atuljha23 Context engineering
Your context window is a working set. Like any working set, it has limits, and the cost of a bad one compounds.
The four moves
- Write — commit durable info out of the transcript and into files the model can re-read later. Plans, decisions, learnings, todos. Don't leave them floating in chat.
- Select — load exactly what you need for the current step. Not the whole repo. The specific files that matter for the specific change.
- Compress — summarize old transcript turns into one line each, and offload details to files. After a long debugging arc: one paragraph of "what we tried and what happened" beats 30 turns of blow-by-blow.
- Isolate — delegate sub-problems to subagents so their discovery doesn't pollute the main context. Research-heavy exploration → Explore agent. Review → code-reviewer agent.
Practical heuristics
- Hot > cold. Keep recent decisions, current plan, and the ~3 files you're actively editing in context. Everything else can be retrieved.
- References, not contents. "See
docs/adr/0007-queue.md" beats pasting 400 lines of ADR. - One plan, one place. If the plan is in context AND in a file, they'll drift. Canonicalize in the file; reference from context.
- Invalidate explicitly. When a decision changes, say so. "We said X, now we're doing Y because Z." Don't let old assumptions linger.
Smells that say "you've lost the plot"
- The model repeats analysis you already gave it
- Transcript is >50 turns and you don't remember what turn 10 decided
- You're answering the same question for a second time
- The model confidently contradicts itself
Act on the smell. Use /holocron:handoff to write a durable doc and start fresh.
Related
/holocron:handoff— persist state before compaction/holocron:learn— persist rules for future sessions/holocron:plan— write the plan to a file rather than carry it in chat