For Sage brand work in Pi, use the paste-ready prompt at docs/prompts/pi-sage-brand-gen-full-pipeline.md. Keep this link instead of copying the full prompt into skill bodies.

Brand-Gen Truncation Audit

Risk addressed: prompt assembly applies caps through 6+ layers in sequence (per-part cap_text_at_sentence → joined-prelude cap → compress_prompt_body → evict_to_budget over PromptBlocks → execution-prompt re-cap → ad-hoc [:N] slices). Most truncation is silent — only review_prompt_architecture reports dropped blocks. Hard-coded limits live as magic numbers across 14+ files. The priority-tier PromptBlock machinery (prompt_block.py) only sees content after earlier caps already fired.

What this skill produces

A single JSON-shaped manifest of cap sites. Each entry:

{
  "site_id": "prompt_assembly.py:387",
  "kind": "cap_text_at_sentence | compress_prompt_body | evict_to_budget | hard_slice | summarize | strip",
  "stage": "per_part | combined_prelude | body | execution | post_block_eviction",
  "input_field": "brand_anchor_rule | inspiration_directive | copy_bank | …",
  "limit_value": 480,
  "limit_source": "constant:NON_INTERFACE_BODY_CAP | json:data/prompt_budget.json#interface | literal:480 | literal:[:5]",
  "configurable": true,
  "tracked": true,
  "tracked_via": "dropped_blocks | recommendation_string | none",
  "fires_before_priority_system": false,
  "snippet": "cap_text_at_sentence(brand_anchor_rule, NON_INTERFACE_BODY_CAP)"
}

Plus a summary: total sites, silent vs tracked ratio, sites firing before evict_to_budget, hard-coded literal counts by file.

When to use

• Before adjusting data/prompt_budget.json — to know which caps actually respect it.
• Before promoting a new PromptBlock priority — to know which caps would defeat it.
• When iteration_memory.json keeps losing entries that the agent expected to survive.
• When the rendered prompt visibly drops a section the planner asked for but dropped_blocks is empty.
• When introducing a new prompt section, to choose where its cap should live.

Inputs

• Repo path (default: $PWD, must contain brand_gen/prompt_assembly.py).
• Optional --filter — restrict to one of: interface, non_interface, card, inspiration, copy.
• Optional --budget-source — path to data/prompt_budget.json (default location auto-detected).

Procedure

1. Locate the cap primitives. Read brand_gen/prompt_assembly.py for cap_text_at_sentence, compress_prompt_body, _get_budget, _DEFAULT_BUDGET. Read brand_gen/prompt_block.py for evict_to_budget, SECTION_PRIORITY, SECTION_STAGE, SECTION_CONSTRAINT. Read brand_gen/card_text.py for _truncate_multiline_copy. Note the function signatures and what each emits as side-effect (returned struct, appended dropped_blocks, logged warning, silent).

1. Enumerate every call site for each primitive. Use Grep for the function names. For each match, capture file:line, the args, the surrounding stage label (per-part vs combined-prelude vs body vs post-block-eviction).

1. Enumerate hard `[:N]` slices on prompt content. Grep prompt_assembly.py, reference_role_packs.py, inspiration_extraction.py, card_builder.py, card_engine.py, sage_generation_contract.py for patterns: \[:\d+\], \[:[A-Z_]+\], \.strip\(\)\[:, \.split[^]]*\)\[:. Filter to those operating on prompt-bearing strings (skip list-of-Tuple slices for typed records).

1. Enumerate raw character truncations. Grep for truncate, max_len, max_chars, summary, condense, [:1200], [:2000], …. For card flows, capture (max_lines, max_chars) tuples in card_engine.py:762,1158,1195,1224,1304.

1. Classify each site:

• kind — which primitive (or hard_slice for raw [:N]).
• stage — read 30 lines of context to decide whether this fires inside build_effective_prompt per-part, combined-prelude, body-cap, after evict_to_budget, or inside the execution-prompt builder.
• limit_value, limit_source — trace the constant. Resolve NON_INTERFACE_BODY_CAP etc. to its _DEFAULT_BUDGET entry and check whether data/prompt_budget.json overrides it.
• tracked — does the output struct of the enclosing function include the drop? Check review_prompt_architecture (prompt_assembly.py:1391-1560) for what gets added to dropped_blocks / recommendations. Anything outside that surface is silent.
• fires_before_priority_system — true if it executes before evict_to_budget(blocks_from_sections(...), budget) runs in build_execution_prompt. Per-part caps in build_effective_prompt are ALL true.

1. Cross-check against `SECTION_PRIORITY`. For each tracked section name in prompt_block.py, list every cap that could touch its content before the priority system gets to weigh it. A brand_anchor_rule (priority 0, hard) silently truncated by a per-part [:N] is the highest-severity finding.

1. Emit manifest. Format as JSON to stdout (or --output <path>). Print summary table:

• Total sites, silent count, tracked count.
• Sites firing before priority system, by section name.
• Hard-coded literal limits (count + file).
• Sites operating on hard-constraint sections (worst risk).

Output schema

{
  "summary": {
    "total_sites": 142,
    "silent": 118,
    "tracked": 24,
    "fires_before_priority_system": 87,
    "hard_section_silent_caps": 11,
    "files_with_literal_limits": 14
  },
  "sites": [ /* entries shaped like the example above */ ],
  "section_coverage": {
    "brand_anchor_rule": { "priority": 0, "constraint": "hard", "pre_priority_caps": 3, "tracked_caps": 1 }
  },
  "duplicate_primitives": [
    { "primitive": "sentence-aware truncation", "implementations": ["prompt_assembly.cap_text_at_sentence", "card_text._truncate_multiline_copy"] }
  ]
}

Reference files

• brand_gen/prompt_assembly.py (lines 387-425, 1391-1560) — compression hot zone.
• brand_gen/prompt_block.py — priority/eviction model, the source of truth this audit verifies.
• brand_gen/card_engine.py (lines 762, 1158, 1195, 1224, 1304) — card-side truncation regime.
• brand_gen/card_text.py:660 — duplicate _truncate_multiline_copy.
• data/prompt_budget.json — externalized limits; verify which caps actually read it.

Don't

• Don't propose fixes inside this skill. The output feeds brand-gen-architecture-pass orchestrator and compound-engineering:ce-plan for remediation planning.
• Don't edit prompt_assembly.py while running this audit — the manifest is a snapshot.
• Don't conflate [:N] on a List of typed records (fine) with [:N] on prompt content (lossy).