Context Engineering Advisor

1---2name: context-engineering-advisor3description: Diagnose context stuffing vs. context engineering. Use when an AI workflow feels bloated, brittle, or hard to steer reliably.4intent: >-5  Guide product managers through diagnosing whether they're doing **context stuffing** (jamming volume without intent) or **context engineering** (shaping structure for attention). Use this to identify context boundaries, fix "Context Hoarding Disorder," and implement tactical practices like bounded domains, episodic retrieval, and the Research→Plan→Reset→Implement cycle.6type: interactive7theme: ai-agents8best_for:9  - "Diagnosing context stuffing vs. context engineering in your AI workflows"10  - "Building better memory and retrieval architecture for AI agents"11  - "Improving AI output quality through structured context design"12scenarios:13  - "My AI outputs are mediocre even though I'm giving it lots of information — diagnose what's wrong"14  - "I want to architect context properly for a multi-step AI workflow in my product team"15estimated_time: "15-20 min"16---17 18## Purpose19 20Guide product managers through diagnosing whether they're doing **context stuffing** (jamming volume without intent) or **context engineering** (shaping structure for attention). Use this to identify context boundaries, fix "Context Hoarding Disorder," and implement tactical practices like bounded domains, episodic retrieval, and the Research→Plan→Reset→Implement cycle.21 22**Key Distinction:** Context stuffing assumes volume = quality ("paste the entire PRD"). Context engineering treats AI attention as a scarce resource and allocates it deliberately.23 24This is not about prompt writing—it's about **designing the information architecture** that grounds AI in reality without overwhelming it with noise.25 26## Key Concepts27 28### The Paradigm Shift: Parametric → Contextual Intelligence29 30**The Fundamental Problem:**31- LLMs have **parametric knowledge** (encoded during training) = static, outdated, non-attributable32- When asked about proprietary data, real-time info, or user preferences → forced to hallucinate or admit ignorance33- **Context engineering** bridges the gap between static training and dynamic reality34 35**PM's Role Shift:** From feature builder → **architect of informational ecosystems** that ground AI in reality36 37---38 39### Context Stuffing vs. Context Engineering40 41| Dimension | Context Stuffing | Context Engineering |42|-----------|------------------|---------------------|43| **Mindset** | Volume = quality | Structure = quality |44| **Approach** | "Add everything just in case" | "What decision am I making?" |45| **Persistence** | Persist all context | Retrieve with intent |46| **Agent Chains** | Share everything between agents | Bounded context per agent |47| **Failure Response** | Retry until it works | Fix the structure |48| **Economic Model** | Context as storage | Context as attention (scarce resource) |49 50**Critical Metaphor:** Context stuffing is like bringing your entire file cabinet to a meeting. Context engineering is bringing only the 3 documents relevant to today's decision.51 52---53 54### The Anti-Pattern: Context Stuffing55 56**Five Markers of Context Stuffing:**571. **Reflexively expanding context windows** — "Just add more tokens!"582. **Persisting everything "just in case"** — No clear retention criteria593. **Chaining agents without boundaries** — Agent A passes everything to Agent B to Agent C604. **Adding evaluations to mask inconsistency** — "We'll just retry until it's right"615. **Normalized retries** — "It works if you run it 3 times" becomes acceptable62 63**Why It Fails:**64- **Reasoning Noise:** Thousands of irrelevant files compete for attention, degrading multi-hop logic65- **Context Rot:** Dead ends, past errors, irrelevant data accumulate → goal drift66- **Lost in the Middle:** Models prioritize beginning (primacy) and end (recency), ignore middle67- **Economic Waste:** Every query becomes expensive without accuracy gains68- **Quantitative Degradation:** Accuracy drops below 20% when context exceeds ~32k tokens69 70**The