Book-Reader-Assist/library/Context Engineering/Front-Matter/Front-Matter-source-text.md

Context Engineering — Front-Matter: Front matter

• Source: /library/Context Engineering/source-file.pdf
• PDF pages: 1–11
• Pages without text: 1

---

Context Engineering: Engineering Information for AI Systems J.C. Ködel

Context Engineering: Engineering Information for AI Systems

1. About the author
2. Map of the trilogy
3. How LLMs use context
4. The model only sees the input
5. Attention: how the model weighs what you sent
6. Nothing survives between calls
7. Where the context hides
8. What changes in your practice
9. Tokens and context windows
10. The model reads tokens, not words
11. What tokenization explains as a bonus
12. The context window is the container
13. A big window is no license to fill it
14. Measure it yourself: what travels with a one-line question
15. The yardstick you take from this chapter
16. Memory and limits
17. The chat’s memory is a replay
18. Where the illusion breaks
19. What about the tools that claim to have memory?
20. Work with the memory that exists, not the one you

imagine 6. The context cycle

1. The shape of the cycle
2. Where the cycle swells
3. The arithmetic of accumulation
4. Reading a session as a cycle
5. Context rot: why large contexts degrade quality
6. The U-shaped curve: “lost in the middle”
7. Needles, haystacks and the test that became a standard
8. Context rot: degradation in tasks that ought to be trivial
9. Diagnosing rot in your session
10. Token economics: the real cost of bad context
11. How the meter runs
12. Agent scale: the multiplier nobody budgets for
13. Do the math yourself
14. Parametric calculation: cost of irrelevant context
15. What to measure tomorrow morning
16. Quality and cost are the same bug
17. Prompt engineering vs context engineering: why the prompt became a second-order variable
18. What prompt engineering really solves
19. The same prompt, opposite results
20. The discipline that takes its place
21. The objections that deserve an answer
22. Where the right context comes from
23. Specifications

1. What a spec carries
2. Examples are the part the model understands best
3. The waterfall objection
4. Living documentation
5. The document that describes the present
6. “All docs rot, so why write them?”
7. What each artifact answers
8. ADRs
9. A record for the why
10. “ADRs are bureaucracy”
11. Three artifacts, three questions
12. Conventions
13. Fewer decisions per task
14. The conventions document
15. Conventions that run in CI
16. Where each kind of information lives
17. Persistent context files
18. The shortcut and what it costs
19. Anatomy of a file that works
20. Anti-patterns, and where each line goes instead
21. “It turns into a dump and nobody maintains it”
22. Project organization
23. The context source you do not write
24. What the technical tree screams
25. What the feature tree screams

1. Modularization
2. Parnas’s criterion
3. Deep modules, small surface
4. A public surface is not the interface keyword
5. Boundary lines across VilaSchedule’s tree
6. “Too much ceremony for a system this size”
7. Context for brownfield projects
8. Step 1: structure and names
9. Step 2: git archaeology
10. Step 3: AI-guided reading
11. Step 4: generating the artifacts incrementally
12. Context layers
13. A layer is a lifetime, not a folder
14. The layers of a session
15. The same session, annotated by layer
16. What the layers let you decide
17. “This is bureaucracy for a twenty-minute session”
18. Context packing
19. Packing is choosing the minimum, and choosing means saying no
20. The inventory of the bloated packet
21. The window is not uniform
22. Four questions that assemble the packet
23. The same request, packed
24. “If I forget the right file, it will make something up”
25. What the packet cannot carry

1. Context recovery
2. Recovery is reassembling what had no address
3. Recovery is not prevention
4. The routine I use to restart a task
5. The state note
6. The two restarts, side by side
7. “In 2026 the agent handles it on its own”
8. Not everything that came back is still true
9. Context validation
10. Checking a belief is not validating input
11. Two ways to state what is not so
12. The statement, the check and the repair
13. The checklist I run
14. When the check fails
15. “If I have to check everything, what is the AI for?”
16. What is left of the check when the history shrinks
17. Context compression
18. Compressing is choosing what is left
19. What a summary optimizes for
20. The anchors you write beforehand
21. “Then turn automatic summarization off”
22. What this chapter assumes is in place
23. One window, one task
24. Context isolation
25. One context per task
26. When splitting is worth the coordination cost

1. Thursday, split again
2. The subtask contract
3. Two subtasks at once, each on its own ground
4. “The subagent loses sight of the whole”
5. “Re-explaining the context to each one is expensive”
6. The packet that fits in no window at all
7. RAG vs direct context
8. Fetching the passage when the question comes up
9. Size, mutability and how each task uses it
10. Embedding is the default until it hurts
11. “RAG retrieves the wrong passage”
12. “Chunking fragments meaning”
13. The column the search does not answer
14. MCP and tools as dynamic context
15. Information you do not read but ask for
16. The name this has in 2026
17. The definition is what the model reads
18. Every tool is context paid for before the question
19. When the data calls for a tool
20. “That is a whole integration to read four times”
21. “And when the tool is down?”
22. What the three decisions still do not say
23. Context security and trust
24. The window has one voice
25. The attack has a name and a test
26. Privilege is granted per tool, not per trust

1. Provenance is origin plus authority
2. The packet is an exposure surface
3. “A good model already resists this”
4. Where to start
5. Development loops with AI
6. Technique is not cadence
7. Pack, run, validate, distill
8. Where recovery comes in
9. One turn on Thursday
10. Calibrate without breaking it
11. “That is ceremony for a ten-minute task”
12. Two weeks later, the same feeling
13. Measuring context: how to evaluate whether your context improves results
14. “Evaluating that is work for a machine learning team”
15. What counts as right the first time
16. Thirty seconds per turn
17. The number on its own decides nothing
18. Two counts that fit in the same file
19. “Fifteen turns prove nothing”
20. Your rate and the team’s
21. Principles applied: chat, IDE, terminal and CI
22. Four questions before any configuration
23. The map of July 2026
24. Chat assistant: the context lives outside the repository
25. IDE agent: the context lives next to the code
26. Terminal agent: the context lives in directory layers

1. Agent in CI: nobody there to correct course
2. One source, four projections
3. “This will age the same way”
4. The context that never leaves your laptop
5. Teams: context as a repository asset
6. What belongs to the repository
7. One owner per artifact
8. An agent’s first day
9. “Nobody is going to maintain this”
10. What happens when all of this meets a project
11. Preparing a project (from scratch and from a legacy system)
12. A caveat about the tool
13. The path from scratch: three files and a tree
14. The legacy path: a packet that shows its evidence
15. How to know the packet is ready
16. A complete AI-guided implementation
17. Session 1: the skeleton, and three sentences that paid for the session
18. Session 2: five statements, three checks and a defect that was not a statement
19. Session 3: the window fills up, and what is left is not what you think
20. Session 4: the same task resumed twice
21. Session 5: the slice that left the main window
22. What the five sessions add up to
23. Post-mortem: where the context failed and how it was recovered

1. The delivery that came back incomplete and said nothing
2. The decision nobody made out loud
3. The reason that died with the session
4. The packet that was wrong
5. The boundary drawn halfway
6. The rule nobody enforced
7. What each failure cost
8. The legacy system as counterpoint
9. Post-mortem script
10. References
11. Papers and articles
12. Vendor documentation and publications
13. Other sources