of key rules BEFORE the read instruction.

Before:

MUST ATTENTION READ .claude/protocols/evidence.md

After:

> **Evidence-Based Reasoning** — Speculation is FORBIDDEN. Every claim requires `file:line` proof.
> Confidence: >95% recommend freely, 80-94% with caveats, <80% DO NOT recommend.

MUST ATTENTION READ .claude/protocols/evidence.md for full details.

Scope rules:

• .claude/ protocol files → always add an inline summary (stable, belongs to framework)
• docs/project-reference/ files → NO inline summary (project-specific). Add: (read directly when relevant; do not rely on hook-injected conversation text)

Transform 2: Top Summary Section

Required structure (first 20 lines after frontmatter):

> **[IMPORTANT]** task tracking instruction...

> **Protocol Name** — [inline summary]. MUST ATTENTION READ `path` for details.

## Quick Summary

**Goal:** [One sentence — what this skill achieves]

**Workflow:**

1. **[Step]** — [description]

**Key Rules:**

- [Most critical constraint]

Transform 3: Bottom Closing Reminders

Add at the very end of the file:

---

## Closing Reminders

**IMPORTANT MUST ATTENTION** [echo rule #1 from the top section]
**IMPORTANT MUST ATTENTION** [echo rule #2]
**IMPORTANT MUST ATTENTION** [echo rule #3]
**IMPORTANT MUST ATTENTION** add a final review task to verify work quality

Pick 3-5 rules AI most commonly violates. Bottom section re-anchors attention after the long middle.

Shared Protocol Duplication Policy — Inline protocol content in skills (wrapped in <!-- SYNC:tag -->) is INTENTIONAL duplication. Do NOT extract, deduplicate, or replace with file references. AI compliance drops significantly when protocols are behind file-read indirection. To update: edit .claude/skills/shared/sync-inline-versions.md first, then grep SYNC:protocol-name and update all occurrences.

AI Mistake Prevention — Failure modes to avoid on every task:

Check downstream references before deleting. Deleting components causes documentation and code staleness cascades. Map all referencing files before removal. Verify AI-generated content against actual code. AI hallucinates APIs, class names, and method signatures. Always grep to confirm existence before documenting or referencing. Trace full dependency chain after edits. Changing a definition misses downstream variables and consumers derived from it. Always trace the full chain. Trace ALL code paths when verifying correctness. Confirming code exists is not confirming it executes. Always trace early exits, error branches, and conditional skips — not just happy path. When debugging, ask "whose responsibility?" before fixing. Trace whether bug is in caller (wrong data) or callee (wrong handling). Fix at responsible layer — never patch symptom site. Assume existing values are intentional — ask WHY before changing. Before changing any constant, limit, flag, or pattern: read comments, check git blame, examine surrounding code. Verify ALL affected outputs, not just the first. Changes touching multiple stacks require verifying EVERY output. One green check is not all green checks. Holistic-first debugging — resist nearest-attention trap. When investigating any failure, list EVERY precondition first (config, env vars, DB names, endpoints, DI registrations, data preconditions), then verify each against evidence before forming any code-layer hypothesis. Surgical changes — apply the diff test. Bug fix: every changed line must trace directly to the bug. Don't restyle or improve adjacent code. Enhancement task: implement improvements AND announce them explicitly. Surface ambiguity before coding — don't pick silently. If request has multiple interpretations, present each with effort estimate and ask. Never assume all-records, file-based, or more complex path.

Critical Thinking Mindset — Apply critical thinking, sequential thinking. Every claim needs traced proof, confidence >80% to act. Anti-hallucination: Never present guess as fact — cite sources for every claim, admit uncertainty freely, self-check output for errors, cross-reference independently, stay skeptical of own confidence — certainty without evidence root of all hallucination.

MUST ATTENTION apply critical thinking — every claim needs traced proof, confidence >80% to act. Anti-hallucination: never present guess as fact.

MUST ATTENTION apply AI mistake prevention — holistic-first debugging, fix at responsible layer, surface ambiguity before coding, re-read files after compaction.