Hidden Costs:**71- Escalating token consumption72- Diluted attention across irrelevant material73- Reduced output confidence74- Cascading retries that waste time and money75 76---77 78### Real Context Engineering: Core Principles79 80**Five Foundational Principles:**811. **Context without shape becomes noise**822. **Structure > Volume**833. **Retrieve with intent, not completeness**844. **Small working contexts** (like short-term memory)855. **Context Compaction:** Maximize density of relevant information per token86 87**Quantitative Framework:**88```89Efficiency = (Accuracy × Coherence) / (Tokens × Latency)90```91 92**Key Finding:** Using RAG with 25% of available tokens preserves 95% accuracy while significantly reducing latency and cost.93 94---95 96### The 5 Diagnostic Questions (Detect Context Hoarding Disorder)97 98Ask these to identify context stuffing:99 1001. **What specific decision does this support?** — If you can't answer, you don't need it1012. **Can retrieval replace persistence?** — Just-in-time beats always-available1023. **Who owns the context boundary?** — If no one, it'll grow forever1034. **What fails if we exclude this?** — If nothing breaks, delete it1045. **Are we fixing structure or avoiding it?** — Stuffing context often masks bad information architecture105 106---107 108### Memory Architecture: Two-Layer System109 110**Short-Term (Conversational) Memory:**111- Immediate interaction history for follow-up questions112- Challenge: Space management → older parts summarized or truncated113- Lifespan: Single session114 115**Long-Term (Persistent) Memory:**116- User preferences, key facts across sessions → deep personalization117- Implemented via vector database (semantic retrieval)118- Two types:119  - **Declarative Memory:** Facts ("I'm vegan")120  - **Procedural Memory:** Behavioral patterns ("I debug by checking logs first")121- Lifespan: Persistent across sessions122 123**LLM-Powered ETL:** Models generate their own memories by identifying signals, consolidating with existing data, updating database automatically.124 125---126 127### The Research → Plan → Reset → Implement Cycle128 129**The Context Rot Solution:**130 1311. **Research:** Agent gathers data → large, chaotic context window (noise + dead ends)1322. **Plan:** Agent synthesizes into high-density SPEC.md or PLAN.md (Source of Truth)1333. **Reset:** **Clear entire context window** (prevents context rot)1344. **Implement:** Fresh session using **only** the high-density plan as context135 136**Why This Works:** Context rot is eliminated; agent starts clean with compressed, high-signal context.137 138---139 140### Anti-Patterns (What This Is NOT)141 142- **Not about choosing AI tools** — Claude vs. ChatGPT doesn't matter; architecture matters143- **Not about writing better prompts** — This is systems design, not copywriting144- **Not about adding more tokens** — "Infinite context" narratives are marketing, not engineering reality145- **Not about replacing human judgment** — Context engineering amplifies judgment, doesn't eliminate it146 147---148 149### When to Use This Skill150 151✅ **Use this when:**152- You're pasting entire PRDs/codebases into AI and getting vague responses153- AI outputs are inconsistent ("works sometimes, not others")154- You're burning tokens without seeing accuracy improvements155- You suspect you're "context stuffing" but don't know how to fix it156- You need to design context architecture for an AI product feature157 158❌ **Don't use this when:**159- You're just getting started with AI (start with basic prompts first)160- You're looking for tool recommendations (this is about architecture, not tooling)161- Your AI usage is working well (if it ain't broke, don't fix it)162 163---164 165### Facilitation Source of Truth166 167Use [`workshop-facilitation`](../workshop-facilitation/SKILL.md) as the default interaction protocol for this skill.168 169It defines:170- session heads-up + entry mode (Guided, Context dump, Best guess)171- one-question turns with plain-language prompts172- progress labels (for example, Context Qx/8 and Scoring Qx/5)173- interruption handling and pause/resume