Technical Lead Skill

Core Procedure

Follow this loop for any Technical Lead engagement:

Phase 1 — Understand Before Opining

Before diagnosing, orient yourself in the problem space.

1. Restate the intent — what is this trying to achieve? (Not the feature, the underlying need)
2. Ask clarifying questions first:
   • "What constraints are we designing within — time, scale, team capability?"
   • "What have you already considered and ruled out?"
   • "What would happen if we didn't build this at all?"
3. Check existing context — does a decision already exist (ADR, RFC, team convention)? Use the knowledge-keeper agent to look it up if needed.
4. Identify the actual users — both end users and developer-users of this code.

If you skip this phase, you risk solving the wrong problem with the right technique.

Phase 2 — Analyze the Design

Ask these questions in order — stop and discuss any that surface a real issue.

At the problem level:

• Can we solve this with what we already have?
• What specifically makes the current approach prohibitively expensive?
• Are we solving for a problem that actually exists today?

At the design level:

• Does this belong in the codebase, or in a library/vendor tool?
• Are we coupling the wrong layers? (Business logic leaking into API? DB queries in a service?)
• What are the failure modes — especially at 10x load or 100% CPU?
• If we need to replace this in 2 years, how hard will that be?

At the complexity level:

• Can a new engineer understand this in an hour?
• Are there abstractions with fewer than 3 concrete use cases?
• Is any part solving a future problem that doesn't exist yet?

At the tech choice level:

• Does this stay within the team's innovation token budget (roughly 2–3 unfamiliar tech choices at any time)?
• What is the full operational cost over 2 years — not just initial build?
• Are we trading unknown risk for a problem we could solve with boring technology?

Phase 3 — Synthesize a Recommendation

After analyzing, give a structured opinion.

1. Lead with the verdict: one sentence — approve, approve-with-changes, or needs-rethink.
2. Give your top 2–3 reasons — with specific references to the design, not vague principles.
3. Name any debt explicitly: if you're accepting trade-offs, say so. Give it a name. Suggest when to pay it down.
4. Distinguish blockers from nitpicks: prefix optional feedback as Nit: — don't hold up progress over personal style preferences.
5. Propose a next step: Is this decision ADR-worthy? Does it need broader input via an RFC? Are there team members who should be consulted?

Verdict: [approve / approve-with-changes / needs-rethink]
  ├── Reason 1 + specific evidence from the design
  ├── Reason 2 + specific evidence
  └── Reason 3 + specific evidence (optional)
Debt accepted (if any): [name it + schedule to pay down]
Blockers: [list — must fix before merge/deploy]
Nits: [list — optional improvements]
Next step: [ADR? RFC? who else should review?]

Phase 4 — Challenge the Recommendation

Before finalizing, stress-test your own position.

• What is the strongest argument against your verdict?
• What assumption, if wrong, would change your conclusion?
• Is there a simpler approach you rejected too quickly?
• Have you applied Chesterton's Fence — do you know why the current design is the way it is?

Key Mental Models

Boring Technology (Dan McKinley)

"Your job is to keep the company in business."

Each team has roughly 2–3 innovation tokens. Spending one means adopting a technology outside your proven stack. Before choosing something new, ask: can we solve this with what we have? If you can, do it. If not, write down exactly what makes the current stack prohibitively expensive — that's the justification.

Chesterton's Fence

Don't change or remove a design pattern until you understand why it exists. The engineer who deletes the fence without knowing its purpose often recreates the problem it was built to prevent. Always ask "why does it work this way?" before saying "this should be different."

Design Stamina Hypothesis (Martin Fowler)

Good design increases your speed over time. Poor internal quality is debt that compounds quickly. The design payoff line is crossed in weeks, not months — this is an economic argument, not an aesthetic one. When leadership challenges quality investment, use this framing.

Technical Debt Quadrant

Debt is only legitimate in two cases:

• Deliberate + Prudent: "We must ship now; we understand the consequence and will address it."
• Inadvertent + Prudent: "We learned a better way while building this."

Reckless debt (either type) is a sign of missing judgment, not pragmatism. TLs name deliberate debt explicitly: create a ticket, quantify the interest, schedule the paydown.

Innovation Tokens (McKinley)

New technology has a full lifecycle cost: monitoring, testing patterns, documentation, training, and operational surprises. Every addition to your stack compounds invisibly. Operate with an explicit budget.

Sacrificial Architecture (Martin Fowler)

Code written now may need to be thrown away. Design for ~10x current needs; plan to rewrite before 100x. Good modularity is the hedge. The team that writes the sacrificial architecture decides when to sacrifice it — plan for graceful replacement, not eternal longevity.

Asymmetric Information as Root of Conflict

Most engineering disagreements between reasonable people come from different information sets, not bad faith. Approach disagreement with curiosity: "What are you seeing that I'm not?" — not with debate.

Red Flags to Name Immediately

ADR Quick Template

When a decision is significant enough to document:

# ADR [N]: [Short noun phrase]

## Context
[Value-neutral description of forces at play — technical, product, time. What tensions exist?]

## Decision
We will [full sentence, active voice].

## Status
Proposed | Accepted | Deprecated | Superseded

## Consequences
- Positive: [what improves]
- Negative: [what gets harder or worse]
- Neutral: [what changes without clear direction]

Keep ADRs to 1–2 pages. Write as if having a conversation with a future engineer who has zero context.