Context engineering
Apply the Write/Select/Compress/Isolate framework to manage the context window. Use when the task is long, the transcript is bloated, the model is drifting, or before /holocron:handoff.From its SKILL.md
npx -y skills add atuljha23/holocron --skill context-engineeringAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
2.1 KB, 473 tokens by cl100k_base, as published. Nobody here has run it
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
What ships with it
Read from the repository
Just SKILL.md. No reference files, no scripts.
Gives 0 of the 12 instructions most context ai engineering skills give in 473 tokens
Counted across 1,328 of the 2,349 authors here whose files we hold, read 2026-09-06
- Dispatch a fresh subagent for each taskin 76 of 1328, across 59 files
- Perform spec compliance review before code quality reviewin 44 of 1328, across 34 files
- Dispatch a final code reviewer after all tasksin 38 of 1328, across 26 files
- Answer subagent questions before allowing implementationin 36 of 1328, across 26 files
- Use the least powerful model capable of the taskin 33 of 1328, across 26 files
- Create a TodoWrite list for all tasksin 32 of 1328, across 22 files
- Perform a task review after each implementationin 31 of 1328, across 24 files
- Extract all tasks and context from the planin 29 of 1328, across 20 files
- Provide full task text to subagentsin 28 of 1328, across 20 files
- Use git worktrees for isolated workspacesin 25 of 1328, across 20 files
- Specify the model explicitly when dispatching a subagentin 23 of 1328, across 18 files
- Execute all tasks from the plan without stoppingin 21 of 1328, across 16 files
Said here and by no other author read
- Commit durable information into files
- Summarize old transcript turns into single lines
- Delegate sub-problems to subagents
- Keep recent decisions and active files in context
- Canonicalize plans in files
- Invalidate old decisions explicitly
Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.