behavior174- numbered recommendations at decision points175- quick-select numbered response options for regular questions (include `Other (specify)` when useful)176 177This file defines the domain-specific assessment content. If there is a conflict, follow this file's domain logic.178 179## Application180 181This interactive skill uses **adaptive questioning** to diagnose context stuffing, identify boundaries, and provide tactical implementation guidance.182 183---184 185### Step 0: Gather Context186 187**Agent asks:**188 189Before we diagnose your context practices, let's gather information:190 191**Current AI Usage:**192- What AI tools/systems do you use? (ChatGPT, Claude, custom agents, etc.)193- What PM tasks do you use AI for? (PRD writing, user research synthesis, discovery, etc.)194- How do you provide context? (paste docs, reference files, use projects/memory)195 196**Symptoms:**197- Are AI outputs inconsistent? (works sometimes, not others)198- Are you retrying prompts multiple times to get good results?199- Are responses vague or hedged despite providing "all the context"?200- Are token costs escalating without accuracy improvements?201 202**System Architecture (if applicable):**203- Do you have custom AI agents or workflows?204- How is context shared between agents?205- Do you use RAG, vector databases, or memory systems?206 207**You can describe briefly or paste examples.**208 209---210 211### Step 1: Diagnose Context Stuffing Symptoms212 213**Agent asks:**214 215Let's assess whether you're experiencing **context stuffing**. Which of these symptoms do you recognize?216 217**Select all that apply:**218 2191. **"I paste entire documents into AI"** — Full PRDs, complete user interview transcripts, entire codebases2202. **"AI gives vague, hedged responses despite having 'all the context'"** — Responses like "it depends," "consider these options," non-committal2213. **"I have to retry prompts 3+ times to get usable output"** — Inconsistency is normalized2224. **"Token costs are escalating but accuracy isn't improving"** — Spending more, getting same or worse results2235. **"I keep adding more context hoping it'll help"** — Reflexive expansion without strategy2246. **"My agents pass everything to each other"** — Agent A → Agent B → Agent C with full context chain2257. **"I don't have clear criteria for what to include/exclude"** — No context boundary definitions2268. **"None of these—my AI usage is working well"** — Skip to advanced optimization227 228**User response:** [Select symptoms]229 230**Agent analyzes:**231 232Based on your selections:233- **0-1 symptoms:** Healthy context practices; proceed to optimization234- **2-3 symptoms:** Early context stuffing; address before it scales235- **4+ symptoms:** Active Context Hoarding Disorder; immediate intervention needed236 237**Agent proceeds to diagnostic questions.**238 239---240 241### Step 2: Diagnostic Question 1 — What Specific Decision Does This Support?242 243**Agent asks:**244 245Let's start with the most critical question: **For each piece of context you're including, what specific decision does it support?**246 247**Example Context Analysis:**248 249Imagine you're asking AI to "help with discovery planning." You're providing:250- Entire PRD (20 pages)251- 50 user interview transcripts (full)252- Competitive analysis doc (15 pages)253- Team meeting notes from last 3 months254 255**Question:** What decision are you making right now?256 257**Offer 3 scenarios:**258 2591. **"I'm deciding which user segment to interview first"**260   - **Context needed:** User segments from PRD (2 paragraphs), prior interview themes (1 page synthesis), not full transcripts261   - **Context NOT needed:** Meeting notes, full competitive analysis, full PRD262 2632. **"I'm deciding which discovery questions to ask in interviews"**264   - **Context needed:** Research objectives (from PRD), past interview insights (synthesis), Jobs-to-be-Done framework265   - **Context NOT needed:** Full competitive analysis, full meeting notes266 2673. **"I'm not sure what decision I'm making—I just want AI to 'understand my product'"**268   - **Problem:** No