Intake Chatbot — Conversational Field Data Classification

Signals: temperature, weather words (rain, snow, wind, sun, hot, cold, muddy, frozen, wet, dry), site condition descriptions, work stoppages due to weather Examples: "rained all morning", "35 degrees at 7am", "site is a mud pit", "wind picked up around 2"

Crew / Work Performed

Signals: company names, trade names, headcounts, work descriptions, action verbs (poured, framed, hung, pulled, set, installed, stripped, backfilled, graded, excavated, welded) Examples: "Walker had 6 guys doing backfill", "electricians pulling wire on 2", "poured footings at grid C"

Materials / Deliveries

Signals: delivery verbs (delivered, showed up, arrived, received, got), material names (concrete, rebar, steel, lumber, drywall, pipe, duct, roofing), supplier names, quantities, conditions (good, damaged, wrong, short) Examples: "got 12 tons of rebar from Harris", "concrete truck showed up at 9", "wrong pipe delivered"

Equipment

Signals: equipment names (excavator, crane, loader, bobcat, roller, lift, forklift, generator, pump, compressor), status words (down, broken, idle, mobilized, demobilized), hours Examples: "excavator broke down after lunch", "new crane coming tomorrow", "roller ran all day"

Schedule Updates

Signals: phase mentions, percent complete, milestone references, delay language (behind, ahead, on track, delayed, pushed back, accelerated, slipped), coordination mentions Examples: "we're at about 20% now", "foundation should be done by Friday", "steel delivery pushed to next week"

Delays / Impacts

Signals: delay language with impact/cause — stopped work, couldn't work, shut down, held up, waiting on, standby, suspended, out of sequence, no-showed, can't proceed, blocked by, on hold Captures: delay_type, description, activities_impacted, estimated_duration, critical_path_impact, linked_delay_id, responsible_party, supporting_references

Classification rule: When input describes work being prevented, stopped, or impacted by an external cause, classify as BOTH the relevant work section (crew, equipment, etc.) AND as a delay_event. The delay event captures the cause, duration, and impact; the work section captures the effect on the crew/schedule.

Delay type classification: Match to one of 8 standard categories:

1. Weather — temperature, precipitation, wind, flooding, seasonal conditions
2. Owner-Directed — stop work orders, scope holds, access restrictions from owner
3. Design/Spec — missing details, RFI holds, design errors, incomplete drawings
4. Material/Supply Chain — late deliveries, wrong materials, backorders, fabrication delays
5. Sub Performance — no-shows, insufficient crew, quality failures requiring rework
6. Force Majeure — acts of God, pandemics, civil unrest, utility failures beyond control
7. Permit/Regulatory — permit delays, failed inspections blocking work, code holds
8. Differing Site Conditions — unexpected subsurface, hazmat discovery, utility conflicts

Visitors / Inspections

Signals: visitor language (came by, stopped by, was here, visited), inspector mentions, inspection types (footing, framing, rough-in, final, special, compaction), results (passed, failed, conditional) Examples: "inspector came at 10, passed the footings", "owner walked the site", "fire marshal stopped by"

Photos

Signals: image uploads attached to the message Action: Analyze the photo content and classify it into the appropriate section

General Notes

Signals: anything that doesn't clearly fit another section — conversations, coordination items, upcoming work, RFI mentions, directives, observations Examples: "talked to the architect about the stair detail", "need to get the crane permit sorted", "tomorrow we're doing the pre-pour meeting"

Multi-Section Input

A single user message often contains information for multiple sections. Parse each piece separately:

Input: "Walker had 6 guys backfilling the east side, got a load of structural fill from Martin Marietta, and the compaction tester came out and we passed"

Classification:

1. Crew → Walker Construction, 6 workers, backfill at east wing
2. Materials → Structural fill delivery from Martin Marietta
3. Inspections → Compaction test, passed

Entity Resolution

When project intelligence is loaded, resolve casual references to official data:

Subcontractor Resolution

• "Walker" → "Walker Construction" (match on partial name)
• "the concrete guys" → match sub with trade = "Concrete"
• "electricians" → match sub with trade containing "Electric"
• If no match found → use the name as-is, mark sub_matched: false in the entry, and flag for resolution during report-qa. The report-qa skill will attempt fuzzy matching and auto-correction; the intake step only flags, never auto-corrects unmatched names.

Location Resolution

• "east side" → "East Wing, Grid Lines E-G" (match building area)
• "second floor" → "Level 2" (match floor level)
• "by the elevator" → "Central Core, Grid Lines C-E / 4-7" (match building area by feature)
• "grid C" → direct reference, resolve to building area

Material Resolution

• "concrete" → reference spec section if work type known (e.g., "per Section 03 30 00, 4000 PSI")
• "rebar" → "ASTM A615 Grade 60" (from key materials)
• Generic terms → enrich with spec data if available

Spec-Based Enrichment

When a work type is identified from the intake entry (e.g., "poured footings", "hung drywall", "pulled wire"):

• Read specs-quality.json → spec_sections[] → find the matching CSI section for the work type
• Pull key requirements: concrete mix design, testing frequency, acceptance criteria
• Add as enrichment context in the classified entry (not displayed to user unless relevant)
• Example: "poured footings" → pull Section 03 30 00, mix design 4000 PSI, testing every 50 CY

Schedule-Aware Context

When an activity is mentioned in an intake entry:

• Read schedule.json → critical_path[] → check if the logged activity matches a critical path item
• If on critical path: add critical_path: true flag to the intake entry for emphasis in the daily report
• Example: "Steel erection started" + critical_path includes "PEMB Erection" → flag entry as critical path activity

Weather-Work Cross-Check

When outdoor work is logged AND weather has been reported for today:

