Implementation

Verification-driven coding with tight feedback loops. Distilled from 21,321 tracked operations across 64+ projects, 612 debugging sessions, and 2,476 conversation histories. These are the patterns that consistently ship working code.

Core insight: 2-3 edits then verify. 73% of fixes go unverified — that's the #1 quality gap. The difference between a clean session and a debugging spiral is verification cadence.

The Sequence

Every implementation follows the same macro-sequence, regardless of scale:

digraph implement {
    rankdir=LR;
    node [shape=box];

    "ORIENT" [style=filled, fillcolor="#e8e8ff"];
    "PLAN" [style=filled, fillcolor="#fff8e0"];
    "IMPLEMENT" [style=filled, fillcolor="#ffe8e8"];
    "VERIFY" [style=filled, fillcolor="#e8ffe8"];
    "COMMIT" [style=filled, fillcolor="#e8e8ff"];

    "ORIENT" -> "PLAN";
    "PLAN" -> "IMPLEMENT";
    "IMPLEMENT" -> "VERIFY";
    "VERIFY" -> "IMPLEMENT" [label="fix", style=dashed];
    "VERIFY" -> "COMMIT" [label="pass"];
    "COMMIT" -> "IMPLEMENT" [label="next chunk", style=dashed];
}

ORIENT — Read existing code before touching anything. Grep -> Read -> Read is the dominant opening. Sessions that read 10+ files before the first edit require fewer fix iterations. Never start with blind changes.

PLAN — Scale-dependent (see below). Skip for trivial fixes, write a task list for features, run a research swarm for epics.

IMPLEMENT — Work in batches of 2-3 edits, then verify. Follow the dependency chain. Edit existing files 9:1 over creating new ones. Fix errors immediately — don't accumulate them.

VERIFY — Typecheck is the primary gate. Run it after every 2-3 edits. Run tests after feature-complete. Run the full suite before commit.

COMMIT — Atomic chunks, committed as you go. Verify, stage specific files, commit, then loop back to the next chunk. Many small commits per session is the norm. See Commit Cadence below for message anatomy.

Code Discipline

Behavioral lens that governs every phase of the loop. Adapted from Karpathy's observations on LLM coding pitfalls — models "make wrong assumptions on your behalf and just run along with them without checking ... they really like to overcomplicate code and APIs, bloat abstractions ... implement a bloated construction over 1000 lines when 100 would do."

These principles bias toward caution over speed. For trivial fixes, use judgment.

Think before coding

Don't assume. Don't hide confusion. Surface tradeoffs.

ORIENT (read the code) is the prerequisite. This principle is what to do with what you find: name the gaps, don't paper over them.

Simplicity first

Minimum code that solves the problem. Nothing speculative.

The test: would a senior engineer call this overcomplicated? If yes, simplify.

Surgical changes

Touch only what you must. Clean up only your own mess.

The test: every changed line should trace directly to the user's request.

Goal-driven execution

Define verifiable success. Loop until it passes.

For multi-step work, state the plan with verification per step:

1. [Step] → verify: [check]
2. [Step] → verify: [check]
3. [Step] → verify: [check]

Strong success criteria let you loop independently. Weak criteria require constant clarification.

Scale Selection

Strategy changes dramatically based on scope. Pick the right weight class:

Skip planning when: Scope is clear, single-file change, fix describable in one sentence.

Plan when: Multiple files, unfamiliar code, uncertain approach.

Dependency Chain

Build things in this order. Validated across fullstack, Rust, and monorepo projects:

Types/Models -> Backend Logic -> API Routes -> Frontend Types -> Hooks/Client -> UI Components -> Tests

Fullstack (Python + TypeScript):

1. Database model + migration
2. Service/business logic layer
3. API routes (FastAPI or tRPC)
4. Frontend API client
5. React hooks wrapping API calls
6. UI components consuming hooks
7. Lint -> typecheck -> test -> commit

Rust:

1. Error types (thiserror enum with #[from])
2. Type definitions (structs, enums)
3. Core logic (impl blocks)
4. Module wiring (mod.rs re-exports)
5. cargo check -> cargo clippy -> cargo test

Key finding: Database migrations are written AFTER the code that needs them. Frontend drives backend changes as often as the reverse.

Verification Cadence

The single most impactful practice. Get this right and everything else follows.

The Edit-Verify-Fix Cycle

The sweet spot: 3 changes -> verify -> 1 fix. This is the most common successful pattern.

The expensive pattern: 2 changes -> typecheck -> 15 fixes (type cascade). Prevent by grepping all consumers before modifying shared types.

Combined gates save time: turbo lint:fix typecheck --filter=pkg runs both in one shot. Scope verification to affected packages, never the full monorepo.