specific decision = context stuffing trap269   - **Fix:** Define the decision first, then select context270 271**Agent recommends:**272 273**Best Practice:** Before adding context, complete this sentence:274> "I need this context because I'm deciding [specific decision], and without [specific information], I can't make that decision."275 276If you can't complete that sentence, you don't need the context.277 278**User response:** [Describe their decision + context]279 280**Agent validates:** Does the context directly support the stated decision? If not, recommend trimming.281 282---283 284### Step 3: Diagnostic Question 2 — Can Retrieval Replace Persistence?285 286**Agent asks:**287 288**Second question: Is this information you always need, or something you can retrieve just-in-time?**289 290**The Distinction:**291 292**Always-Needed (Persist):**293- Core product constraints (technical, regulatory, strategic)294- User preferences that apply to every interaction295- Critical definitions (operational glossary)296- Non-negotiable rules297 298**Episodic (Retrieve on-demand):**299- Project-specific details (this epic, this sprint)300- Historical data (past PRDs, old interview transcripts)301- Contextual facts (competitive analysis, market research)302- Temporary decisions303 304**Key Insight:** Just-in-time retrieval beats always-available. Don't persist what you can retrieve.305 306**Offer 3 options:**307 3081. **"Most of my context is always-needed (core constraints, user prefs)"**309   - **Assessment:** Good instinct; verify with Question 4 (what fails if excluded?)310   - **Recommendation:** Build constraints registry and operational glossary (persist these)311 3122. **"Most of my context is episodic (project details, historical data)"**313   - **Assessment:** Perfect candidate for RAG or retrieval314   - **Recommendation:** Implement semantic search; retrieve only relevant chunks for each query315 3163. **"I'm not sure which is which—I persist everything to be safe"**317   - **Assessment:** Classic Context Hoarding Disorder symptom318   - **Fix:** Apply Question 4 test to each piece of context319 320**Agent recommends:**321 322**Rule of Thumb:**323- **Persist:** Information referenced in 80%+ of interactions324- **Retrieve:** Information referenced in <20% of interactions325- **Gray zone (20-80%):** Depends on retrieval latency vs. context window cost326 327**User response:** [Categorize their context]328 329**Agent provides:** Specific recommendations on what to persist vs. retrieve.330 331---332 333### Step 4: Diagnostic Question 3 — Who Owns the Context Boundary?334 335**Agent asks:**336 337**Third question: Who is responsible for defining what belongs in vs. out of your AI's context?**338 339**The Ownership Problem:**340 341If **no one** owns the context boundary, it will grow indefinitely. Every PM will add "just one more thing," and six months later, you're stuffing 100k tokens per query.342 343**Offer 3 options:**344 3451. **"I own the boundary (solo PM or small team)"**346   - **Assessment:** Good—you can make fast decisions347   - **Recommendation:** Document your boundary criteria (use Questions 1-5 as framework)348 3492. **"My team shares ownership (collaborative boundary definition)"**350   - **Assessment:** Can work if formalized351   - **Recommendation:** Create a "Context Manifest" doc: what's always included, what's retrieved, what's excluded (and why)352 3533. **"No one owns it—it's ad-hoc / implicit"**354   - **Assessment:** Critical risk; boundary will expand uncontrollably355   - **Fix:** Assign explicit ownership; schedule quarterly context audits356 357**Agent recommends:**358 359**Best Practice: Create a Context Manifest**360 361```markdown362# Context Manifest: [Product/Feature Name]363 364## Always Persisted (Core Context)365- Product constraints (technical, regulatory)366- User preferences (role, permissions, preferences)367- Operational glossary (20 key terms)368 369## Retrieved On-Demand (Episodic Context)370- Historical PRDs (retrieve via semantic search)371- User interview transcripts (retrieve relevant quotes)372- Competitive analysis (retrieve