Skip to content

context-routing.md

Latest commit

 

History

History
217 lines (160 loc) · 7.93 KB

context-routing.md

File metadata and controls

217 lines (160 loc) · 7.93 KB
Raw

Context Routing Guide

purpose: Detailed routing rules for AI agents navigating this knowledge base audience: AI agents (Cursor, Claude Code) and humans designing agent workflows summary: Explains HOW to load context efficiently — tiered loading, task stacks, sensitivity rules, and navigation patterns. token_estimate: large depends_on: [] related: ../CLAUDE.md, ../context/README.md, ../pipeline/README.md domain: company confidence: canonical context_tier: 1 last_updated: 2026-02-18

CLAUDE.md tells agents what exists and when to load it. This doc explains HOW to load context efficiently and what to watch out for.

Tiered Context Loading

Not all files are equal. Load the minimum context needed for the task.

Tier When Loaded What's There Token Budget
0 — Always Every conversation CLAUDE.md (~50 lines) ~500 tokens
1 — Task-triggered When task type is identified Style guide, relevant role, key docs 2,000-5,000 tokens
2 — On-demand When agent hits a specific question Individual docs, study guides, research 1,000-3,000 per file
3 — Search-only Only via grep/search, never bulk-loaded Transcripts, downloads, customer records Variable

The rule: Start at Tier 0. Escalate only when you need more information. Never skip to Tier 3 without checking Tier 1-2 first.

Task Routing Stacks

Each task type has a recommended loading sequence. Load files in order — stop when you have enough context.

Writing Content

1. context/voice/writing-style-context-v2.md     (always — the definitive guide)
2. [FILL IN: Additional style guides for specific channels]
3. docs/business/business-model.md                (if about [YOUR COMPANY])
4. docs/products/product-overview-template.md     (if about a specific product)
5. knowledge/domain/                              (if about your domain expertise)

Answering Company Questions

1. docs/start-here.md                             (orientation)
2. docs/company/strategy-overview.md              (if strategy question)
3. docs/business/business-model.md                (if business model question)
4. docs/business/ideal-customer-profile.md        (if ICP/sales question)
5. docs/delivery/                                 (if about how we deliver)
6. docs/finance/                                  (if financial — flag sensitivity)

Making Strategic Decisions

1. context/roles/[relevant-role]-v1.md            (load the right persona)
2. docs/company/strategy-overview.md              (strategic context)
3. docs/business/business-model.md                (business context)
4. docs/finance/fiscal-plan-template.md           (if financial implications)
5. context/personal/[user-manual].md              (if advising the founder directly)

Research / Deep Dives

1. knowledge/[relevant-topic]/README.md           (start at the hub)
2. pipeline/research/                             (check existing research)
3. sources/sources-index.md                       (find reference material)
4. records/downloads/                             (search — don't bulk-load)

Client Work

1. records/customers/[client]/[client]-client-context-v1.md  (client context)
2. records/transcripts/                           (search for relevant meetings)
3. docs/delivery/                                 (how we deliver)
4. context/voice/writing-style-context-v2.md      (if writing for client)

Creating Deliverables

1. Load the relevant task stack above FIRST
2. Save drafts to: pipeline/scratchpad/
3. Save finished work to: pipeline/outputs/
4. Never save directly to docs/ or knowledge/ without explicit instruction

Sensitivity Classification

Some content requires caution. When in doubt, ask before generating.

Level What It Means Directories Action
Public Can be shared externally knowledge/, docs/products/[public content] Generate freely
Internal Company-internal, not secret docs/company/, docs/business/, docs/delivery/, context/ Generate freely, don't share externally
Leadership-only Financial, strategic, sensitive docs/finance/, context/personal/ Flag before generating. Don't include in outputs without asking.

Hard rules:

  • Never generate legal, compliance, or financial content without flagging
  • Never include fiscal plans, board meeting content, or valuation data in client-facing outputs
  • Never share personal context (psych profile, user manual) externally
  • When information conflicts between docs, prefer the newer document

Pipeline Rules

Files flow forward through the pipeline. Never backward.

pipeline/research/  →  pipeline/scratchpad/  →  pipeline/outputs/
    (raw material)       (working drafts)        (finished work)
Stage What Goes Here What "Done" Means Next Step
research/ Raw research, competitive briefs, data gathering Material is organized and ready to inform a deliverable Move key findings to scratchpad/ as a draft
scratchpad/ Drafts, working docs, experiments Document is polished enough to share or publish Graduate to outputs/
outputs/ Finished deliverables, published content Delivered to client or published externally Archive or reference

Process traces: When a skill produces output, include a trace in the file's metadata:

<metadata>
source_skill: research-to-study-guide
input_files: [pipeline/research/topic-raw-notes.md]
output_stage: scratchpad
last_updated: 2026-02-18
</metadata>

Records: Search Protocol

The records/ directory can contain hundreds of files. Never bulk-load them.

Directory How to Access
records/transcripts/ Search by date (YYYY-MM-DD prefix) or grep for keywords
records/customers/ Navigate to records/customers/[client]/ for client context
records/downloads/ Grep for topics. Never read all at once.

Search strategy:

  1. Use grep/search with specific keywords — never read *
  2. Start with README.md or INDEX.md in each directory to understand what's there
  3. Read only the specific files that match your query
  4. If you need meeting context, search by date or participant name

Navigation Patterns

"I don't know where to look"

  1. Read CLAUDE.md (Tier 0 — you already have this)
  2. Check the root INDEX.md for a global sitemap
  3. Check the README.md in the most relevant directory
  4. Search/grep for keywords across the repo

"I need to understand [YOUR COMPANY]"

Start at docs/start-here.md. It links to everything else.

"I need to write something"

Always load context/voice/writing-style-context-v2.md before writing anything. No exceptions.

"I need a persona for analysis"

Pick from context/roles/. The README.md lists all roles with when to use each one.

"I found conflicting information"

  1. Check last_updated in metadata — newer wins
  2. Check confidence in metadata — canonical > current > research > aspirational
  3. If still unclear, flag the conflict and ask

Metadata Schema

Files use <metadata> XML tags for machine-readable context:

<metadata>
purpose: [What this file does]
audience: [Who reads this]
summary: [ONE sentence summary]
token_estimate: [small/medium/large]
depends_on: [Files that must co-load, if any]
related: [Cross-references]
domain: [company/product/writing/etc.]
confidence: [canonical/current/draft]
context_tier: [0/1/2/3]
last_updated: [YYYY-MM-DD]
</metadata>

Confidence levels:

  • canonical — Source of truth, actively maintained
  • current — Accurate as of last_updated, may drift
  • research — Findings and analysis, not official position
  • aspirational — Future plans, not yet implemented