lore
The working SpecStory skill, packaged with its mining scripts, workflows, approval gate, and upstream provenance.
Show Us Your Agent Skills / EP 07 / guest dossier
GREG CECCARELLI
SPECSTORY
LORE
DEAD RECKON
LOOP ENGINEERING
GREG CECCARELLI
Greg turns the record of agent work into infrastructure for the next run. Lore mined 516 saved sessions, used each next user response as evidence of acceptance or correction, and presented recurring practices with citations before Greg approved any new skill. Dead Reckon protects the other end of the process: Claude Code, Cursor, or Codex keeps working until deterministic checks hidden from the coding model say the outcome is complete. Greg, co-founder and CPO of SpecStory, connects both systems to a new software-development lifecycle built around intent, parallel agent work, and verification.
516saved sessions parsed in the live Lore run
3 TURNSprompt, response, and next prompt form one evidence-bearing beat
MONTHLYGreg's suggested cadence for deeper Lore mining
0 AUTOskills installed without explicit human approval
20 HOURSan unattended Dead Reckon run can continue toward its gate
TRUNKshared integration; branches reserved for experiments
EP 07 · GREG CECCARELLI · Lore, Dead Reckon, loop engineering, and the agentic software-development lifecycle, live on stream
"A lot of the patterns and heuristics that you use are actually in those agent session logs."
The trace contains the why behind a change and the practices a builder repeats without having named them. 00:05:40
SESSION EXHAUST BECOMES A SKILL LIBRARY
the skills-from-agent-history workflow, demonstrated through Lore
SpecStory normalizes histories from Claude Code, Codex, Cursor, Gemini, and other coding agents into markdown inside the project. Lore builds a local database over that corpus instead of sending hundreds of long session files straight to a model.
The deterministic pass discovers sessions, parses turns, and creates beats. Each beat includes the next user prompt, so acceptance, rejection, and correction become evidence about the preceding response. Model-driven theme sweeps and deeper mining then look for recurring practices across projects.
The output is a candidate dossier with citations, evidence strength, corroboration, and adversarial checks. Greg reads the sources, chooses what deserves to become a skill, and can uninstall a forged skill after inspecting the result.
"It presents stuff for your review. It will never install anything unless you actually approve it."
Corroboration narrows the list. Greg still decides whether a recurring behavior belongs in his active skill library. 00:34:00
MINE, CORROBORATE, APPROVE
the reusable workflow behind the live Lore run
Collect histories in one formatSync saved sessions from each coding agent into markdown inside the project. 00:12:15
Index before asking a modelParse the corpus locally so the reasoning model receives candidates and evidence, not every raw token. 00:19:05
Include the next user turnBuild each beat from prompt, response, and follow-up so correction or acceptance remains visible. 00:20:35
Mine across projectsRun theme sweeps, deeper analysis, and corroboration across the full saved corpus. 00:30:10
Inspect the dossier and citationsRead the practice, its evidence, and the sessions Lore cites before deciding it is reusable. 00:31:20
Forge only the useful candidatesApprove the skill explicitly, inspect what Lore generated, and uninstall it if it does not earn a place. 00:36:20
"The model will not be the one determining its end run."
Dead Reckon keeps the stopping decision in a deterministic gate the coding model cannot read or modify. 01:20:05
THE WORKER DOES NOT CONTROL THE FINISH LINE
the verifiable loop-engineering workflow, built around Dead Reckon
A normal coding-agent run stops when the model stops calling tools, returns a completion, or hits a turn limit. Dead Reckon wraps Claude Code, Cursor, or Codex and invokes it again until a protected definition-of-done gate passes.
The worker-facing goal states the outcome, acceptance criteria, phases, and reference material. The harness-side checks live somewhere the worker cannot edit. That separation lets a long run continue without letting the same model build the artifact, alter the yardstick, and declare success.
Greg pairs that gate with cross-model review and functional verification. One provider can implement a change, another can review it, and the final criteria can require a real end-to-end run rather than a confident completion message.
Write the outcome and riderKeep the goal concise and place detailed source intent, context, and phased work in its paired rider. 01:26:00
Define independent checksName the deterministic evidence that must exist before the harness can break the loop. 01:19:14
Hide the stopping gateStore the decisive checks outside the coding model's readable and writable scope. 01:19:55
Run the harness in an outer loopLet the coding CLI plan, edit, test, and retry until the protected gate passes. 01:30:30
Switch model families for reviewGreg implements with Claude Code and reviews a commit range with Codex. 01:34:21
Exercise the real behaviorRequire the program to run and preserve evidence that the full handoff actually worked. 01:35:20
"The unit of work that you're driving these agents towards in agentic engineering is completely different."
Greg expects small teams to spend less time assigning implementation tasks and more time articulating goals, steering parallel agents, and building verification infrastructure. 01:54:40
THE SOFTWARE-DEVELOPMENT LIFECYCLE CHANGES
goals replace task decomposition; verification and integration become the limiting disciplines
Greg describes an operating model built for the amount of code agents can produce. Small teams articulate goals, run agents in parallel, steer them briefly, and verify the resulting behavior. The human work shifts toward source intent, environment design, review criteria, and the infrastructure that proves the result.
Pull requests become a bottleneck when generation outruns human review. Greg integrates continuously into a shared trunk and reserves branches for experiments. When a branch drifts too far, an agent performs a semantic rebase: it reads the branch's intent and reimplements that intent against the current codebase.
Git remains useful as a hardened file-system protocol. The collaboration layer around it can change because agent-written summaries, cross-model review, hidden gates, and semantic reimplementation no longer assume that every unit of work ends in a conventional pull request.
"PRs are the limiting gate. You can produce so much code, who's going to ever review it all?"
The bottleneck moves from producing a patch to proving that a flood of patches satisfies the intended outcome. 01:55:55
GREG'S FIELD KIT
two repo skills, two workflows, the harness, and the book behind the operating model
goal-rider-author
Write a concise outcome document and a detailed rider carrying phases, reference context, and verification criteria.
skills-from-agent-history
Normalize histories, build beats, mine recurring practices, inspect cited evidence, and forge only approved candidates.
verifiable-agent-loops
Keep stopping checks outside the worker, constrain its tools, inspect the trajectory, and repair any gate it learns to game.
Dead Reckon
Greg's outer harness for unattended coding runs controlled by hidden, deterministic definition-of-done checks.
25 Patterns in Agentic Engineering
Greg's free book on the practices behind sustained product development with coding agents.
"I've done this loop now 1,000 times, and I can't take it anymore."
Repeated friction is a signal to extract a skill, build a harness, or change the development process around the agent. 01:49:10
GREG'S STACK
session capture, skill mining, coding harnesses, and verification