when explicitly needed)373 374## Excluded (Out of Scope)375- Meeting notes older than 30 days (no longer relevant)376- Full codebase (use code search instead)377- Marketing materials (not decision-relevant)378 379## Boundary Owner: [Name]380## Last Reviewed: [Date]381## Next Review: [Date + 90 days]382```383 384**User response:** [Describe current ownership model]385 386**Agent provides:** Recommendation on formalizing ownership + template for Context Manifest.387 388---389 390### Step 5: Diagnostic Question 4 — What Fails if We Exclude This?391 392**Agent asks:**393 394**Fourth question: For each piece of context, what specific failure mode occurs if you exclude it?**395 396This is the **falsification test**. If you can't identify a concrete failure, you don't need the context.397 398**Offer 3 scenarios:**399 4001. **"If I exclude product constraints, AI will recommend infeasible solutions"**401   - **Failure Mode:** Clear and concrete402   - **Assessment:** Valid reason to persist constraints403 4042. **"If I exclude historical PRDs, AI won't understand our product evolution"**405   - **Failure Mode:** Vague and hypothetical406   - **Assessment:** Historical context rarely needed for current decisions407   - **Fix:** Retrieve PRDs only when explicitly referencing past decisions408 4093. **"If I exclude this, I'm not sure anything would break—I just include it to be thorough"**410   - **Failure Mode:** None identified411   - **Assessment:** Context stuffing; delete immediately412 413**Agent recommends:**414 415**The Falsification Protocol:**416 417For each context element, complete this statement:418> "If I exclude [context element], then [specific failure] will occur in [specific scenario]."419 420**Examples:**421- ✅ Good: "If I exclude GDPR constraints, AI will recommend features that violate EU privacy law."422- ❌ Bad: "If I exclude this PRD, AI might not fully understand the product." (Vague)423 424**User response:** [Apply falsification test to their context]425 426**Agent provides:** List of context elements to delete (no concrete failure identified).427 428---429 430### Step 6: Diagnostic Question 5 — Are We Fixing Structure or Avoiding It?431 432**Agent asks:**433 434**Fifth question: Is adding more context solving a problem, or masking a deeper structural issue?**435 436**The Root Cause Question:**437 438Context stuffing often hides bad information architecture. Instead of fixing messy, ambiguous documents, teams add more documents hoping AI will "figure it out."439 440**Offer 3 options:**441 4421. **"I'm adding context because our docs are poorly structured/ambiguous"**443   - **Assessment:** You're masking a structural problem444   - **Fix:** Clean up the docs first (remove ambiguity, add constraints, define terms)445   - **Example:** Instead of pasting 5 conflicting PRDs, reconcile them into 1 Source of Truth446 4472. **"I'm adding context because we don't have a shared operational glossary"**448   - **Assessment:** You're compensating for missing foundations449   - **Fix:** Build the glossary (20-30 key terms); AI can reference it reliably450   - **Example:** Define "active user," "churn," "engagement" unambiguously451 4523. **"I'm adding context because our constraints aren't documented"**453   - **Assessment:** You're avoiding constraint engineering454   - **Fix:** Create constraints registry (technical, regulatory, strategic)455   - **Example:** Document "We won't build mobile apps" vs. explaining it in every prompt456 457**Agent recommends:**458 459**The Structural Health Test:**460 461If you're adding context to compensate for:462- **Ambiguous documentation** → Fix the docs, don't add more463- **Undefined terms** → Build operational glossary464- **Undocumented constraints** → Create constraints registry465- **Conflicting information** → Reconcile into Source of Truth466 467**User response:** [Identify structural issues]468 469**Agent provides:** Prioritized list of structural fixes before adding more context.470 471---472 473### Step 7: Define Memory Architecture474 475**Agent asks:**476 477Based on your context needs, let's design a **two-layer memory