Practical tips:

• Run lint:fix BEFORE lint check to reduce iterations
• cargo check over cargo build (2-3x faster, same error detection)
• Truncate verbose output: 2>&1 | tail -20
• Wrap tests with timeout: timeout 120 uv run pytest

Decision Trees

Read vs Edit

Familiar file you edited this session?
  Yes -> Edit directly (verify after)
  No  -> Read it this session?
    Yes -> Edit
    No  -> Read first (79% of quick fixes start with reading)

Subagents vs Direct Work

Self-contained with a clear deliverable?
  Yes -> Produces verbose output (tests, logs, research)?
    Yes -> Subagent (keeps context clean)
    No  -> Need frequent back-and-forth?
      Yes -> Direct
      No  -> Subagent
  No -> Direct (iterative refinement needs shared context)

Refactoring Approach

Can changes be made incrementally?
  Yes -> Move first, THEN consolidate (separate commits)
        New code alongside old, remove old only after tests pass
  No  -> Analysis phase first (parallel review agents)
        Gap analysis: old vs new function-by-function
        Implement gaps as focused tasks

Bug Fix vs Feature vs Refactor

Error Recovery

65% of debugging sessions resolve in 1-2 iterations. The remaining 35% risk spiraling into 6+ iterations.

Quick Resolution (Do This)

1. Read relevant code first (79% success correlation)
2. Form explicit hypothesis: "The issue is X because Y"
3. Make ONE targeted fix
4. Verify the fix worked

Spiral Prevention (Avoid This)

1. Separate error domains — fix ALL type errors first, THEN test failures. Never interleave.
2. 3-strike rule — after 3 failed attempts on same error: change approach entirely, or escalate.
3. Cascade depth > 3 — pause, enumerate ALL remaining issues, fix in dependency order.
4. Context rot — after ~15-20 iterations, /clear and start fresh. A clean session with a better prompt beats accumulated corrections every time.

The Two-Correction Rule

If you've corrected the same issue twice, /clear and restart. Accumulated context noise defeats accuracy.

Commit Cadence

Commit as you go. Each commit captures one logical chunk that has been built, verified, and tested. Many small commits per session is the norm — never accumulate hours of unrelated work into a single mega-commit. The COMMIT step in the macro-sequence loops back to IMPLEMENT for the next chunk; that loop is the rhythm.

When to commit

Rule of thumb: if a reviewer would want to read it as a separate diff, it's a separate commit.

Local style first

Before the first commit in any repo, detect the local style. Repos have personalities — your commits should match.

git log -10 --oneline

Observe the actual patterns and mirror them:

Mirror format, NOT quality. If existing commits are terse one-liners, you still write a descriptive subject and body — you're elevating the standard, not lowering yours to match. If no clear pattern exists, default to Conventional Commits.

Conventional Commits (default)

Format: type(scope): subject. Scope is optional but encouraged when the change is localized.

Message anatomy

Subject line:

Body (always include one):

• Explain why, not what — the diff shows what
• Wrap at ~72 chars, separate from subject with one blank line
• Two sentences is often enough; a few short paragraphs when warranted
• Mention load-bearing context: hidden constraints, related issues, what a future bisect would want to know
• No uncertain language. Banish "likely", "probably", "might", "seems", "appears to", "presumably". You wrote the code — state facts. If you don't know what a change does, read more before committing.

Co-Authored-By

When an agent does meaningful work in a commit, add a Co-Authored-By: trailer with a descriptive identity that names the model. This makes attribution legible across multi-agent sessions and git log.

Co-Authored-By: Nova (Claude Opus 4.7) <noreply@anthropic.com>

Use a name that signals which model or persona contributed — not just "Claude".

The HEREDOC pattern

Always pass commit messages via HEREDOC to preserve formatting and avoid shell-quoting bugs:

git commit -m "$(cat <<'EOF'
fix(auth): guard against null session in token refresh

Refresh requests racing with logout were dereferencing a freed session
pointer, surfacing as a 500 with no log trail. Added an early return
that emits a single warn log so the failure mode is visible without
spamming on every refresh attempt.

Co-Authored-By: Nova (Claude Opus 4.7) <noreply@anthropic.com>
EOF
)"

Examples

Multi-agent hygiene

Other agents may be working in parallel. Stage with care:

git status                # See the full picture first
git diff --staged         # Review what you're about to commit
git add <specific-files>  # Only files you personally touched
git commit -m "..."       # HEREDOC for the message

Anti-Patterns

Cross-Model Review

For high-stakes changes, use /hyperskills:codex-review after implementation. A fresh model context eliminates implementation bias and catches real bugs: migration idempotency, PII in debug logging, empty array edge cases, missing batch limits.

References

For quantitative benchmarks and implementation archetype templates, consult references/benchmarks.md.

What This Skill is NOT

• Not a gate. Don't follow all five phases for a typo fix. Scale selection exists for a reason.
• Not a replacement for reading code. This skill tells you HOW to implement, not WHAT to implement.
• Not a planning tool. Use /hyperskills:plan for task decomposition.
• Not an excuse to skip tests. "Verify" means running actual checks, not eyeballing the diff.