TLC AI Dev Flow · v1 · Skills for agentic software factories

The skills that run the factory, not the session.

An agentic software factory is the development loop as an event-driven system, not a manual chat. Writing code is no longer the bottleneck. Verification is. The human does not leave the loop. They move to the ends: intent and direction.

npx @tech-leads-club/agent-skills install --skill tlc-discover tlc-plan tlc-implement the-judge

Works with Cursor, Claude Code, Copilot, Windsurf, Cline and 14 others.

Event-driven, not a session

Work arrives from an issue, a user request, an incident alert or the backlog. Every item enters in the same shape: scope, acceptance criteria, owner.

Human at the ends

The human leaves the middle (execution) and holds the ends: define intent before code exists, then validate direction. A decision now costs a sentence. After two thousand lines, it costs a rewrite.

Verification is the bottleneck

Code got cheap. Proof did not. The factory spends its expensive steps on layered verification, not on typing. The human reads the PR for direction, not to hunt bugs the suite already caught.

Vertical slices

Build a thin end-to-end flow first, then deepen it. Left alone, a model builds horizontally and leaves nothing testable in the middle.

Layered verification

Cheap first (lint, types, unit). Expensive near the PR (integration, browser, mutation, security). Two sanity tests: run the agent-written test against pre-patch code, and discard edits it made to existing tests.

v1 is the core loop

Discover, plan, implement and the-judge ship today. Intake, triage and production feedback are stations on the same line. They come next.

How the factory runs

Seven stations on one line. v1 lights up research, plan, implement and the PR gate. Intake, triage and production are the same factory, still being built.

1
ENTRYcoming

Work arrives as an event

Issue, user request, incident alert or backlog item. Same format every time: scope, acceptance criteria, owner.

2
TRIAGEcoming

Label is queue, lock and pause

An agent classifies: ready to build, needs a spec, missing information, wait. The label stops two agents grabbing the same issue, and it is the human pause point.

3
RESEARCH

Define the problem

tlc-discover. Bug often skips this. A feature does not. Problem, success metric, architecture and signatures, before code exists.

4
PLAN

Cut vertical slices

tlc-plan. Observable outcomes with concrete values. Each slice proves something a later agent can build without guessing.

5
IMPLEMENT

Build and prove

tlc-implement. Isolated sandbox, own branch, vertical slice, cheap checks first. An independent verifier proves every check. The author never verifies their own work.

6
GATE

Judge the PR

the-judge. Evidence-first review, one consolidated GitHub review. Deterministic checks first, then judgment. Branch protection still requires a human. They read for direction.

7
PRODUCTIONcoming

Ship, watch, re-enter

Deploy and monitor. An incident or a piece of feedback becomes a new issue and re-enters stage 1.

The v1 skills

Four skills, one core loop. Discover decides. Plan cuts. Implement proves. The Judge reviews.

1DISCOVER

Interview the idea

Turns an unshaped problem into a verdict and a design document with literal decisions. The expensive judgment lives here: the problem, the success metric, the architecture and the signatures, before a line of code exists.

tlc-discover
2PLAN

Cut decided work into tasks

Takes work that is already decided and writes tasks a builder can act on without guessing. Each slice proves something. Criteria are observable outcomes with concrete values, not adjectives.

tlc-plan
3IMPLEMENT

Build, then prove it

Extracts a checklist from the plan, builds in vertical slices, and proves every check with an independent verifier. The author is never the verifier. Done is an exit code, not a self-report.

tlc-implement
4JUDGE

Review with evidence

Evidence-first PR judge. Runs lint, types and tests first, researches current docs, then posts one consolidated GitHub review. Every finding carries a citation. The verdict is APPROVE, COMMENT or REQUEST_CHANGES. The human still merges.

the-judge
The bottleneck moved

Software factories are not new

The idea dates to 1968. What changed is which station on the line became an agent, and where the bottleneck went. Code got cheap. Proof did not.

Before

  • Human in the middle

    Execution is the job. The expensive hour is writing the code.

  • Work lives in a session

    A chat is the queue, the lock and the memory. Nothing arrives in a shared format.

  • Code is the bottleneck

    Verification is a review bolted on at the end, if it happens at all.

Now

  • Human at the ends

    Define intent before code exists. Validate direction after proof. Execution is an agent station.

  • Work arrives as an event

    Issue, Slack, alert, backlog. Same shape every time, so the next station does not improvise.

  • Verification is the bottleneck

    Cheap checks first, expensive ones near the PR. The human reads for direction, not to hunt bugs.

Two sanity tests at the verification station

The pre-patch test

Run the test the agent wrote against the code from before the patch. If it still passes, it tests nothing.

Do not let it rewrite the suite

Discard edits the agent made to existing tests. A green suite that it loosened is not proof.

The human stays in the loop

The human doesn't leave the loop. They change position.

v1 covers research, plan, implement and PR review. Intake, triage and production land in later versions of the same line.

Define intentValidate directionLayered verificationEvidence-first review

Start the factory with the core loop

Four skills, one command, works with your agent today.

npx @tech-leads-club/agent-skills install --skill tlc-discover tlc-plan tlc-implement the-judge