Use when implementing a feature or story, especially with acceptance criteria to verify. Triggers on /implement, build, create feature, add functionality.
From shieldnpx claudepluginhub infraspecdev/tesseract --plugin shieldThis skill uses the workspace's default tool permissions.
Implements structured self-debugging workflow for AI agent failures: capture errors, diagnose patterns like loops or context overflow, apply contained recoveries, and generate introspection reports.
Compares coding agents like Claude Code and Aider on custom YAML-defined codebase tasks using git worktrees, measuring pass rate, cost, time, and consistency.
Designs and optimizes AI agent action spaces, tool definitions, observation formats, error recovery, and context for higher task completion rates.
Plan sidecar: {output_dir}/*/plan.json (searches all feature plans, updates story status in place)
Structured feature implementation workflow with test-driven development, acceptance criteria confirmation, and continuous lightweight review.
The user provides one or more of:
EPIC-1-S1) — looked up in the plan sidecar JSONWhen NOT to use:
At startup, call execute-steps to register these steps. Execute them in order, updating status after each.
| Step | Action | Condition | Mandatory |
|---|---|---|---|
| 1 | Load story from plan.json | skip if no plan context | No |
| 2 | Confirm acceptance criteria | always | Yes |
| 3 | Write failing test | always (TDD) | Yes |
| 4 | Implement to pass test | always | Yes |
| 5 | Per-step review | always | Yes |
| 6 | Commit + update AC status in plan.json | always | Yes |
| 7 | Repeat 3-6 for next AC | loop until all AC done | Yes |
| 8 | Update story status in plan.json | always | Yes |
Per-AC tracking: Step 6 updates each AC's status in plan.json immediately after commit — not just at the end (step 8). This means plan.json is the per-AC source of truth. On resume, the skill reads plan.json to determine which ACs are already done, then continues the loop from the next incomplete AC. steps.json tracks coarse workflow position; plan.json tracks per-AC completion.
1. Load context (sidecar, PM card, or user description)
2. Confirm acceptance criteria with user
3. Plan implementation steps
4. Get user approval
5. TDD: write failing tests, implement, per-step review
6. Final verification against acceptance criteria
If plan sidecars exist in {output_dir}/*/plan.json:
If no plans found in {output_dir}/*/plan.json, check for legacy shield/docs/plans/<name>.json and suggest migration.
If a PM card URL/ID is provided and a PM tool is configured:
pm_get_capabilities to verify the adapter is availableIf no sidecar or PM card, gather requirements by asking:
ALWAYS confirm acceptance criteria before starting implementation.
Present the criteria to the user:
Story: EPIC-1-S1 — IPAM Pool Hierarchy
Acceptance Criteria:
1. Regional pools allocate /20 CIDRs
2. No CIDR overlap across regions
3. Rollback without data loss
Confirm before starting implementation:
[a] Proceed as-is
[b] Edit criteria (add/remove/modify)
[c] Skip — implement without formal criteria
If the user edits criteria:
Do NOT skip this step. The confirmed criteria are what /review will verify against.
Explore the codebase to understand existing patterns, then create a plan:
If superpowers is available, delegate to superpowers:writing-plans for the plan structure.
Show the plan summary. NEVER proceed without explicit approval.
For each step:
Write a test that captures the expected behavior. Run it — confirm it FAILS.
Write the minimal code to make the test pass.
After each step passes its test:
terraform/review), run domain-specific checks on the changed filesCommit the step immediately:
git add <specific files>
git commit -m "feat(<feature-name>): step N — <description>"
If implementing from a sidecar story, update the story status:
"status": "in-progress""status": "in-review"
Re-render HTML after status change.After all steps complete:
Run all tests — every test must pass
Run linters/formatters — code must be clean
Check each acceptance criterion against the implementation:
| Criteria | Status | Evidence |
|---|---|---|
| Regional pools allocate /20 CIDRs | Met | main.tf:42 — netmask_length = 20 |
| No CIDR overlap | Met | tests/overlap.tftest.hcl |
Update sidecar — mark story as "in-review" if all criteria met
Invoke shield:summarize — produce an implementation summary
/review — run full agent-based code review with AC verification/pm-sync — update story status in the PM tool| Mistake | Fix |
|---|---|
| Skipping AC confirmation | Always confirm before implementing |
| Implementing before tests | Write test first, confirm it fails |
| No per-step review | Check each step before committing |
| Not updating sidecar | Update status and re-render HTML |
| Giant end-of-feature commit | Commit after each step |
| Skipping verification | Check every AC before claiming done |