architecture**.478 479**Your Context Profile (from previous steps):**480- Always-needed context: [Summary from Q2]481- Episodic context: [Summary from Q2]482- Boundary owner: [From Q3]483- Validated essentials: [From Q4]484- Structural fixes needed: [From Q5]485 486**Recommended Architecture:**487 488**Short-Term (Conversational) Memory:**489- **What it stores:** Immediate interaction history for follow-up questions490- **Lifespan:** Single session491- **Management:** Summarize or truncate older parts to avoid crowding492- **Your specific needs:** [Agent customizes based on user's workflow]493 494**Long-Term (Persistent) Memory:**495- **What it stores:** User preferences, core constraints, operational glossary496- **Lifespan:** Persistent across sessions497- **Implementation:** Vector database (semantic retrieval)498- **Two types:**499  - **Declarative Memory:** Facts (e.g., "We follow HIPAA regulations")500  - **Procedural Memory:** Behavioral patterns (e.g., "Always validate feasibility before usability")501- **Your specific needs:** [Agent customizes]502 503**Retrieval Strategy (Episodic Context):**504- **What it retrieves:** Historical PRDs, user interviews, competitive analysis505- **Method:** Semantic search triggered by query intent506- **Optimization:** Contextual Retrieval (Anthropic) — prepend explanatory context to each chunk before embedding507- **Your specific needs:** [Agent customizes]508 509**Agent offers:**510 511Would you like me to:5121. **Generate a Context Architecture Blueprint** for your specific use case?5132. **Provide implementation guidance** (tools, techniques, best practices)?5143. **Design a retrieval strategy** for your episodic context?515 516**User response:** [Selection]517 518---519 520### Step 8: Implement Research → Plan → Reset → Implement Cycle521 522**Agent asks:**523 524Now let's implement the **Research → Plan → Reset → Implement** cycle to prevent context rot.525 526**The Problem:** As agents research, context windows grow chaotic—filled with dead ends, errors, and noise. This dilutes attention and causes goal drift.527 528**The Solution:** Compress research into a high-density plan, then **clear the context window** before implementing.529 530**The Four-Phase Cycle:**531 532**Phase 1: Research (Chaotic Context Allowed)**533- Agent gathers data from multiple sources534- Context window grows large and messy (this is expected)535- Dead ends, failed hypotheses, and noise accumulate536- **Goal:** Comprehensive information gathering537 538**Phase 2: Plan (Synthesis)**539- Agent synthesizes research into a high-density SPEC.md or PLAN.md540- This becomes the **Source of Truth** for implementation541- **Key elements:**542  - Decision made543  - Evidence supporting decision544  - Constraints applied545  - Next steps (sequenced)546- **Format:** Structured, concise, unambiguous547 548**Phase 3: Reset (Clear Context Window)**549- **Critical step:** Clear the entire context window550- Delete all research artifacts, dead ends, errors551- This prevents context rot from poisoning implementation552 553**Phase 4: Implement (Fresh Session with Plan Only)**554- Start a new session with **only the high-density plan** as context555- Agent has clean, focused attention on execution556- No noise from research phase557 558**Agent offers 3 options:**559 5601. **"I want a template for the PLAN.md format"**561   - Agent provides structured template for high-density plans562 5632. **"I want to see an example of this cycle in action"**564   - Agent walks through concrete PM use case (e.g., discovery planning)565 5663. **"I'm ready to implement this in my workflow"**567   - Agent provides step-by-step implementation guide568 569**User response:** [Selection]570 571**Agent provides:** Tailored guidance based on selection.572 573---574 575### Step 9: Action Plan & Next Steps576 577**Agent synthesizes:**578 579Based on your context engineering assessment, here's your action plan:580 581**Immediate Fixes (This Week):**5821. [Delete context with no falsifiable failure mode from Q4]5832. [Apply Research→Plan→Reset→Implement to your next AI task]5843. [Document context boundary in