Vol. 1 · Issue 01 A CDD.DEV PRODUCTION Issue 02 · Swarm → 10¢ · still cheaper than re-deriving it
Pulp sci-fi panorama: a brass robot kneels at a lunar pool seeing its upgraded chrome self, beneath a vast glowing ring of stars.

Astounding tales of

SELF‑IMPROVINGMACHINES

Starring reflect — agent skills that turn one session's friction into every future session's head start.

npx skills add ckorhonen/reflect
Source on GitHub →

Every session with a coding agent produces two things: the work, and a trail of friction — corrections you repeated, facts rediscovered the hard way, tokens burned on output nobody read. The work ships. The friction evaporates, and you pay for it again tomorrow.

reflect captures it instead. And compounds it.

This issue's feature

THE LOOP

Five panels. Runs at the end of a session, or on a schedule. Entirely hands‑off: safe improvements apply themselves, risky ones queue for your approval.

  1. 1

    The session ends

    A transcript full of evidence: what you corrected, what stalled, what the agent figured out the hard way.

  2. 2

    Signals surface

    Repeated corrections. Wrong first guesses. The same three commands reconstructed from scratch. One tool call hogging half the context.

  3. 3

    Each signal gets a lane

    A rule, a skill, a slash command, a script, a delegation, a context diet — one signal, one lane, one smallest-possible change.

  4. 4

    The gate

    Evidence-backed? Durable? Not already written down? Testable? Anything that can't answer yes four times gets logged as a hunch — or dropped.

  5. 5

    Apply & ledger

    Changes land in your repo; every one gets a ledger row. The next session starts already knowing what this one learned.

↻ and the machine wakes up a little sharper tomorrow.

Collect all four

THE BACK ISSUES

One package, four skills. Install the set, or pick your issues.

A brass robot studies a ticker tape of its own actions through a magnifying glass.

Issue No. 1

reflect

“The Machine That Read Its Own Mind”

The core loop. Reviews the session, routes findings into the eight lanes, applies what clears the gate, ledgers everything. Ships a transcript analyzer so it reads evidence, not the whole log.

npx skills add ckorhonen/reflect@reflect
A chrome robot checks a wall of analog gauges, clipboard in hand.

Issue No. 2

reflect‑feedback

“The Gauges Never Lie”

Reflection without feedback just accumulates. Audits the ledger: verifies improvements that demonstrably fire, watches the young ones, prunes dead weight. Never invents usage.

npx skills add ckorhonen/reflect@reflect-feedback
A robot librarian on a ladder prunes an endless card-catalog archive; discarded cards dissolve into light.

Issue No. 3

reflect‑memory

“The Archive That Ate Itself”

Compounding's failure mode is unbounded growth. Dedupes, merges, expires, and de-contradicts accumulated rules — and reports the context-token load before and after.

npx skills add ckorhonen/reflect@reflect-memory
A robot mechanic swaps interchangeable rocket tools from a hangar rack.

Issue No. 4

reflect‑skills

“The Hangar of a Thousand Tools”

Scans weeks of local session logs, clusters the recurring failures, and treats your skill collection as the fix surface: sharpen a trigger, install from the ecosystem, author new, or retire a misfirer.

npx skills add ckorhonen/reflect@reflect-skills

Field manual, plate II

THE EIGHT LANES

Every signal a session produces gets routed into exactly one of these.

01

Agent rules

One-line additions to AGENTS.md / CLAUDE.md the agent should never rediscover.

02

Skills

Workflows with enough judgment to deserve their own trigger — authored, or installed from the ecosystem.

03

Slash commands

The request you keep typing, reduced to one keystroke of intent.

04

Deterministic scripts

Fragile hand-repeated parsing and checking, encoded as code. If an LLM isn't needed to verify it, an LLM shouldn't do it twice.

05

Delegation

Bounded, checkable work routed to smaller, cheaper models. The main thread graduates to orchestrator.

06

Context diet

Verbose tool output wrapped, filtered, or capped — measured in tokens, not vibes.

07

Observations

Maybe-patterns logged with a sighting count until they earn promotion.

08

No action

A confident, recorded “not worth it.” The lane that keeps the other seven honest.

Exhibit A

THE LEDGER

Cross-session memory is a .reflect/ folder of plain markdown — readable, reviewable, committable. Team memory, not a black box. Every improvement carries its evidence and earns its status.

Reader mail

LETTERS TO THE EDITOR

“What IS reflect, exactly?”

Four agent skills that review your sessions and extract durable improvements — rules, skills, commands, scripts, delegation, context diets — then log each one to a ledger so the gains compound. Your agent, but with a memory for its own mistakes.

“How do I get it?”

npx skills add ckorhonen/reflect — works with Claude Code, Codex, and any harness that reads the open SKILL.md format. Issues sold separately, too.

“Can I trust it hands-off?”

It only applies additive, low-risk changes on its own. Anything destructive or config-shaped becomes a proposed ledger row awaiting your yes. No secrets, no transcript text, ever stored.

“What's in the .reflect folder?”

Plain markdown, committed to your repo: observations.md (hunches with sighting counts) and applied.md (every improvement with status and evidence). Team memory, not a black box.

A fleet of retro rockets rises in formation at dawn while brass robots watch from a mesa.

Don't miss the next issue

YOUR AGENT COULD BE
BETTER TOMORROW

npx skills add ckorhonen/reflect

TRANSMISSION

Music
Sound FX