Ad Spike

/ad-spike

Implements WORKFLOW.md §14 (Staged Spikes With Golden Fixtures) end-to-end. The skill is for cases where the spec is clear but the technique is uncertain across multiple plausible approaches — library choice, CV approach, multi-stage transformation. WORKFLOW §9 (TDG) assumes the path is known and validates end-to-end; §14 assumes the path is unknown and validates per stage. Different uncertainty regimes; this skill is for the unknown one.

The skill creates a working directory under spikes/NNNN-<slug>/ and fills it stage-by-stage. The directory is throwaway by design — when the spike concludes, an ADR records the decision (/ad-adr) and the spike directory is deleted. See ADR-0017 for the promote-or-delete lifecycle rationale.

Step 0 — Confirm uncertainty

The skill is for unknown technique across multiple plausible approaches, not for non-trivial work in general. If a single happy path is obvious, do not start a spike. If the how is knowable from official docs / implementation references / in-repo patterns / git history, route to ad-ground and stop.

Concrete tests to run before starting:

• Could ad-ground's four-source research surface a single happy path with a defensible deviation gate? If yes, run that instead.
• Are there ≥2 candidate techniques with materially different trade-offs that no source resolves? If no, this is not a spike.
• Is end-to-end validation against expected outputs feasible without per-stage debug? If yes, this is ad-task + ad-philosophy Goal-Driven Execution territory, not a spike.

If the spike is warranted, confirm with the user the recortte (the specific surface where uncertainty sits — not the whole feature) and proceed.

Step 1 — Discovery

List canonical approaches grounded in official docs and real examples. Pick one (or a small set, ≤3) by an explicit criterion.

Candidate-listing process:

1. Search official documentation for the language / library / domain in question. Cite URL + version.
2. Search public implementation references (open-source repos, Stack Overflow / forum answers, blog posts, gists) for solutions to the same technical recortte. Cite <source>:<locator> — <repo>:<path>:<line-range> for repos, <URL> for Stack Overflow / blog / gist — and fetch via tools; never paraphrase from training memory.
3. Survey in-repo for analogous patterns the codebase already uses. Cite <file>:<line> or "no analog found".
4. Survey git history for prior attempts at the same problem. Cite <commit-sha> or "no prior attempt".

Output format:

## Discovery — <recortte>

### Candidate techniques
1. **<name>** — <one-line description>. Source: <URL or repo:path>. Trade-offs: <pros / cons>.
2. **<name>** — ...
3. **<name>** — ...

### Selection criterion
<one-line criterion: latency / accuracy / readability / dependencies / etc>

### Picked
<technique X>, picked by criterion <Y>. Alternatives held in reserve: <list>.

The output of this step is information, not code. No spike directory is created yet. The user reviews the candidate list and confirms the picked approach (or revises) before Step 2.

Step 2 — Golden fixture

Curate inputs with rich expected outputs. The fixture is the ground truth the staged pipeline validates against; richer fixtures catch more failure modes.

Create the spike directory:

mkdir -p spikes/NNNN-<slug>/{fixtures,debug,eval}

Where NNNN is the next available 4-digit number (mirrors ADR / task / spec numbering). List spikes/ and pick the next slot.

The fixture format is JSON keyed by input path (recommended) or whatever shape the domain demands. For computer vision: bounding boxes, sizes, lighting condition, difficulty tag, edge case markers. For multi-stage transformations: intermediate states. For library choice: representative inputs covering typical and edge cases.

Example fixture file (spikes/0001-detect-circles/fixtures/golden.json):

{
  "inputs/easy-01.jpg": {
    "expected": [
      { "bbox": [120, 80, 240, 200], "label": "circle", "size": "large", "lighting": "even" }
    ],
    "difficulty": "easy",
    "edge_cases": []
  },
  "inputs/hard-01.jpg": {
    "expected": [
      { "bbox": [50, 60, 90, 100], "label": "circle", "size": "small", "lighting": "low" },
      { "bbox": [200, 80, 260, 140], "label": "circle", "size": "medium", "lighting": "even", "occluded": true }
    ],
    "difficulty": "hard",
    "edge_cases": ["low-light", "partial-occlusion", "multiple-objects"]
  }
}

Curation principles:

• Include edge cases (low light, partial occlusion, malformed inputs, large inputs, empty inputs) — not just "happy path" examples. The fixture's job is to surface where a technique fails, not just whether it succeeds on easy cases.
• Include difficulty tags so per-stage evaluation can report performance segmented by difficulty.
• Keep the fixture as data, not code. JSON / YAML / CSV — anything that diffs cleanly and survives a refactor.

The fixture is the contract the pipeline validates against. Treat it like spec text — it should not change once the spike runs unless ground truth itself changes.