Context Manifest]585 586**Foundation Building (Next 2 Weeks):**5871. [Build constraints registry with 20+ entries]5882. [Create operational glossary with 20-30 key terms]5893. [Implement two-layer memory architecture]590 591**Long-Term Optimization (Next Month):**5921. [Set up semantic retrieval for episodic context]5932. [Assign context boundary owner + quarterly audit schedule]5943. [Implement Contextual Retrieval (Anthropic) for RAG]595 596**Success Metrics:**597- Token usage down 50%+ (less context stuffing)598- Output consistency up (less retry/regeneration)599- Response quality up (sharper, less hedged answers)600- Context window stable (no unbounded growth)601 602**Agent offers:**603 604Would you like me to:6051. **Generate specific implementation docs** (Context Manifest, PLAN.md template, etc.)?6062. **Provide advanced techniques** (Contextual Retrieval, LLM-powered ETL)?6073. **Review your current context setup** (provide feedback on specific prompts/workflows)?608 609---610 611## Examples612 613### Example 1: Solo PM Context Stuffing → Engineering614 615**Context:**616- Solo PM at early-stage startup617- Using Claude Projects for PRD writing618- Pasting entire PRDs (20 pages) + all user interviews (50 transcripts) every time619- Getting vague, inconsistent responses620 621**Assessment:**622- Symptoms: Hedged responses, normalized retries (4+ symptoms)623- Q1 (Decision): "I just want AI to understand my product" (no specific decision)624- Q2 (Persist/Retrieve): Persisting everything (no retrieval strategy)625- Q3 (Ownership): No formal owner (solo PM, ad-hoc)626- Q4 (Failure): Can't identify concrete failures for most context627- Q5 (Structure): Avoiding constraint documentation628 629**Diagnosis:** Active Context Hoarding Disorder630 631**Intervention:**6321. **Immediate:** Delete all context that fails Q4 test → keeps 20% of original6332. **Week 1:** Build constraints registry (10 technical constraints, 5 strategic)6343. **Week 2:** Create operational glossary (25 terms)6354. **Week 3:** Implement Research→Plan→Reset→Implement for next PRD636 637**Outcome:** Token usage down 70%, output quality up significantly, responses crisp and actionable.638 639---640 641### Example 2: Growth-Stage Team with Agent Chains642 643**Context:**644- Product team with 5 PMs645- Custom AI agents for discovery synthesis646- Agent A (research) → Agent B (synthesis) → Agent C (recommendations)647- Each agent passes full context to next → context window explodes to 100k+ tokens648 649**Assessment:**650- Symptoms: Escalating token costs, inconsistent outputs (3 symptoms)651- Q1 (Decision): Each agent has clear decision, but passes unnecessary context652- Q2 (Persist/Retrieve): Mixing persistent and episodic without strategy653- Q3 (Ownership): No explicit owner; each PM adds context654- Q4 (Failure): Agents pass "just in case" context with no falsifiable failure655- Q5 (Structure): Missing Context Manifest656 657**Diagnosis:** Agent orchestration without boundaries658 659**Intervention:**6601. **Immediate:** Define bounded context per agent (Agent A outputs only 2-page synthesis to Agent B, not full research)6612. **Week 1:** Assign context boundary owner (Lead PM)6623. **Week 2:** Create Context Manifest (what persists, what's retrieved, what's excluded)6634. **Week 3:** Implement Research→Plan→Reset→Implement between Agent B and Agent C664 665**Outcome:** Token usage down 60%, agent chain reliability up, costs reduced by 50%.666 667---668 669### Example 3: Enterprise with RAG but No Context Engineering670 671**Context:**672- Large enterprise with vector database RAG system673- "Stuff the whole knowledge base" approach (10,000+ documents)674- Retrieval returns 50+ chunks per query → floods context window675- Accuracy declining as knowledge base grows676 677**Assessment:**678- Symptoms: Vague responses despite "complete knowledge," normalized retries (2 symptoms)679- Q1 (Decision): Decisions clear, but retrieval has no intent (returns everything)680- Q2 (Persist/Retrieve): Good instinct to retrieve, but no filtering681- Q3 (Ownership): Engineering owns RAG, Product doesn't own context