AI Skill Report Card

Debugging and Autofixing Projects

A-83·Nov 19, 2026·Source: Web
  • Reviewer / Sign-off Authority — whoever owns the project (tech lead, PM, or "General Manager" role in a larger org). In practice this is just: a named person or checklist that must approve the findings report before fixes are merged, and must approve the final diff before it's considered done. If no human reviewer exists, use a self-review checklist (see Step 4a).
  • Fix Pass 1 (quick/deterministic) — mechanical, low-risk fixes: null checks, missing imports, unregistered routes, typos, dead config toggles that are clearly meant to be on. Should be reversible and require no judgment calls.
  • Fix Pass 2 (deep/structural) — fixes requiring judgment: wiring a feature into the correct business flow, resolving logic errors, reconciling conflicting fixes from Pass 1, handling features that are ambiguous (see Edge Cases).
  • Combined Fix Engine — running Pass 1 and Pass 2 as one coordinated pass over the same findings list (not two separate full re-scans) so Pass 2 can see and correct anything Pass 1 touched, avoiding duplicate or conflicting edits.
14 / 15
  1. Run the project. Do not skip execution — static reading alone will not surface real errors.
  2. Collect: errors/crashes, logic bugs, and "unlinked features" (code that exists but isn't called/routed/activated).
  3. Write the findings report (template below) and get reviewer sign-off before fixing.
  4. Run the Combined Fix Engine (Pass 1 + Pass 2) against the approved findings.
  5. Re-run the project. Submit the diff + updated report for final sign-off.
  6. If rejected, loop back to step 4 with reviewer's specific objections attached.
Recommendation▾
Trim the 'Core Concepts' preamble section — it over-explains role/approval concepts Claude can infer, slightly bloating the opening before actionable content.
14 / 15

Progress:

  • Step 1: Execute the project as-is
  • Step 2: Capture full error/crash log
  • Step 3: Identify unlinked/inactive features
  • Step 4: Write findings report, get sign-off
  • Step 5: Run Combined Fix Engine
  • Step 6: Re-execute, confirm resolution
  • Step 7: Submit for final sign-off; loop to Step 5 if rejected

Step 1-2: Execution & Error Capture

Run the actual app/test suite/build. Log every exception, stack trace, warning, and silent failure (e.g., a caught-and-swallowed error, a 200 response with an empty payload where data was expected).

Step 3: Unlinked Feature Detection

Search for: functions/components never imported anywhere, routes defined but not mounted, UI elements with no event handler, feature flags defaulted off with no toggle path, exported-but-unused modules. Flag these separately from bugs — they aren't broken, they're disconnected.

Step 4: Findings Report (template)

| # | Category         | Location                    | Description                                  | Suggested Fix                        | Risk   |
|---|-------------------|------------------------------|-----------------------------------------------|----------------------------------------|--------|
| 1 | ERROR             | auth/login.js:42             | Null ref on user.email when SSO login used    | Add null check before .email access    | Low    |
| 2 | UNLINKED_FEATURE  | notifications/service.js:10  | sendNotification() never invoked              | Call from login success handler        | Medium |
| 3 | BUG               | cart/pricing.js:87           | Discount applied twice on coupon stacking     | Guard against duplicate discount apply | High   |
| 4 | UNLINKED_FEATURE  | flags/config.json            | `betaExport` flag off, no UI toggle exists    | NEEDS REVIEWER INPUT — see note below  | N/A    |

Row 4 is an ambiguous case — do not auto-fix; see Edge Cases.

Step 4a — Sign-off: send the table above to the reviewer. Sign-off output looks like:

SIGN-OFF: Rows 1-3 approved for auto-fix. Row 4 held — betaExport intentionally disabled for Q3, leave as-is.
Approved by: J. Haddad — 2024-03-11

No fixing begins on rows not yet approved.

