dleerdefi/claude-code-construction/rfi-drafter

RFI Drafter

Purpose

Two responsibilities: (1) draft formal RFIs from identified issues, and (2) manage the issue detection registry where potential problems surface during other skill workflows. Detection is ambient. Drafting is user- triggered. No RFI is ever created without explicit user instruction.

Design: RIGID output format and quality checks. GUIDED research sequence before drafting. FLEXIBLE across any trade, CSI division, or document conflict type.

Does NOT: autonomously discover issues and write RFIs, send RFIs without user review, or determine whether an issue is actually an error vs. intentional design.

Permitted Scripts

Run every script through ${CLAUDE_PLUGIN_ROOT}/bin/construction-python, which selects the toolkit's Python on Windows, macOS and Linux. Never call bare python.

Do NOT create custom scripts during execution. All output goes through the scripts above.

Mode 1: User-Initiated RFI Drafting

The user identifies an issue and instructs Claude to draft an RFI. This is the primary workflow — a 30-minute task compressed to minutes.

Step 0: Gather RFI Template (first RFI per project)

Before the first RFI, ask the user for their firm's RFI form template. Construction teams have a branded template they use for all official RFIs — the output MUST match their format exactly.

If .docx template provided:

1. Read with python-docx to identify field locations (table cells,

placeholders like [PROJECT NAME], content controls)

1. Build a field mapping JSON linking each RFI data field to its

location in the template

1. Store mapping at .construction/skills/rfi-drafter/rfi_template_map.json
2. Store a SHA-256 hash of the template for change detection

If PDF template provided:

1. Rasterize each page at 200 DPI
2. Read with vision to understand field layout and structure
3. Build the mapping JSON from visual analysis
4. Store at .construction/skills/rfi-drafter/rfi_template_map.json

If no template available: Fall back to the generic format built into rfi_export.py. Inform the user that providing their template is recommended for production RFIs.

On subsequent RFIs, reuse the stored mapping. If the template file hash differs from the stored hash, re-map.

Step 1: Understand the Issue

Gather from the user (or from an issue registry record if escalating):

• What: the conflict, ambiguity, or missing information
• Where: sheet number(s), grid location, room/area, detail ref
• Between what: which documents disagree (plan vs detail, arch vs

structural, drawing vs spec, etc.)

If the user's description is vague, ask clarifying questions. Never draft from ambiguous input — an RFI that doesn't clearly state the problem wastes everyone's time.

Step 2: Research Context

Before drafting, systematically gather evidence. Read references/research-checklist.md for the full checklist.

Minimum research before any RFI:

1. Read the specific area on the source sheet(s)
2. Check related details/sections for the same condition
3. Check the relevant spec section for requirements
4. Check if general notes address the condition
5. Check addenda and ASIs for superseding changes
6. Check existing RFI log for duplicates (if available)

Read the sheet by rasterizing it and cropping to the conflict zone (the ${CLAUDE_PLUGIN_ROOT} paths below are substituted only here in SKILL.md, not in the reference files):

mkdir -p ".construction/skills/rfi-drafter/pages"
"${CLAUDE_PLUGIN_ROOT}/bin/construction-python" "${CLAUDE_PLUGIN_ROOT}/scripts/pdf/rasterize_page.py" "{pdf}" {page} --dpi 200 --output ".construction/skills/rfi-drafter/pages/page.png"
"${CLAUDE_PLUGIN_ROOT}/bin/construction-python" "${CLAUDE_PLUGIN_ROOT}/scripts/pdf/crop_region.py" ".construction/skills/rfi-drafter/pages/page.png" --box x1,y1,x2,y2 --normalized --output ".construction/skills/rfi-drafter/pages/conflict_area.png"

Page images are working data: run these from the project folder and keep them under .construction/skills/rfi-drafter/pages/, never in the project root, never in /tmp or any other folder outside the project.

{page} is 1-based: 1 for a split sheet PDF, page_index + 1 for a page in a bound set (sheet_index.yaml counts page_index from 0). The box is 0–1 fractions of the page.

First read .construction/project.yaml: present and readable means AgentCM manages the folder. A listing that shows the file does not decide it: read it. In AgentCM mode, read its files for cross-references to the affected area. See references/research-checklist.md for which files, and follow ${CLAUDE_PLUGIN_ROOT}/reference/agentcm-mode.md (which source wins, missing files, drawings through crop.sh, corrections of AgentCM's data).

If research reveals the issue is already resolved (by addendum, ASI, or existing RFI response), inform the user — no RFI needed.

Step 3: Draft the RFI

Use the format in references/rfi-format.md. Every RFI must include:

• Project identification (name, number, RFI number)
• Addressee (the design professional who owns the document)
• Specific drawing/spec references with grid locations
• Clear description of the issue
• Suggested resolution (always — accelerates response)
• Schedule impact statement
• Attachment list

Read references/rfi-format.md for the full template, field-by-field guidance, and an example RFI.

