A Prompt Library for Building Running Plans With AI

Most runners’ first prompt is some version of “write me a 16-week marathon plan for a sub-3:30.” What comes back looks like a training plan. It has Tuesday intervals, a Sunday long run, a three-week taper. It also, quite often, ramps you from 40km to 70km a week in six weeks, puts two threshold sessions 48 hours apart in week 9, and never once mentions what to do when your calf tightens up.

The prompts collected here are for the stage after that. You already have a plan, from Runna or Garmin Coach or a ChatGPT session in February, and the real question is no longer “what should I run” but “is this thing actually progressing me, or is it just progressing the numbers?” For the ground-up version of this work, the plan architecture itself, the pillar on building your own plan with ChatGPT and Claude covers structure, periodisation and how to pick a goal pace. This page is narrower: it is the library of prompts you keep in a note and paste weekly.

Start With a Context Block, Not a Question

Every prompt below assumes a block of your real data sits above it. Without that, the model answers from the average of every training article ever written, which is how you end up with the 10% rule applied to someone running 12km a week.

Keep this in a saved note (or a Claude Project, or a ChatGPT custom instruction) and update the numbers each Sunday night:

ATHLETE CONTEXT
Age 42. Running 5 years, consistent for the last 14 months.
Current volume: 4 runs/week, 38-44km, one long run 16-18km.
Recent race results (chip times):
  5K   22:40  (Mar 2026)
  10K  47:15  (Jun 2026)
  HM   1:44:52 (Aug 2026, warm, positive split ~90s)
Watch: Garmin Forerunner 265. Garmin VO2max estimate 47.
Garmin 7-day acute load 618, chronic load 471 (ratio 1.31).
Strava Fitness (CTL) 52, up from 44 three weeks ago.
Threshold HR ~168bpm, max observed 186, resting 48.
Injury history: right posterior tibial tendon grumble Nov 2025,
  6 weeks reduced running. Nothing since. Bunion, left foot, mild.
Life: two kids, no weekday session possible before 6am,
  Wednesday is a hard no (childcare).
Goal: Manchester Marathon, 12 April 2027. Target 3:29:xx.
Current plan source: Runna, "Sub 3:30 Marathon", 18 weeks, started week 3.

Notice what is in there that a plan generator never asks: the positive split in the half (evidence the goal pace may be optimistic), the specific tendon that gave way, and Wednesday. Models are reasonably good at respecting hard constraints when you state them as constraints rather than preferences.

Prompt 1: The Progression Audit

This is the one to run before you commit to any block. It asks the model to check the plan against itself, arithmetically, rather than to praise it.

Below is my context block and the full week-by-week volume from my plan. Use your code execution tool to build a table with: week number, total km, week-on-week percentage change, long run km, long run as a percentage of weekly total, number of quality sessions, and the 28-day rolling average. Then flag every week that meets any of these: week-on-week increase over 10%; two consecutive increases over 7%; long run over 33% of weekly volume; more than two quality sessions; no down week within any four-week stretch; long run increasing by more than 3km in a single step. List the flags as a numbered table with the week number and the specific trigger. Do not smooth over anything, and do not offer encouragement.

The code execution instruction matters. Asked to do this in prose, GPT-5 and Claude both drift on arithmetic across an 18-row table, quietly rounding a 14% jump into “a modest increase.” Asked to run Python, they get it right and you can see the working. A real output on a Runna sub-3:30 block looked like this:

FLAGS (6)
wk  trigger                                  value
 5  week-on-week increase                    +13.2% (46.5 -> 52.6 km)
 6  second consecutive increase >7%          +8.9%
 7  long run % of weekly volume              35.4% (21 of 59.3 km)
 9  quality sessions                         3 (tempo Tue, hills Thu, MP long Sun)
11  long run single-step increase            +4 km (26 -> 30)
14  no down week in weeks 11-14              min week 61.2 km

Six flags is not a broken plan. It is a plan built for the median user of an app, which you are not. Weeks 5 and 6 back to back are the pair worth changing, and week 9 is the one most likely to hurt you.

Prompt 2: The Sunday Load Check

Short, run weekly, takes ninety seconds.

