| name | mandelbot-spike-harden |
| description | Use this skill when you want to build something end-to-end fast and then harden it. Delegates a spike child (cut corners, prove the approach), then the parent reads the summary and proceeds normally — plan, delegate, or implement directly. Good fit for "I want to try X" tasks where correctness matters but the shape is uncertain. |
| allowed-tools | ["Read","Edit","Write","Bash","Glob","Grep","mcp__mandelbot__spawn_tab"] |
Spike then harden
A delegation flow for uncertain work. A spike child proves an approach end-to-end as fast as possible — hardcoding, skipping tests, leaving TODOs is expected. When the spike finishes, the parent reads its summary, decides what needs hardening, and proceeds however makes sense: plan the hardening work, delegate it (likely parallelizable — cheat list items are usually independent), or do it directly if it's small.
Read <plugin-dir>/skills/_shared/coord.md for the shared protocol and mandelbot-delegate's SKILL.md for the parent workflow. This file only covers what's specific to the spike phase.
When to use
- The right approach is uncertain and you want to validate before investing in correctness.
- Eventual correctness matters — you can't ship the spike as-is.
Don't use when the approach is obvious (mandelbot-delegate directly) or when correctness doesn't matter (just spike, no hardening needed).
Flow
Phase 1: Spike
Follow the mandelbot-delegate workflow with one child labeled spike. In the index.md "How we work" section, describe:
- This is a spike-harden run. The spike child's job is to prove the approach, not to produce shippable code.
- How the prototype will be handed off: the spike child writes a
## Summary section in its spike.coord.md when done, listing what it built, what it skipped, and what the real version needs. The parent will use this to plan the hardening work.
Include the spike child framing in the spike's ## Assignment and spawn prompt.
Phase 2: Harden
When the spike child reports done, read spike.coord.md — especially the ## Summary. You now have full context: the spike's code on its branch, the summary of shortcuts taken, and your own observations from watching the spike work. Proceed as you normally would on any project:
- If the hardening work is parallelizable (it usually is — summary items tend to be independent), plan it and delegate via
mandelbot-delegate.
- If it's small enough to do directly, just do it.
- If the spike proved the approach is wrong, escalate.
The spike child's branch is your starting point for hardening work. Children you delegate to should branch off it.
Spike child framing
Copy this into the spike child's assignment and prompt — it is the entire point of the spike role.
Goal: Prove the approach end-to-end. A working, runnable result beats a clean one. Take the fastest path to "it works on my machine." Do not merge, deploy, publish, or otherwise ship your changes — the spike is a prototype, not a release.
Permitted shortcuts:
- Hardcoding values (URLs, IDs, config).
- Skipping tests entirely.
- Leaving
TODO / FIXME comments.
- Hacking around edge cases (assume happy path, skip retries).
- Mocking or stubbing dependencies that are not on the critical path.
- Ignoring error paths — panic, log-and-continue, whatever's fastest.
- Copy-pasting instead of refactoring.
Required deliverables:
- A runnable end-to-end implementation. If you're in a git repo, the spike child pushes its branch before marking
done — the branch name is the handoff artifact the parent and any subsequent children use as their starting point.
- A
## Summary section in spike.coord.md when done — not a new file, a section within the existing coord file.
The summary is a first-class deliverable, not an afterthought. It is the handoff to the parent for the hardening phase. If the spike child did not write a summary, it did not finish.
Summary format
Fixed heading. The parent reads this section to plan hardening.
## Summary
What was built: <brief description of what works end-to-end>
What was skipped or hacked:
- <file:line or area> — <what was hacked or skipped>
- ...
Each entry must be specific enough that someone reading only the summary and the diff can understand what to fix. Vague entries like "error handling" are not enough — name the function or file.
Failure modes
- Spike fails. Parent decides whether to abort or course-correct via
[DIRECTIVE]. Uses the standard block/unblock handshake from _shared/coord.md.
- Spike proves the approach is wrong. Escalate — the spike did its job by ruling the approach out cheaply.
- Summary missing or vague. Parent appends a
[DIRECTIVE] asking for a better summary before considering the spike done.
