CO
context-preflight
Estimate working-context demand before substantial work, detect context-shortage risk early, and trigger pre-work compaction or a reduced working set before the task becomes unstable. Use when a request spans multiple files, research artifacts, or large documentation rewrites.
Install
mkdir -p .claude/skills/context-preflight && curl -L -o skill.zip "https://agentskills.codes/api/skills/download/14997" && unzip -o skill.zip -d .claude/skills/context-preflight && rm skill.zipInstalls to .claude/skills/context-preflight
Activation
This is the description your AI agent reads to decide when to run this skill — the better it matches your request, the more reliably it fires.
Estimate working-context demand before substantial work, detect context-shortage risk early, and trigger pre-work compaction or a reduced working set before the task becomes unstable. Use when a request spans multiple files, research artifacts, or large documentation rewrites.277 chars✓ has a “when” triggerlonger than Claude Code's old 250-char listing cap (fine on current versions)
About this skill
Context Preflight
Use this skill before substantial research, planning, implementation, or documentation rewrites.
Goals
- Estimate whether the upcoming task can fit safely in the agent's working context.
- Prefer pre-work compaction over emergency compaction in the middle of implementation.
- Preserve important constraints and decisions before context pressure becomes risky.
Why This Exists
- Mid-task compaction can drop important implementation details, decisions, or constraints.
- A brief preflight is cheaper than re-deriving lost context after work has already started.
- Prompt-engineering guidance favors clear structure, relevant context only, and explicit separation of instructions versus supporting material.
Capacity Model
- Estimate file-heavy technical material at roughly 1 token per 4 bytes.
- Estimate prose-heavy documentation at roughly 1 token per 3 bytes.
- Reserve working room for instructions, the active user request, interim notes, validation output, and the final response.
- Keep a safety margin of at least 25 percent of the practical working budget.
Risk Bands
- Low risk: projected working set uses less than 55 percent of the practical budget.
- Medium risk: projected working set uses 55 to 75 percent of the practical budget.
- High risk: projected working set uses more than 75 percent of the practical budget or requires many large files plus long artifacts.
Procedure
- Identify the likely working set before deep reading.
- Estimate code and prose token demand from file sizes or search results.
- Add buffers for instructions, user intent, validation output, and planned artifact writing.
- Classify the request as low, medium, or high risk.
- If risk is high, compact before substantive work begins.
- If direct
/compactexecution is unavailable, reduce the working set and refresh request artifacts with a concise summary before continuing. - Re-run the estimate if the scope expands materially during the task.
Compaction Rules
- Prefer compaction before implementation, not during debugging or after multiple edits.
- Preserve active constraints, open questions, selected defaults, target files, and validation status in request artifacts before compaction.
- If a request can be split into phases instead of compacting, phased execution is acceptable.
Fallback When /compact Is Unavailable
- Update the active request artifacts with a concise state summary.
- Narrow the working set to only the currently relevant files.
- Defer non-critical side quests or cleanup to a later request.
Output Template
Use assets/context-risk-decision.template.md when recording the preflight result.