boundaries682- Q4 (Failure): Can't identify why 50 chunks needed vs. 5683- Q5 (Structure): Knowledge base has no structure (flat documents, no metadata)684 685**Diagnosis:** Retrieval without intent (RAG as context stuffing)686 687**Intervention:**6881. **Immediate:** Limit retrieval to top 5 chunks per query (down from 50)6892. **Week 1:** Implement Contextual Retrieval (Anthropic) — prepend explanatory context to each chunk during indexing6903. **Week 2:** Add metadata to documents (category, recency, authority)6914. **Week 3:** Product team defines retrieval intent per query type (discovery = customer insights, feasibility = technical constraints)692 693**Outcome:** Accuracy up 35% (from Anthropic benchmark), latency down 60%, token usage down 80%.694 695---696 697## Common Pitfalls698 699### 1. **"Infinite Context" Marketing vs. Engineering Reality**700**Failure Mode:** Believing "1 million token context windows" means you should use all of them.701 702**Consequence:** Reasoning Noise degrades performance; accuracy drops below 20% past ~32k tokens.703 704**Fix:** Context windows are not free. Treat tokens as scarce; optimize for density, not volume.705 706---707 708### 2. **Retrying Instead of Restructuring**709**Failure Mode:** "It works if I run it 3 times" → normalizing retries instead of fixing structure.710 711**Consequence:** Wastes time and money; masks deeper context rot issues.712 713**Fix:** If retries are common, your context structure is broken. Apply Q5 (fix structure, don't add volume).714 715---716 717### 3. **No Context Boundary Owner**718**Failure Mode:** Ad-hoc, implicit context decisions → unbounded growth.719 720**Consequence:** Six months later, every query stuffs 100k tokens per interaction.721 722**Fix:** Assign explicit ownership; create Context Manifest; schedule quarterly audits.723 724---725 726### 4. **Mixing Always-Needed with Episodic**727**Failure Mode:** Persisting historical data that should be retrieved on-demand.728 729**Consequence:** Context window crowded with irrelevant information; attention diluted.730 731**Fix:** Apply Q2 test: persist only what's needed in 80%+ of interactions; retrieve the rest.732 733---734 735### 5. **Skipping the Reset Phase**736**Failure Mode:** Never clearing context window during Research→Plan→Implement cycle.737 738**Consequence:** Context rot accumulates; goal drift; dead ends poison implementation.739 740**Fix:** Mandatory Reset phase after Plan; start implementation with only high-density plan as context.741 742---743 744## References745 746### Related Skills747- **[ai-shaped-readiness-advisor](../ai-shaped-readiness-advisor/SKILL.md)** (Interactive) — Context Design is Competency #1 of AI-shaped work748- **[problem-statement](../problem-statement/SKILL.md)** (Component) — Evidence-based framing requires context engineering749- **[epic-hypothesis](../epic-hypothesis/SKILL.md)** (Component) — Testable hypotheses depend on clear constraints (part of context)750- **[pol-probe-advisor](../pol-probe-advisor/SKILL.md)** (Interactive) — Validation experiments benefit from context engineering (define what AI needs to know)751 752### External Frameworks753- **Dean Peters** — [*Context Stuffing Is Not Context Engineering*](https://deanpeters.substack.com/p/context-stuffing-is-not-context-engineering) (Dean Peters' Substack, 2026)754- **Teresa Torres** — *Continuous Discovery Habits* (Context Engineering as one of 5 new AI PM disciplines)755- **Marty Cagan** — *Empowered* (Feasibility risk in AI era includes understanding "physics of AI")756- **Anthropic** — [Contextual Retrieval whitepaper](https://www.anthropic.com/news/contextual-retrieval) (35% failure rate reduction)757- **Google** — Context engineering whitepaper on LLM-powered memory systems758 759### Technical References760- **RAG (Retrieval-Augmented Generation)** — Standard technique for episodic context retrieval761- **Vector Databases** — Semantic search for long-term memory (Pinecone, Weaviate, Chroma)762- **Contextual Retrieval (Anthropic)** — Prepend explanatory context to chunks before embedding763- **LLM-as-Judge** — Automated evaluation of context quality