zipai-optimizer – skill Agent & Tự động hóa cho agent AI

ZipAI: Context & Token Optimizer

When to Use

Use this skill when the request needs context-window-aware triage, prompt caching optimizations, concise technical output, ambiguity handling, or selective reading of logs, source files, JSON/YAML payloads, VCS output, or MCP tool results.

Rules

Rule 1 — Adaptive Verbosity (No Filler)

• Fixes: technical only. ZERO filler (e.g., "Certainly", "I understand", "Here is", "Sure").
• Analysis: provide concise conclusions, assumptions and supporting evidence; do not request or reveal private hidden reasoning.
• Direct Ask: max 15 words in ultra-dense telegraphic style. Omit grammatical helper constructs.
• Long Sessions: never re-summarize past thread context.
• Reviews: use structured headers: [ISSUE], [SUGGESTION], [NITPICK].

Rule 2 — Ambiguity-First Execution

• Ask exactly ONE question if 2+ interpretations exist. Never stack questions.
• Default to minimal intervention for minor changes.
• Scope ambiguous requests to narrowest boundary.

Rule 3 — Prompt Caching & Prefix Stability

• Static-First Ordering: Structure prompts to place invariant components (system instructions, core rules, static tool schemas) at the top of the prompt.
• Isolate Dynamic Context: Append dynamic and volatile elements (active conversation history, recently read file contents, CLI execution outputs) at the very end of the prompt to protect and reuse the cached prefix.
• Prefix Integrity: Avoid interleaving new queries or dynamic variables inside static system blocks. Keep the static instructions strictly invariant.
• Cached Files Reuse: Reuse already loaded file contents present in the conversation history; do not re-read files unless explicitly updated.

Rule 4 — Semantic Input Pruning & Log Compression

• Traceback Extraction: When handling error or build outputs, parse and filter logs using grep/regex to extract only tracebacks, error statements, and a maximum of 3-5 lines of context around them. Strip all info logs, successful build tasks, and redundant progress messages.
• Skeletal Code Viewing (AST): For large files (>300 lines), locate relevant declarations first; do not skip required bodies, error paths or supporting files. Use grep -nE "^(class|def|async def|function|const|let|var).*=" (or language equivalents) to view class and function headers first, then target specific ranges with view_file.
• Smart JSON/YAML Crusher: Minify structured inputs. Strip pretty-printing whitespaces, comments, and unused fields from JSON/YAML payloads before placing them in context. Convert large arrays to dense CSV or key-value listings if they are queried.

Rule 5 — Surgical & Compact Output

• Local Replacements: Perform edits using surgical tools (str_replace or single-hunk diffs). Never reprint unchanged surrounding code or perform full-file reprints.
• Batch Modifies: Consolidate multiple non-contiguous edits in a single file into a single multi-replace chunk operation, ordered from leaf dependencies upward.
• Differential Output: Limit conversational responses to the exact modified blocks, avoiding conversational code repetition.

Rule 6 — Telegraphic Grammar & Density

• Syntax Compression: Strip articles ("a", "an", "the"), redundant helper verbs ("to be", "to have", "do"), and politeness/softening modifiers ("please", "simply", "just", "easy").
• Structure: Format output blocks into dense semantic mappings (key: val), short bullet lists, and compact tables. Avoid paragraphs of text.

Rule 7 — Token-Budget Reasoning (CoT Optimization)

• Direct Mode: Skip long planning/thinking cycles for trivial, deterministic edits (typos, formatting, import adjustments).
• Abbreviated Thoughts: Keep thought blocks compact. Never reprint code snippets or copy-paste file blocks inside thoughts. Reference files via path and lines (e.g. file.py#L12-18).

---

Negative Constraints

• No filler: "Here is", "I understand", "Let me", "Great question", "Certainly", "Of course", "Happy to help".
• No blind truncation of stacktraces or error logs.
• Use navigation to target relevant content, but perform complete-file/tree reads whenever review scope, correctness or policy requires them; pruning is not evidence of review.
• Reuse unchanged verified content; reread when it changed, was truncated or relevant context is uncertain.
• No multi-question clarification dumps.
• No silent bundling of unrelated changes.
• No full git diff ingestion on large changesets — extract hunks only.
• No git log beyond 20 entries unless a specific range is requested.
• No full MCP object inspection when field-level access suffices.
• No MCP mutations without prior read of current resource state.
• No SHA reuse across sessions for file updates.

---

Example

User request:

> Use @zipai-optimizer for this task: Ultra-dense token optimizer skill for prompt caching, log pruning, AST-based inspection, and minified JSON payloads.

Limitations

• Brainstorming: disable during creative/open-ended design phases.
• Grep Blindness: key context may fall outside filter boundaries.
• Overshadowing: aggressive pruning may drop micro-variables in long sessions.