Closing Reminders

IMPORTANT MUST ATTENTION apply caveman compression FIRST before any structural enhancement — never skip Phase 1 IMPORTANT MUST ATTENTION never compress code blocks, YAML frontmatter, structured tables, or SYNC tags IMPORTANT MUST ATTENTION verify rule density post-compression ≥ pre-compression — compression must not dilute signal IMPORTANT MUST ATTENTION state the action to take, not only what to avoid — pair every NEVER with the right path, and append a terse — why: to each non-obvious rule — why: affirmative directives + carried rationale are followed more reliably and survive compression (principles #10/#11) IMPORTANT MUST ATTENTION apply primacy-recency anchoring — 3 critical rules in first 5 AND last 5 lines of every enhanced file IMPORTANT MUST ATTENTION add inline summaries only for .claude/ protocol files, never for docs/ project files IMPORTANT MUST ATTENTION cite file:line evidence for every claim (confidence >80% to act). NEVER speculate without proof. IMPORTANT MUST ATTENTION READ CLAUDE.md before starting

[TASK-PLANNING] Before acting, analyze task scope and systematically break it into small todo tasks and sub-tasks using task tracking.

Hookless Prompt Protocol Mirror (Auto-Synced)

Source: .claude/hooks/lib/prompt-injections.cjs + .claude/.ck.json

[WORKFLOW-EXECUTION-PROTOCOL] [BLOCKING] Workflow Execution Protocol — MANDATORY IMPORTANT MUST CRITICAL. Do not skip for any reason.

Generic portability boundary: Reusable skills and protocol text stay project-neutral; project-specific conventions are discovered from docs/project-config.json and docs/project-reference/. Apply shared AI-SDD from shared/sdd-artifact-contract.md. Read docs/project-config.json and docs/project-reference/docs-index-reference.md, then open the project reference docs named there. Any supported AI tool may execute when this shared context and local docs are available.

1. DETECT: Match prompt against workflow catalog
2. ANALYZE: Find best-match workflow AND evaluate if a custom step combination would fit better
3. ASK (REQUIRED FORMAT): Use a direct user question with this structure unless the user explicitly invoked a workflow/skill and the local protocol treats explicit invocation as confirmation:
   • Question: "Which workflow do you want to activate?"
   • Option 1: "Activate [BestMatch Workflow] (Recommended)"
   • Option 2: "Activate custom workflow: [step1 → step2 → ...]" (include one-line rationale)
4. ACTIVATE (if confirmed): Call $workflow-start <workflowId> for standard; sequence custom steps manually
5. CREATE TASKS: task tracking for ALL workflow steps
6. EXECUTE: Follow each step in sequence [CRITICAL-THINKING-MINDSET] Apply critical thinking, sequential thinking. Every claim needs traced proof, confidence >80% to act. Anti-hallucination principle: Never present guess as fact — cite sources for every claim, admit uncertainty freely, self-check output for errors, cross-reference independently, stay skeptical of own confidence — certainty without evidence root of all hallucination. AI Attention principle (Primacy-Recency): Put the 3 most critical rules at both top and bottom of long prompts/protocols so instruction adherence survives long context windows. Goal-driven execution: Define success criteria first, loop until verified, and stop only when observable checks pass. Tests verify intent: Tests must protect business rules/invariants and fail when the protected intent breaks, not only mirror current behavior.

[LESSON-LEARNED-REMINDER] [BLOCKING] Task Planning & Continuous Improvement — MANDATORY. Do not skip.

Break work into small tasks (task tracking) before starting. Add final task: "Analyze AI mistakes & lessons learned".

Extract lessons — ROOT CAUSE ONLY, not symptom fixes:

1. Name the FAILURE MODE (reasoning/assumption failure), not symptom — "assumed API existed without reading source" not "used wrong enum value".
2. Generality test: does this failure mode apply to ≥3 contexts/codebases? If not, abstract one level up.
3. Write as a universal rule — strip project-specific names/paths/classes. Useful on any codebase.
4. Consolidate: multiple mistakes sharing one failure mode → ONE lesson.
5. Recurrence gate: "Would this recur in future session WITHOUT this reminder?" — No → skip $learn.
6. Auto-fix gate: "Could $code-review/$code-simplifier/$security/$lint catch this?" — Yes → improve review skill instead.
7. BOTH gates pass → ask user to run $learn. [TASK-PLANNING] [MANDATORY] BEFORE executing any workflow or skill step, create/update task tracking for all planned steps, then keep it synchronized as each step starts/completes.