Here is last week actual versus planned, plus my Garmin acute/chronic load and Strava Fitness. Tell me in under 150 words whether I am (a) under-recovered, (b) progressing appropriately, or (c) under-stimulated. Base the call on the load ratio, my sleep and resting HR trend, and whether I hit the prescribed paces or drifted. If you cannot tell from the data, say which single number you would need.

That last sentence does more work than it looks like it does. It stops the model bluffing. A useful answer reads like: “Acute/chronic 1.31 with resting HR up 4bpm and two sessions run 8s/km slower than prescribed. That combination is under-recovered, not progressing. Ratio alone I would ignore at 1.31; ratio plus pace drift plus HR I would not.”

Prompt 3: The Adversarial Second Opinion

The single highest-value prompt in the library, because AI plan builders are agreeable by default.

You are a sceptical physiotherapist who has seen a lot of recreational marathoners break down in weeks 9 to 13. I am about to start the block below. Argue against it. Give me the three most likely ways this specific plan injures this specific athlete, ranked by probability, each with the week it would most likely happen and the early warning sign I would notice first. Then give me the minimum change to each that preserves the training effect.

What comes back on the example block above, correctly, centres on the posterior tibial tendon and the jump into week 11’s 30km long run, and suggests splitting it as 24km with 8km at marathon pace instead. That is a better answer than “great plan, good luck.”

Prompt 4: Translating Paces Into Something Your Watch Understands

Plans give you paces. Your watch gives you laps, and your legs give you effort. Bridging those three is where most self-coached runners lose the session.

Convert this workout into: (1) a Garmin structured workout I can build in Garmin Connect, with each step as distance-or-time plus a target range, not a single number; (2) the HR ceiling I should not cross in the easy portions, from my threshold HR of 168; (3) a plain-English “what this should feel like” line per step; (4) the abort criteria, at what point do I bin the rest of the session.

Range targets rather than point targets are the fix for a specific failure: a watch beeping at you for being 3s/km off a number you invented six weeks ago. For the example athlete, 5 x 1km at threshold becomes 4:22 to 4:31/km, HR cap 172 on the reps, jog recoveries under 145, abort if rep 4 is more than 8s slower than rep 2.

Prompt 5: The Niggle Triage

Not medical advice, and worth saying so in the prompt itself so the model stops hedging in every paragraph.

I have a niggle. Details: [location, what makes it worse, what makes it better, day it started, what I ran the three days before, pain 0-10 walking / first 2km / after]. I am not asking for a diagnosis and I will see a physio if this persists past 10 days. Give me: the two or three structures most consistent with this pattern, a single test I can do at home to discriminate between them, what this week’s plan should become for each possibility, and the specific red flags that mean stop running today.

Which Prompt For Which Question

The Goal-Pace Reality Check

Halfway through a block, the honest question arrives: is 3:29 still on? Feed the model your actual session data rather than your hopes.

From these six sessions (paces, HR, RPE, conditions attached) and my race results, estimate my current marathon capability three ways: Riegel from the August half, a VDOT-equivalent from the 10K, and a threshold-based estimate from my recent tempo work. Show each number separately and tell me where they disagree and why. Then tell me what my last 8 weeks would need to look like for 3:29 to be a 70% proposition rather than a 30% one.

Three separate estimates that disagree is more informative than one confident number. On the example data, Riegel from a warm, positive-split half gives roughly 3:39, the 10K gives around 3:34, and threshold work suggests 3:36. All three land north of target, which is useful eleven weeks out and useless on race morning.

Where These Prompts Still Fail You

Models invent training-load thresholds with total confidence. ACWR between 0.8 and 1.3 comes from cricket and Australian football research, and its application to recreational distance running is genuinely contested, so treat a 1.35 as a prompt to look at your sleep rather than a verdict. Ask any model for a citation and check whether the study exists.

Long conversations are the other trap. By message forty, the model has absorbed your optimism about goal pace and will defend it. Start a clean session for the Adversarial Second Opinion every time, with only the context block pasted in.

Keep a running note of what each model got wrong and how you found out. Six weeks of that, and the corrections themselves become the most valuable thing in your prompt library.