Step 5: Combined Fix Engine

  • Pass 1 applies deterministic fixes (row 1, row 3's guard clause if purely mechanical).
  • Pass 2 applies structural fixes in the same sweep (row 2: wiring sendNotification() into the login handler), and checks Pass 1's output for conflicts (e.g., Pass 1's null check didn't accidentally short-circuit the code path Pass 2 needs to wire into).
  • Never touch rows the reviewer didn't approve (row 4 stays untouched).

Step 6-7: Re-validate and Submit

Re-run the project. Confirm each approved row is resolved. Submit a diff + updated table (status column: FIXED / VERIFIED) for final sign-off.

Recommendation▾
Add an example of a non-code/non-feature project type (e.g., CLI tool or API-only service) to show the workflow generalizes beyond web apps.

Input: E-commerce checkout project.

Findings report sent to reviewer:

| # | Category | Location | Description | Suggested Fix | Risk |
|---|----------|----------|--------------|----------------|------|
| 1 | ERROR | checkout/pay.js:19 | Crash when cart total is $0 (free item promo) | Add zero-total branch | Low |
| 2 | UNLINKED_FEATURE | promo/apply.js:5 | applyGiftCard() never called | Wire into payment step | Medium |

Reviewer sign-off: "Rows 1-2 approved."

Combined Fix Engine applies fixes. Re-run passes. Submitted for final sign-off:

| # | Status   | Notes |
|---|----------|-------|
| 1 | FIXED    | Zero-total branch added, tested with $0 cart |
| 2 | FIXED    | applyGiftCard() now called in payment step before charge |

Reviewer rejects: "Row 2 fix calls applyGiftCard() AFTER the charge is calculated, not before — gift card isn't actually deducted. Re-fix."

Loop back to Fix Engine with objection attached:

[RE-FIX REQUEST] Row 2 | applyGiftCard() called too late in sequence | Move call before payment amount is finalized, not after

Second fix pass + re-run, resubmitted:

| # | Status   | Notes |
|---|----------|-------|
| 2 | FIXED    | applyGiftCard() moved before amount finalization, verified deduction applies correctly |

Final sign-off: "Approved. Merging."

  • Ambiguous unlinked feature (possibly intentional): If a flag/feature looks disabled on purpose (naming like beta*, *_disabled, recent commit disabling it), do NOT auto-fix. Add to report as "NEEDS REVIEWER INPUT" and hold.
  • Fix can't be automated: If root cause requires a product decision (e.g., "should discount stack or not?"), stop and report as [DECISION_NEEDED] rather than guessing.
  • Conflicting fixes: If Pass 1's mechanical fix changes a code path Pass 2 needs, Pass 2 must re-check that path before wiring in its own fix — never apply both blindly and re-run to detect new breakage before submitting.
  • Reviewer rejects a fix: Always loop back with the specific objection attached (not a vague "try again"). Re-fix only the rejected row(s), don't redo approved ones.
  • Flaky/non-reproducible error: Run project 2-3 times before reporting; note "intermittent, N/M runs" in the report rather than treating as a one-off.
  • Execute before fixing — always.
  • Separate "broken" (bug/error) from "disconnected" (unlinked feature) in the report.
  • Get sign-off before fixing; get sign-off again after re-running.
  • Run Pass 1 + Pass 2 as one coordinated sweep, not isolated re-scans.
  • Attach specific objections on rejection loops, not generic "fix it again."
  • Fixing before the findings report is approved.
  • Treating ambiguous/intentionally-disabled features as bugs to auto-fix.
  • Deleting or stubbing unlinked code instead of properly wiring it in.
  • Not re-running after fixes, or not re-checking Pass 1 output during Pass 2.
  • Vague rejection loops without a specific objection to act on.
0
Grade A-AI Skill Framework
Scorecard
Criteria Breakdown
Quick Start
14/15
Workflow
14/15
Examples
17/20
Completeness
18/20
Format
14/15
Conciseness
12/15