| name | context-bootstrapper |
| description | Receiving-agent protocol. Boots a new agent into the minimum correct context (select + isolate) and produces a clear next-step plan without guessing. |
Context Bootstrapper — Receiving Agent Protocol
Use this skill when starting work on this repo, resuming after a break, switching to a new “feature-agent”, or when context feels bloated/confusing.
Goal: reach a confident, bounded next step using minimal context.
Bootstrap order (read in this order)
-
Project baseline (always-on context)
.cursor/rules/always.mdc
AGENT_HANDOFF.md
-
Current phase / feature truth (choose the one that matches the user’s goal)
PM_PLAN.md (phase/scope)
TEST_PLAN.md (Tier 1 / Tier 2 validation gates)
-
Most recent session handoff note (if present)
.cursor/handoff/NNNN-handoff-YYYY-MM-DD_HHmm.md (highest NNNN, tie-break by timestamp)
- and/or
doc/handoff/NNNN-HANDOFF-YYYY-MM-DD_HHmm.md (same rule)
-
If the task is code-touching: read the smallest set of files necessary to act safely.
Produce the “Receiver Brief” (what you must write next)
After reading, produce a short brief with:
- Objective: one sentence; restate the user goal precisely.
- Scope:
- Constraints: branch policy, “no guessing”, validation tier.
- Acceptance criteria: 2–5 verifiable checks.
- Next steps: 3–7 bite-sized steps, each with a validation hook.
- Open questions: only if something blocks safe progress.
If any required inputs are missing, stop and ask before acting.
Token budget guidance (cheap but effective)
Preferred context payload for a new session:
- Tracked truth:
AGENT_HANDOFF.md + the relevant plan doc(s) (project state)
- One handoff note: latest only (session delta)
- Only the files you’re editing (and their direct dependencies)
Avoid:
- Full transcript dumps
- Large logs unless they directly change decisions
- Repeating old execution details already captured in tracked docs