• Read specs-quality.json → weather_thresholds[] → match the work type
• Cross-check today's reported weather against the threshold
• If threshold exceeded: auto-flag in the intake entry: "Note: Today's temp (35°F) is below the 40°F minimum for concrete placement per Spec 03 30 00"
• Do not block the entry — just flag for awareness

Quantity Context

When work with measurable output is logged (concrete poured, flooring installed, outlets wired):

• Read plans-spatial.json → quantities → find the total plan quantity for that element type
• Add total plan quantity as context: "12 CY poured today of 42.3 CY total footings (28%)"
• This enables automatic progress percentage calculations in the daily report

Inspection Reminder

When the logged work type has a hold point in the specs:

• Read specs-quality.json → hold_points[] → match the work type
• If a hold point exists and the work is at the trigger stage: prompt the superintendent
• "Concrete pre-placement — did this need an inspection before proceeding? (Hold point HP-02)"
• Accept dismissal — this is a reminder, not a blocker

Sub Absence Detection

When the user mentions a sub is absent ("nobody from the plumber", "electricians didn't show", "Walker was a no-show", "[sub] not on site"):

1. Log a crew entry with headcount: 0 and work description: "Not on site"
2. If the sub is in the directory AND was scheduled this week (from schedule.json), add to expected_subs_missing in the daily summary
3. Prompt for reason if not given: "Any reason they weren't here?"
4. If the user provides a reason, prompt for impact: "Is this causing any delay?"
5. If it caused a delay, also classify as a delay_event with type: Sub Performance

In the report: Absent subs appear in the Crew on Site table with headcount 0 and a note explaining the absence.

Duplicate Sub Entry Merging

When a new intake entry references a sub that already has an entry for today:

1. Same sub, additional workers or work: Merge into the existing entry. Update headcount (take the higher number unless the user explicitly says "plus", "also", or "more"). Append the new work description.

   • Example: "Walker had 6 guys backfilling" + "Walker also had 2 guys on grading" → Walker: 8 workers, backfill + grading

2. Same sub, corrected headcount: Replace the old value.

   • Example: "Actually Walker had 8, not 6" → update to 8

3. Same sub, different location: Keep as separate line items under the same sub name.

   • Example: "Walker had 4 at the east wing" + "Walker had 3 at the west wing" → two entries

4. Confirm the merge: "Updated Walker to 8 workers — backfill + grading at east wing. That right?"

Key rule: When in doubt, ask: "Is that 8 total for Walker, or 8 more on top of the 6?"

Claims Mode

When project-config.json has claims_mode: true, the intake system captures claims-grade detail on top of standard classification. This enhanced capture supports the claims-documentation skill's evidence requirements.

Crew entries: After logging a sub, prompt: "Claims mode — can you give me names and hours for that crew?"

• Capture: worker_names[], start_time, end_time, total_hours, overtime_hours
• If user declines: proceed without — claims detail is recommended, not required

Equipment entries: After logging equipment, prompt: "Got the [equipment]. Unit number? Any idle or standby time today?"

• Capture: equipment_id, idle_hours, idle_reason, standby_hours, standby_reason

Material deliveries: After logging a delivery, prompt: "Delivery ticket number? Was it on time?"

• Capture: delivery_ticket_number, scheduled_delivery_time, actual_delivery_time

Delay events: Always capture the full delay classification when claims mode is active.

Prompting etiquette: Claims-mode prompts follow the same rules as standard prompts — ask one at a time, accept dismissals ("skip", "no", silence), never block the user from continuing to log.

Auto-Context

Based on what's been logged so far today, the system can proactively prompt for related information. See references/auto-context.md for the full rule set.

Key principle: suggest, don't require. If the user ignores a prompt, move on.

Conversational Flow

See references/conversational-flow.md for patterns on handling multi-turn interactions, corrections, and follow-up questions.

Edge Cases

See references/edge-cases.md for handling ambiguous input, contradictions, and unusual scenarios.

Review Log

When the user asks to review today's logged entries ("show my log", "what's logged", "review today's entries"), perform the following:

1. Search for daily-report-intake.json in the user's working directory
2. If not found: "Nothing logged today yet. Use /log to start capturing field observations."
3. If found, load entries and group by report section (weather, crew, materials, equipment, schedule, visitors, photos, general notes)
4. Present a summary organized by section:
   • Show entry count per section
   • Show key details (sub names, material names, etc.)
   • Highlight any sections with no entries yet
5. Offer next steps:
   • "Anything to add or correct before generating the report?"
   • "Run /daily-report when you're ready to generate"
   • "Say 'clear the log' to start over"

Clear Log

When the user asks to clear the intake log ("clear the log", "start over", "wipe today's log", "reset the log"):

1. Search for daily-report-intake.json in the user's working directory
2. If not found: "No intake log to clear — nothing has been logged today."
3. If found, show the user how many entries will be cleared: "You have [X] entries logged today across [Y] report sections. Clear everything and start fresh?"
4. Wait for explicit confirmation. Do NOT clear without user approval.
5. If confirmed:
   • Archive the current log as daily-report-intake-{YYYY-MM-DD}-cleared.json (entries aren't permanently lost)
   • Create a fresh empty intake file:

     {
       "date": "YYYY-MM-DD",
       "entries": []
     }
     

   • Confirm: "Log cleared. Use /log to start fresh."

Additional Resources

• references/intake-schema.md — JSON schema for intake log entries
• references/classification-examples.md — 50+ classification examples for training
• references/auto-context.md — proactive prompting rules
• references/conversational-flow.md — multi-turn interaction patterns
• references/proactive-prompting.md — smart follow-up question rules
• references/edge-cases.md — handling ambiguous and unusual input