Manual-to-Skill Loop
Turn a successful AI session into a repeatable operating procedure
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 6
- Confidence
- 98%
The Manual-to-Skill Loop treats automation as the final stage of discovery rather than the starting point. First, an operator completes a real task by conversing with an agent, adding tools, APIs, constraints, and corrections until the result passes a human taste test. That interaction provides evidence about the inputs, decisions, sequence, guardrails, and output standard that the system actually needs. Only after reaching a strong result does the operator ask the agent to inspect the complete session and encode what worked as a reusable skill. The resulting procedure can then reproduce the workflow with less prompting. This mechanism lets non-engineers develop systems empirically: execution reveals the process, human judgment validates it, and the agent formalizes it.
Origin
James Dickerson described this method on Marketing Against The Grain after explaining that node-based automation taught him linear workflows but agentic tools changed how he builds reusable systems.
Core principles
- 01Discover the workflow through real execution
- 02Judge outputs with human taste
- 03Iterate before attempting automation
- 04Encode the entire successful path
- 05Let the agent identify the reproducible system
How to run it
- 1
Select a Small Real Task
Choose a bounded task with a concrete deliverable and a result you can personally evaluate. Avoid designing an abstract automation before attempting the work.
Pro tip Start with something you already understand well enough to judge.
Watch out A vague task produces a vague and unreliable skill.
- 2
Work the Task Manually
Collaborate with the agent through natural-language prompts. Give it access to relevant files, APIs, references, and tools as those needs emerge.
Pro tip Tell the agent your experience level so it can guide you appropriately.
Watch out Do not treat the first output as proof that the workflow works.
- 3
Iterate to the Quality Bar
Prompt, reprompt, and adjust the available context until the output is genuinely strong. Use human taste to decide when the result is good enough.
Pro tip Record the corrections that materially improve the result.
Watch out Automating a mediocre session only makes mediocre output reproducible.
- 4
Review the Full Session
Ask the agent to analyze everything performed during the session, including successful decisions, failed attempts, dependencies, and guardrails.
Pro tip Ask explicitly what worked, what did not, and why.
Watch out Reviewing only the final prompt hides the iterative knowledge that created the result.
- 5
Encode the Skill
Have the agent turn the proven workflow into a persistent skill with triggers, instructions, references, and quality constraints.
Pro tip Include examples that demonstrate what good output looks like.
Watch out Do not omit tool requirements or failure-handling instructions.
- 6
Validate Reproducibility
Run the new skill against a fresh but comparable task and compare its output with the original standard. Refine the skill when it misses important decisions.
Pro tip Test with different inputs rather than replaying the original example.
Watch out A skill that succeeds only on its source example is not reusable.
In the wild
A marketer asks an agent to research several competitors, supplies Firecrawl and Perplexity, corrects weak comparisons, and iterates until the analysis reveals a credible market gap. The marketer then asks the agent to review the session and encode the sources, comparison criteria, gap analysis, and output format as a reusable competitor-research skill.
→ Future competitor assessments follow the validated process with less prompting and more consistent outputs.
A team member builds one weekly content brief interactively, refining topic selection, evidence requirements, voice, and formatting until stakeholders approve it. The agent converts the complete interaction into a skill and tests it against the following week's source material.
→ The team gains a repeatable briefing process while retaining human approval as the quality gate.
Common mistakes
Automating Before Discovering
Designing the system before doing the task can encode imagined steps instead of the process that actually produces a good result.
Accepting the First Output
A first draft rarely reveals all required context, guardrails, or quality criteria. Iterate until human judgment confirms the result.
Ignoring Failed Attempts
Failures reveal important warnings and constraints. Excluding them from the review makes the resulting skill brittle.
Is it for you?
Best for
Non-technical operators who repeatedly perform a task but do not yet know its ideal automated workflow.
Not ideal for
Unproven tasks whose desired output and quality standard remain unclear.
From the transcript
“So, what I do is I go through manually and do it just go back and forth and prompt and reprompt give it different APIs,…”
“And then once I get there, I say, "Hey, review everything that we did in this session. Turn that into a skill. There's the system."”
“So, yeah. So, the the way that you think in systems is just about learning to work with the agent until you arrive at the…”
From the episode
Claude Code: Landing Page to Lead Magnet in 50 Minutes