Step 4: Quality Check

Before presenting to user, verify against the checklist in references/quality-checks.md. Key gates:

• Specific sheet numbers and grid locations cited (never "the plan")
• Issue clearly described (a stranger could understand it)
• Suggested resolution provided
• Schedule impact stated with a needed-by date
• Spec section referenced if applicable
• No duplicate of an existing RFI
• Addressee is the correct design professional for this scope

Step 5: Present for Review

Present the draft RFI to the user in conversation. The user may:

• Approve and export — populate their template and generate .docx
• Edit and refine — iterate on language, add context
• Reject — issue isn't actually an issue
• Defer — log it back to the issue registry for later

On export, write the RFI data with the Write tool to .construction/skills/rfi-drafter/rfi_draft.json (working data, never the project root), then invoke from the project folder; the .docx is the deliverable and goes in the project root (or the folder the PE names), never under .construction/:

# Template mode (preferred — matches firm's format):
"${CLAUDE_PLUGIN_ROOT}/bin/construction-python" "${CLAUDE_SKILL_DIR}/scripts/rfi_export.py" \
  --template "path/to/firm_rfi_form.docx" \
  --mapping .construction/skills/rfi-drafter/rfi_template_map.json \
  --data .construction/skills/rfi-drafter/rfi_draft.json \
  --output RFI-026.docx
 
# Generic mode (fallback — no template):
"${CLAUDE_PLUGIN_ROOT}/bin/construction-python" "${CLAUDE_SKILL_DIR}/scripts/rfi_export.py" \
  --data .construction/skills/rfi-drafter/rfi_draft.json \
  --output RFI-026.docx

Output is always .docx so the user can make final edits before sending. The script uses safe_output_path() — never overwrites existing files.

Mode 2: Issue Detection Registry

Other skills (tag-audit-and-takeoff, pe-review, submittal-review, etc.) surface potential issues during their normal work. These issues accumulate in .construction/skills/issues/ as JSON records, NOT as RFIs.

How Issues Get Created

Any skill can write an issue record. Read references/issue-schema.md for the full schema. An issue record contains:

• Source skill that found it
• Severity (info / warning / conflict / safety)
• Description of what was observed
• Affected sheets, spec sections, elements
• Confidence level
• Timestamp

Critical rule: no skill writes an RFI directly. Skills write issue records. The user reviews the queue. The user decides which issues become RFIs.

Reviewing the Issue Queue

When the user asks to review issues ("check the issues list", "what issues have been found", "any problems detected"):

1. Load issues from .construction/skills/issues/
2. Sort by severity (safety > conflict > warning > info), then

by confidence (high > medium > low)

1. Present a summary table: severity, source skill, description,

affected sheets, confidence

1. For each issue the user wants to escalate: switch to Mode 1

drafting with the issue record as input context

# List all open issues (sorted by severity, then confidence)
"${CLAUDE_PLUGIN_ROOT}/bin/construction-python" "${CLAUDE_PLUGIN_ROOT}/scripts/issue_manager.py" list
 
# List with human-readable table format
"${CLAUDE_PLUGIN_ROOT}/bin/construction-python" "${CLAUDE_PLUGIN_ROOT}/scripts/issue_manager.py" list --table
 
# Filter by severity or source skill
"${CLAUDE_PLUGIN_ROOT}/bin/construction-python" "${CLAUDE_PLUGIN_ROOT}/scripts/issue_manager.py" list --severity conflict
"${CLAUDE_PLUGIN_ROOT}/bin/construction-python" "${CLAUDE_PLUGIN_ROOT}/scripts/issue_manager.py" list --source-skill "pe-review"
 
# Get a specific issue
"${CLAUDE_PLUGIN_ROOT}/bin/construction-python" "${CLAUDE_PLUGIN_ROOT}/scripts/issue_manager.py" get --id ISS-2026-0001
 
# Escalate an issue to RFI
"${CLAUDE_PLUGIN_ROOT}/bin/construction-python" "${CLAUDE_PLUGIN_ROOT}/scripts/issue_manager.py" update \
  --id ISS-2026-0001 --status escalated --rfi-number RFI-026
 
# Summary statistics
"${CLAUDE_PLUGIN_ROOT}/bin/construction-python" "${CLAUDE_PLUGIN_ROOT}/scripts/issue_manager.py" stats

Writing Issues from Other Skills

Other skills log issues without loading this skill: the issue_manager.py add command, the severity mapping and the pattern vocabulary are in construction-guide's Issue Registry paragraph. This skill only reads the registry and escalates.

Output

RFI documents are produced as Word (.docx) via scripts/rfi_export.py. Two modes:

• Template mode (default): Clones the firm's .docx template and

populates fields per the stored mapping. Output matches the firm's branded format exactly.

• Generic mode (fallback): Builds a standard RFI from scratch

when no template is available.

Output is always .docx — the user must be able to edit before sending. Read references/rfi-format.md for field-by-field content guidance.

Template mapping is stored at .construction/skills/rfi-drafter/rfi_template_map.json. Created on first RFI, reused for all subsequent RFIs in the project.

Issue records are JSON files in .construction/skills/issues/. Managed via ${CLAUDE_PLUGIN_ROOT}/scripts/issue_manager.py.

Allowed Scripts