Business Context Loader

Runs on: local-writes - creates or edits files in your OS folder; needs an agent with write access.

Scans state, reports gaps, picks the next question, and suggests an action based on what's already filled.

The companion file is templates/business-context.template.md (the schema). Filled instances live at companies/<slug>-business.md.

When this skill fires

• User opens a companies/<slug>-business.md file
• User says "what's next for <company>", "what's missing on <company>", "what's stale on <company>"
• User starts work on a company workstream and the loader needs to confirm context is fresh enough
• User runs the skill explicitly: business-context-loader companies/<slug>-business.md

What it does (run order)

Step 1 - Scan

Read the target file. Count markers:

• [FILL] (blocking, not yet answered)
• [PENDING] (deferred)
• [SKIPPED] (not applicable)
• [STALE] (older than refresh cadence)

Walk every section header and check the Last refreshed: line. Compare against today's date.

Staleness thresholds:

• Sections 1-10: stale if Last refreshed is more than 365 days old
• Sections 11-12: stale if Last refreshed is more than 90 days old
• Any section with [FILL] in the date itself: counts as never-refreshed (treat as worse than stale)

Step 2 - Report

One block, max 12 lines. Use this format:

business-context-loader: <slug>

State: <X / 12 sections filled. Y must-fill blocking. Z stale.>

Filled (must-fill 1-7): <list section numbers>
Pending / Fill (must-fill 1-7): <list section numbers + 1-word topic each>
Mature stage (8-10): <state per section>
Quarterly refresh (11-12): <state + age in days>

Gap profile:
- Highest-impact next question: <pick one, see Step 3 priority rules>
- Next action you can take based on what IS filled: <pick one, see Step 4 routing>

Do not pad. Do not narrate. The user reads this in 10 seconds or it's failed.

Step 3 - Pick the next question

Apply this priority order to find the highest-impact gap:

1. Section 1 (Identity & Stage) - if anything [FILL] here, this is first. Nothing else makes sense without it.
2. Section 4 (ICP) - if Section 1 is filled but ICP is [FILL], ICP is next. Without ICP, sections 5-7 cannot be evaluated for sense.
3. Section 2 (Problem & Thesis) - once ICP is filled, problem/thesis must align with what the ICP says they need.
4. Section 5 (Offer & Delivery) - ICP + problem must exist before offer is meaningful.
5. Section 3 (Value Model) - runs alongside Section 5.
6. Section 6 (Pricing) - needs offer + value model first.
7. Section 7 (Market Position) - last of the must-fill seven; competitive context is sharpest once the rest is clear.
8. Section 8-10 - mature stage; only flag if business is past month-3.
9. Sections 11-12 - quarterly refresh; flag if stale.

If multiple sections at the same priority level are blocking, pick the one with the most [FILL] markers.

Ask the next question in the user's words, not template language. Example: don't say "fill in section 4 ICP firmographic axis" - say "we don't have a clear ICP yet for <company>. Who do you think actually buys this - what size, what sector, what geography?"

Step 4 - Suggest an action based on what IS filled

Cross-reference filled sections against the rest of the OS to surface a concrete next move. Use these routing rules:

If nothing routes cleanly, suggest the simplest follow-up: "Want to walk Section <N> now? Takes ~5 minutes."

Step 5 - Refresh stamping

Every time the user fills or updates a section through this skill, update the Last refreshed: line for that section to today's date. Append a one-line entry to the Refresh log at the bottom of the file.

Never silently update a section without changing its date.

Branches

Branch A - File doesn't exist yet

If the user mentions a company that has no companies/<slug>-business.md:

"There's no business-context file for <slug> yet. I can copy the template to companies/<slug>-business.md and we walk it together (~30 min for the must-fill seven). Or you can have <person> rant the answers into their own AI assistant using the questionnaire format - see companies/<slug>-questionnaire.md if it exists, otherwise I can generate one. Which?"

Branch B - File exists but mostly empty (less than 3 sections filled)

Treat as a fresh interview. Walk Sections 1-7 in order, one question at a time. Use the brand-interview / voice-interview pattern: rambly answers are fine, extract intent, never ask the user to be more concise.

Branch C - File partially filled (3-9 sections filled)

Run the standard scan-report-pick-suggest flow above. Default to Step 3 (next question) unless the user asks for an action - then default to Step 4.

Branch D - File mostly filled (10+ sections filled, no STALE)

Don't ask for fills. Instead:

• Run staleness check
• Suggest the most useful action from Step 4
• Ask if there's a workstream the user wants to scope using the filled context

Branch E - Stale sections only

Walk the stale sections one by one with this opener: "Section <N> on <topic> was last refreshed <X> days ago. Here's what it currently says: <content>. What's changed?"

Hard rules

• Never invent values. Never fill [FILL] with a guess. If the user doesn't know, write [PENDING] and move on.
• Never silently overwrite an existing answer. Always show the current answer and ask for confirmation before replacing.
• Always update the per-section Last refreshed: date when a section changes.
• One question at a time. Wait for the answer.
• Use the user's language back to them. Don't translate "we sell to small business owners" into "ICP firmographic axis is SMB."
• If the user rambles, extract the answer and reflect it in one line. "Sounds like the ICP is X. Right?"

Output rules

• Plain language. The user is not a strategist.
• The opening report block (Step 2) is the most-read part of this skill. Keep it tight.
• Never narrate "I'll now scan the file" - just scan it and report.

What this skill does NOT do

• Does not write the proposal, deliverable, or content based on the context. Other skills do that - this skill makes sure they have the context to do their job.
• Does not run automated financial analysis. It surfaces the trigger and lets the user choose.
• Does not auto-refresh stale sections. It surfaces them. The user decides what changed.

Re-run behaviour

If the user invokes this skill on a file that has been scanned in the same session, skip the report block and ask: "We already scanned this. Want to keep filling, jump to actions, or check staleness?"